US20190068464A1

US20190068464A1 - Technologies for machine learning schemes in dynamic switching between adaptive connections and connection optimization

Info

Publication number: US20190068464A1
Application number: US15/858,305
Authority: US
Inventors: Francesc Guim Bernat; Susanne M. Balle; Rahul Khanna; Evan Custodio
Original assignee: Intel Corp
Current assignee: Intel Corp
Priority date: 2017-08-30
Filing date: 2017-12-29
Publication date: 2019-02-28
Also published as: EP3676708A4; CN109428889B; US20190065260A1; US20190069433A1; US11416309B2; US20190067848A1; US11030017B2; US20200257566A1; CN109428841A; US20190065261A1; WO2019045929A1; CN109426630A; US20190068693A1; CN109426646A; CN109426646B; US20190068698A1; DE112018004798T5; EP3676708A1; US11422867B2; US20190068521A1

Abstract

Technologies for adapting a communication protocol (e.g., TCP/IP, UDP, etc.) to network communications between endpoints (e.g., accelerated kernels configured within accelerator devices) include a sled having a compute engine. The compute engine monitors telemetry data associated with one or more network communications between a given kernel and another kernel. The network communications are established via a given communication protocol. The compute engine determines, as a function of the monitored telemetry data, that a condition to change the network communications from the communication protocol to another communication protocol is triggered. The compute engine shifts the network communications to the other communication protocol.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

The present application claims the benefit of Indian Provisional Patent Application No. 201741030632, filed Aug. 30, 2017 and U.S. Provisional Patent Application No. 62/584,401, filed Nov. 10, 2017.

BACKGROUND

In systems that distribute workloads among multiple compute devices (e.g., in a data center), a centralized server may compose nodes of compute devices to process the workloads. Each node represents a logical aggregation of resources (e.g., compute, storage, acceleration, and the like) provided by each compute device. For instance, the node may include a compute device configured with hardware accelerators, such as field-programmable gate array (FPGA) devices and/or graphical processing units (GPUs). Generally, the hardware accelerator improves the execution speed of workload functions. To accelerate a given function of a workload, such as of an application, the centralized server may configure an accelerator device with an accelerated kernel that is suitable for accelerating the task.
Once complete, the accelerator device returns data resulting from the accelerated function to the application. In some cases, the system may provide a kernel-to-kernel network that allows a given kernel to transmit the resulting data to another kernel to further process the workload. A kernel in an accelerator device may establish a network communication with another kernel device via some network communication protocol, such as TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol). Based on a present resource load in the system, using a given network communication protocol may be more efficient than using another protocol.

BRIEF DESCRIPTION OF THE DRAWINGS

The concepts described herein are illustrated by way of example and not by way of limitation in the accompanying figures. For simplicity and clarity of illustration, elements illustrated in the figures are not necessarily drawn to scale. Where considered appropriate, reference labels have been repeated among the figures to indicate corresponding or analogous elements.

FIG. 1 is a simplified diagram of at least one embodiment of a data center for executing workloads with disaggregated resources;

FIG. 2 is a simplified diagram of at least one embodiment of a pod of the data center of FIG. 1;

FIG. 3 is a perspective view of at least one embodiment of a rack that may be included in the pod of FIG. 2;

FIG. 4 is a side plan elevation view of the rack of FIG. 3;

FIG. 5 is a perspective view of the rack of FIG. 3 having a sled mounted therein;

FIG. 6 is a is a simplified block diagram of at least one embodiment of a top side of the sled of FIG. 5;

FIG. 7 is a simplified block diagram of at least one embodiment of a bottom side of the sled of FIG. 6;

FIG. 8 is a simplified block diagram of at least one embodiment of a compute sled usable in the data center of FIG. 1;

FIG. 9 is a top perspective view of at least one embodiment of the compute sled of FIG. 8;

FIG. 10 is a simplified block diagram of at least one embodiment of an accelerator sled usable in the data center of FIG. 1;

FIG. 11 is a top perspective view of at least one embodiment of the accelerator sled of FIG. 10;

FIG. 12 is a simplified block diagram of at least one embodiment of a storage sled usable in the data center of FIG. 1;

FIG. 13 is a top perspective view of at least one embodiment of the storage sled of FIG. 12;

FIG. 14 is a simplified block diagram of at least one embodiment of a memory sled usable in the data center of FIG. 1; and

FIG. 15 is a simplified block diagram of a system that may be established within the data center of FIG. 1 to execute workloads with managed nodes composed of disaggregated resources.

FIG. 16 is a simplified block diagram of at least one embodiment of a system for adapting a communication protocol to network communications between endpoints;

FIG. 17 is a simplified block diagram of at least one embodiment of an accelerator sled of the system of FIG. 16;

FIG. 18 is a simplified block diagram of at least one embodiment of an environment that may be established by the accelerator sled of FIGS. 16 and 17;

FIGS. 19A and 19B are diagrams of an example embodiment of a kernel-to-kernel communication network; and

FIGS. 20 and 21 are simplified flow diagrams of at least one embodiment of a method for adapting a communication protocol to network communications between endpoints.

DETAILED DESCRIPTION OF THE DRAWINGS

While the concepts of the present disclosure are susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawings and will be described herein in detail. It should be understood, however, that there is no intent to limit the concepts of the present disclosure to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives consistent with the present disclosure and the appended claims.
References in the specification to “one embodiment,” “an embodiment,” “an illustrative embodiment,” etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may or may not necessarily include that particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described. Additionally, it should be appreciated that items included in a list in the form of “at least one A, B, and C” can mean (A); (B); (C); (A and B); (A and C); (B and C); or (A, B, and C). Similarly, items listed in the form of “at least one of A, B, or C” can mean (A); (B); (C); (A and B); (A and C); (B and C); or (A, B, and C).
The disclosed embodiments may be implemented, in some cases, in hardware, firmware, software, or any combination thereof. The disclosed embodiments may also be implemented as instructions carried by or stored on a transitory or non-transitory machine-readable (e.g., computer-readable) storage medium, which may be read and executed by one or more processors. A machine-readable storage medium may be embodied as any storage device, mechanism, or other physical structure for storing or transmitting information in a form readable by a machine (e.g., a volatile or non-volatile memory, a media disc, or other media device).
In the drawings, some structural or method features may be shown in specific arrangements and/or orderings. However, it should be appreciated that such specific arrangements and/or orderings may not be required. Rather, in some embodiments, such features may be arranged in a different manner and/or order than shown in the illustrative figures. Additionally, the inclusion of a structural or method feature in a particular figure is not meant to imply that such feature is required in all embodiments and, in some embodiments, may not be included or may be combined with other features.
Referring now to FIG. 1, a data center 100 in which disaggregated resources may cooperatively execute one or more workloads (e.g., applications on behalf of customers) includes multiple pods 110, 120, 130, 140, each of which includes one or more rows of racks. As described in more detail herein, each rack houses multiple sleds, which each may be embodied as a compute device, such as a server, that is primarily equipped with a particular type of resource (e.g., memory devices, data storage devices, accelerator devices, general purpose processors). In the illustrative embodiment, the sleds in each pod 110, 120, 130, 140 are connected to multiple pod switches (e.g., switches that route data communications to and from sleds within the pod). The pod switches, in turn, connect with spine switches 150 that switch communications among pods (e.g., the pods 110, 120, 130, 140) in the data center 100. In some embodiments, the sleds may be connected with a fabric using Intel Omni-Path technology. As described in more detail herein, resources within sleds in the data center 100 may be allocated to a group (referred to herein as a “managed node”) containing resources from one or more other sleds to be collectively utilized in the execution of a workload. The workload can execute as if the resources belonging to the managed node were located on the same sled. The resources in a managed node may even belong to sleds belonging to different racks, and even to different pods 110, 120, 130, 140. Some resources of a single sled may be allocated to one managed node while other resources of the same sled are allocated to a different managed node (e.g., one processor assigned to one managed node and another processor of the same sled assigned to a different managed node). By disaggregating resources to sleds comprised predominantly of a single type of resource (e.g., compute sleds comprising primarily compute resources, memory sleds containing primarily memory resources), and selectively allocating and deallocating the disaggregated resources to form a managed node assigned to execute a workload, the data center 100 provides more efficient resource usage over typical data centers comprised of hyperconverged servers containing compute, memory, storage and perhaps additional resources). As such, the data center 100 may provide greater performance (e.g., throughput, operations per second, latency, etc.) than a typical data center that has the same number of resources.
Referring now to FIG. 2, the pod 110, in the illustrative embodiment, includes a set of rows 200, 210, 220, 230 of racks 240. Each rack 240 may house multiple sleds (e.g., sixteen sleds) and provide power and data connections to the housed sleds, as described in more detail herein. In the illustrative embodiment, the racks in each row 200, 210, 220, 230 are connected to multiple pod switches 250, 260. The pod switch 250 includes a set of ports 252 to which the sleds of the racks of the pod 110 are connected and another set of ports 254 that connect the pod 110 to the spine switches 150 to provide connectivity to other pods in the data center 100. Similarly, the pod switch 260 includes a set of ports 262 to which the sleds of the racks of the pod 110 are connected and a set of ports 264 that connect the pod 110 to the spine switches 150. As such, the use of the pair of switches 250, 260 provides an amount of redundancy to the pod 110. For example, if either of the switches 250, 260 fails, the sleds in the pod 110 may still maintain data communication with the remainder of the data center 100 (e.g., sleds of other pods) through the other switch 250, 260. Furthermore, in the illustrative embodiment, the switches 150, 250, 260 may be embodied as dual-mode optical switches, capable of routing both Ethernet protocol communications carrying Internet Protocol (IP) packets and communications according to a second, high-performance link-layer protocol (e.g., Intel's Omni-Path Architecture's, Infiniband) via optical signaling media of an optical fabric.
It should be appreciated that each of the other pods 120, 130, 140 (as well as any additional pods of the data center 100) may be similarly structured as, and have components similar to, the pod 110 shown in and described in regard to FIG. 2 (e.g., each pod may have rows of racks housing multiple sleds as described above). Additionally, while two pod switches 250, 260 are shown, it should be understood that in other embodiments, each pod 110, 120, 130, 140 may be connected to different number of pod switches (e.g., providing even more failover capacity).
Referring now to FIGS. 3-5, each illustrative rack 240 of the data center 100 includes two elongated support posts 302, 304, which are arranged vertically. For example, the elongated support posts 302, 304 may extend upwardly from a floor of the data center 100 when deployed. The rack 240 also includes one or more horizontal pairs 310 of elongated support arms 312 (identified in FIG. 3 via a dashed ellipse) configured to support a sled of the data center 100 as discussed below. One elongated support arm 312 of the pair of elongated support arms 312 extends outwardly from the elongated support post 302 and the other elongated support arm 312 extends outwardly from the elongated support post 304.
In the illustrative embodiments, each sled of the data center 100 is embodied as a chassis-less sled. That is, each sled has a chassis-less circuit board substrate on which physical resources (e.g., processors, memory, accelerators, storage, etc.) are mounted as discussed in more detail below. As such, the rack 240 is configured to receive the chassis-less sleds. For example, each pair 310 of elongated support arms 312 defines a sled slot 320 of the rack 240, which is configured to receive a corresponding chassis-less sled. To do so, each illustrative elongated support arm 312 includes a circuit board guide 330 configured to receive the chassis-less circuit board substrate of the sled. Each circuit board guide 330 is secured to, or otherwise mounted to, a top side 332 of the corresponding elongated support arm 312. For example, in the illustrative embodiment, each circuit board guide 330 is mounted at a distal end of the corresponding elongated support arm 312 relative to the corresponding elongated support post 302, 304. For clarity of the Figures, not every circuit board guide 330 may be referenced in each Figure.
Each circuit board guide 330 includes an inner wall that defines a circuit board slot 380 configured to receive the chassis-less circuit board substrate of a sled 400 when the sled 400 is received in the corresponding sled slot 320 of the rack 240. To do so, as shown in FIG. 4, a user (or robot) aligns the chassis-less circuit board substrate of an illustrative chassis-less sled 400 to a sled slot 320. The user, or robot, may then slide the chassis-less circuit board substrate forward into the sled slot 320 such that each side edge 414 of the chassis-less circuit board substrate is received in a corresponding circuit board slot 380 of the circuit board guides 330 of the pair 310 of elongated support arms 312 that define the corresponding sled slot 320 as shown in FIG. 4. By having robotically accessible and robotically manipulable sleds comprising disaggregated resources, each type of resource can be upgraded independently of each other and at their own optimized refresh rate. Furthermore, the sleds are configured to blindly mate with power and data communication cables in each rack 240, enhancing their ability to be quickly removed, upgraded, reinstalled, and/or replaced. As such, in some embodiments, the data center 100 may operate (e.g., execute workloads, undergo maintenance and/or upgrades, etc.) without human involvement on the data center floor. In other embodiments, a human may facilitate one or more maintenance or upgrade operations in the data center 100.
It should be appreciated that each circuit board guide 330 is dual sided. That is, each circuit board guide 330 includes an inner wall that defines a circuit board slot 380 on each side of the circuit board guide 330. In this way, each circuit board guide 330 can support a chassis-less circuit board substrate on either side. As such, a single additional elongated support post may be added to the rack 240 to turn the rack 240 into a two-rack solution that can hold twice as many sled slots 320 as shown in FIG. 3. The illustrative rack 240 includes seven pairs 310 of elongated support arms 312 that define a corresponding seven sled slots 320, each configured to receive and support a corresponding sled 400 as discussed above. Of course, in other embodiments, the rack 240 may include additional or fewer pairs 310 of elongated support arms 312 (i.e., additional or fewer sled slots 320). It should be appreciated that because the sled 400 is chassis-less, the sled 400 may have an overall height that is different than typical servers. As such, in some embodiments, the height of each sled slot 320 may be shorter than the height of a typical server (e.g., shorter than a single rank unit, “1 U”). That is, the vertical distance between each pair 310 of elongated support arms 312 may be less than a standard rack unit “1 U.” Additionally, due to the relative decrease in height of the sled slots 320, the overall height of the rack 240 in some embodiments may be shorter than the height of traditional rack enclosures. For example, in some embodiments, each of the elongated support posts 302, 304 may have a length of six feet or less. Again, in other embodiments, the rack 240 may have different dimensions. Further, it should be appreciated that the rack 240 does not include any walls, enclosures, or the like. Rather, the rack 240 is an enclosure-less rack that is opened to the local environment. Of course, in some cases, an end plate may be attached to one of the elongated support posts 302, 304 in those situations in which the rack 240 forms an end-of-row rack in the data center 100.
In some embodiments, various interconnects may be routed upwardly or downwardly through the elongated support posts 302, 304. To facilitate such routing, each elongated support post 302, 304 includes an inner wall that defines an inner chamber in which the interconnect may be located. The interconnects routed through the elongated support posts 302, 304 may be embodied as any type of interconnects including, but not limited to, data or communication interconnects to provide communication connections to each sled slot 320, power interconnects to provide power to each sled slot 320, and/or other types of interconnects.
The rack 240, in the illustrative embodiment, includes a support platform on which a corresponding optical data connector (not shown) is mounted. Each optical data connector is associated with a corresponding sled slot 320 and is configured to mate with an optical data connector of a corresponding sled 400 when the sled 400 is received in the corresponding sled slot 320. In some embodiments, optical connections between components (e.g., sleds, racks, and switches) in the data center 100 are made with a blind mate optical connection. For example, a door on each cable may prevent dust from contaminating the fiber inside the cable. In the process of connecting to a blind mate optical connector mechanism, the door is pushed open when the end of the cable enters the connector mechanism. Subsequently, the optical fiber inside the cable enters a gel within the connector mechanism and the optical fiber of one cable comes into contact with the optical fiber of another cable within the gel inside the connector mechanism.
The illustrative rack 240 also includes a fan array 370 coupled to the cross-support arms of the rack 240. The fan array 370 includes one or more rows of cooling fans 372, which are aligned in a horizontal line between the elongated support posts 302, 304. In the illustrative embodiment, the fan array 370 includes a row of cooling fans 372 for each sled slot 320 of the rack 240. As discussed above, each sled 400 does not include any on-board cooling system in the illustrative embodiment and, as such, the fan array 370 provides cooling for each sled 400 received in the rack 240. Each rack 240, in the illustrative embodiment, also includes a power supply associated with each sled slot 320. Each power supply is secured to one of the elongated support arms 312 of the pair 310 of elongated support arms 312 that define the corresponding sled slot 320. For example, the rack 240 may include a power supply coupled or secured to each elongated support arm 312 extending from the elongated support post 302. Each power supply includes a power connector configured to mate with a power connector of the sled 400 when the sled 400 is received in the corresponding sled slot 320. In the illustrative embodiment, the sled 400 does not include any on-board power supply and, as such, the power supplies provided in the rack 240 supply power to corresponding sleds 400 when mounted to the rack 240.
Referring now to FIG. 6, the sled 400, in the illustrative embodiment, is configured to be mounted in a corresponding rack 240 of the data center 100 as discussed above. In some embodiments, each sled 400 may be optimized or otherwise configured for performing particular tasks, such as compute tasks, acceleration tasks, data storage tasks, etc. For example, the sled 400 may be embodied as a compute sled 800 as discussed below in regard to FIGS. 8-9, an accelerator sled 1000 as discussed below in regard to FIGS. 10-11, a storage sled 1200 as discussed below in regard to FIGS. 12-13, or as a sled optimized or otherwise configured to perform other specialized tasks, such as a memory sled 1400, discussed below in regard to FIG. 14.
As discussed above, the illustrative sled 400 includes a chassis-less circuit board substrate 602, which supports various physical resources (e.g., electrical components) mounted thereon. It should be appreciated that the circuit board substrate 602 is “chassis-less” in that the sled 400 does not include a housing or enclosure. Rather, the chassis-less circuit board substrate 602 is open to the local environment. The chassis-less circuit board substrate 602 may be formed from any material capable of supporting the various electrical components mounted thereon. For example, in an illustrative embodiment, the chassis-less circuit board substrate 602 is formed from an FR-4 glass-reinforced epoxy laminate material. Of course, other materials may be used to form the chassis-less circuit board substrate 602 in other embodiments.
As discussed in more detail below, the chassis-less circuit board substrate 602 includes multiple features that improve the thermal cooling characteristics of the various electrical components mounted on the chassis-less circuit board substrate 602. As discussed, the chassis-less circuit board substrate 602 does not include a housing or enclosure, which may improve the airflow over the electrical components of the sled 400 by reducing those structures that may inhibit air flow. For example, because the chassis-less circuit board substrate 602 is not positioned in an individual housing or enclosure, there is no backplane (e.g., a backplate of the chassis) to the chassis-less circuit board substrate 602, which could inhibit air flow across the electrical components. Additionally, the chassis-less circuit board substrate 602 has a geometric shape configured to reduce the length of the airflow path across the electrical components mounted to the chassis-less circuit board substrate 602. For example, the illustrative chassis-less circuit board substrate 602 has a width 604 that is greater than a depth 606 of the chassis-less circuit board substrate 602. In one particular embodiment, for example, the chassis-less circuit board substrate 602 has a width of about 21 inches and a depth of about 9 inches, compared to a typical server that has a width of about 17 inches and a depth of about 39 inches. As such, an airflow path 608 that extends from a front edge 610 of the chassis-less circuit board substrate 602 toward a rear edge 612 has a shorter distance relative to typical servers, which may improve the thermal cooling characteristics of the sled 400. Furthermore, although not illustrated in FIG. 6, the various physical resources mounted to the chassis-less circuit board substrate 602 are mounted in corresponding locations such that no two substantively heat-producing electrical components shadow each other as discussed in more detail below. That is, no two electrical components, which produce appreciable heat during operation (i.e., greater than a nominal heat sufficient enough to adversely impact the cooling of another electrical component), are mounted to the chassis-less circuit board substrate 602 linearly in-line with each other along the direction of the airflow path 608 (i.e., along a direction extending from the front edge 610 toward the rear edge 612 of the chassis-less circuit board substrate 602).
As discussed above, the illustrative sled 400 includes one or more physical resources 620 mounted to a top side 650 of the chassis-less circuit board substrate 602. Although two physical resources 620 are shown in FIG. 6, it should be appreciated that the sled 400 may include one, two, or more physical resources 620 in other embodiments. The physical resources 620 may be embodied as any type of processor, controller, or other compute circuit capable of performing various tasks such as compute functions and/or controlling the functions of the sled 400 depending on, for example, the type or intended functionality of the sled 400. For example, as discussed in more detail below, the physical resources 620 may be embodied as high-performance processors in embodiments in which the sled 400 is embodied as a compute sled, as accelerator co-processors or circuits in embodiments in which the sled 400 is embodied as an accelerator sled, storage controllers in embodiments in which the sled 400 is embodied as a storage sled, or a set of memory devices in embodiments in which the sled 400 is embodied as a memory sled.
The sled 400 also includes one or more additional physical resources 630 mounted to the top side 650 of the chassis-less circuit board substrate 602. In the illustrative embodiment, the additional physical resources include a network interface controller (NIC) as discussed in more detail below. Of course, depending on the type and functionality of the sled 400, the physical resources 630 may include additional or other electrical components, circuits, and/or devices in other embodiments.
The physical resources 620 are communicatively coupled to the physical resources 630 via an input/output (I/O) subsystem 622. The I/O subsystem 622 may be embodied as circuitry and/or components to facilitate input/output operations with the physical resources 620, the physical resources 630, and/or other components of the sled 400. For example, the I/O subsystem 622 may be embodied as, or otherwise include, memory controller hubs, input/output control hubs, integrated sensor hubs, firmware devices, communication links (e.g., point-to-point links, bus links, wires, cables, light guides, printed circuit board traces, etc.), and/or other components and subsystems to facilitate the input/output operations. In the illustrative embodiment, the I/O subsystem 622 is embodied as, or otherwise includes, a double data rate 4 (DDR4) data bus or a DDR5 data bus.
In some embodiments, the sled 400 may also include a resource-to-resource interconnect 624. The resource-to-resource interconnect 624 may be embodied as any type of communication interconnect capable of facilitating resource-to-resource communications. In the illustrative embodiment, the resource-to-resource interconnect 624 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the resource-to-resource interconnect 624 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to resource-to-resource communications.
The sled 400 also includes a power connector 640 configured to mate with a corresponding power connector of the rack 240 when the sled 400 is mounted in the corresponding rack 240. The sled 400 receives power from a power supply of the rack 240 via the power connector 640 to supply power to the various electrical components of the sled 400. That is, the sled 400 does not include any local power supply (i.e., an on-board power supply) to provide power to the electrical components of the sled 400. The exclusion of a local or on-board power supply facilitates the reduction in the overall footprint of the chassis-less circuit board substrate 602, which may increase the thermal cooling characteristics of the various electrical components mounted on the chassis-less circuit board substrate 602 as discussed above. In some embodiments, power is provided to the processors 820 through vias directly under the processors 820 (e.g., through the bottom side 750 of the chassis-less circuit board substrate 602), providing an increased thermal budget, additional current and/or voltage, and better voltage control over typical boards.
In some embodiments, the sled 400 may also include mounting features 642 configured to mate with a mounting arm, or other structure, of a robot to facilitate the placement of the sled 600 in a rack 240 by the robot. The mounting features 642 may be embodied as any type of physical structures that allow the robot to grasp the sled 400 without damaging the chassis-less circuit board substrate 602 or the electrical components mounted thereto. For example, in some embodiments, the mounting features 642 may be embodied as non-conductive pads attached to the chassis-less circuit board substrate 602. In other embodiments, the mounting features may be embodied as brackets, braces, or other similar structures attached to the chassis-less circuit board substrate 602. The particular number, shape, size, and/or make-up of the mounting feature 642 may depend on the design of the robot configured to manage the sled 400.
Referring now to FIG. 7, in addition to the physical resources 630 mounted on the top side 650 of the chassis-less circuit board substrate 602, the sled 400 also includes one or more memory devices 720 mounted to a bottom side 750 of the chassis-less circuit board substrate 602. That is, the chassis-less circuit board substrate 602 is embodied as a double-sided circuit board. The physical resources 620 are communicatively coupled to the memory devices 720 via the I/O subsystem 622. For example, the physical resources 620 and the memory devices 720 may be communicatively coupled by one or more vias extending through the chassis-less circuit board substrate 602. Each physical resource 620 may be communicatively coupled to a different set of one or more memory devices 720 in some embodiments. Alternatively, in other embodiments, each physical resource 620 may be communicatively coupled to each memory devices 720.
The memory devices 720 may be embodied as any type of memory device capable of storing data for the physical resources 620 during operation of the sled 400, such as any type of volatile (e.g., dynamic random access memory (DRAM), etc.) or non-volatile memory. Volatile memory may be a storage medium that requires power to maintain the state of data stored by the medium. Non-limiting examples of volatile memory may include various types of random access memory (RAM), such as dynamic random access memory (DRAM) or static random access memory (SRAM). One particular type of DRAM that may be used in a memory module is synchronous dynamic random access memory (SDRAM). In particular embodiments, DRAM of a memory component may comply with a standard promulgated by JEDEC, such as JESD79F for DDR SDRAM, JESD79-2F for DDR2 SDRAM, JESD79-3F for DDR3 SDRAM, JESD79-4A for DDR4 SDRAM, JESD209 for Low Power DDR (LPDDR), JESD209-2 for LPDDR2, JESD209-3 for LPDDR3, and JESD209-4 for LPDDR4 (these standards are available at www.jedec.org). Such standards (and similar standards) may be referred to as DDR-based standards and communication interfaces of the storage devices that implement such standards may be referred to as DDR-based interfaces.
In one embodiment, the memory device is a block addressable memory device, such as those based on NAND or NOR technologies. A memory device may also include next-generation nonvolatile devices, such as Intel 3D XPoint™ memory or other byte addressable write-in-place nonvolatile memory devices. In one embodiment, the memory device may be or may include memory devices that use chalcogenide glass, multi-threshold level NAND flash memory, NOR flash memory, single or multi-level Phase Change Memory (PCM), a resistive memory, nanowire memory, ferroelectric transistor random access memory (FeTRAM), anti-ferroelectric memory, magnetoresistive random access memory (MRAM) memory that incorporates memristor technology, resistive memory including the metal oxide base, the oxygen vacancy base and the conductive bridge Random Access Memory (CB-RAM), or spin transfer torque (STT)-MRAM, a spintronic magnetic junction memory based device, a magnetic tunneling junction (MTJ) based device, a DW (Domain Wall) and SOT (Spin Orbit Transfer) based device, a thyristor based memory device, or a combination of any of the above, or other memory. The memory device may refer to the die itself and/or to a packaged memory product. In some embodiments, the memory device may comprise a transistor-less stackable cross point architecture in which memory cells sit at the intersection of word lines and bit lines and are individually addressable and in which bit storage is based on a change in bulk resistance.
Referring now to FIG. 8, in some embodiments, the sled 400 may be embodied as a compute sled 800. The compute sled 800 is optimized, or otherwise configured, to perform compute tasks. Of course, as discussed above, the compute sled 800 may rely on other sleds, such as acceleration sleds and/or storage sleds, to perform such compute tasks. The compute sled 800 includes various physical resources (e.g., electrical components) similar to the physical resources of the sled 400, which have been identified in FIG. 8 using the same reference numbers. The description of such components provided above in regard to FIGS. 6 and 7 applies to the corresponding components of the compute sled 800 and is not repeated herein for clarity of the description of the compute sled 800.
In the illustrative compute sled 800, the physical resources 620 are embodied as processors 820. Although only two processors 820 are shown in FIG. 8, it should be appreciated that the compute sled 800 may include additional processors 820 in other embodiments. Illustratively, the processors 820 are embodied as high-performance processors 820 and may be configured to operate at a relatively high power rating. Although the processors 820 generate additional heat operating at power ratings greater than typical processors (which operate at around 155-230 W), the enhanced thermal cooling characteristics of the chassis-less circuit board substrate 602 discussed above facilitate the higher power operation. For example, in the illustrative embodiment, the processors 820 are configured to operate at a power rating of at least 250 W. In some embodiments, the processors 820 may be configured to operate at a power rating of at least 350 W.
In some embodiments, the compute sled 800 may also include a processor-to-processor interconnect 842. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the processor-to-processor interconnect 842 may be embodied as any type of communication interconnect capable of facilitating processor-to-processor interconnect 842 communications. In the illustrative embodiment, the processor-to-processor interconnect 842 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the processor-to-processor interconnect 842 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications.
The compute sled 800 also includes a communication circuit 830. The illustrative communication circuit 830 includes a network interface controller (NIC) 832, which may also be referred to as a host fabric interface (HFI). The NIC 832 may be embodied as, or otherwise include, any type of integrated circuit, discrete circuits, controller chips, chipsets, add-in-boards, daughtercards, network interface cards, other devices that may be used by the compute sled 800 to connect with another compute device (e.g., with other sleds 400). In some embodiments, the NIC 832 may be embodied as part of a system-on-a-chip (SoC) that includes one or more processors, or included on a multichip package that also contains one or more processors. In some embodiments, the NIC 832 may include a local processor (not shown) and/or a local memory (not shown) that are both local to the NIC 832. In such embodiments, the local processor of the NIC 832 may be capable of performing one or more of the functions of the processors 820. Additionally or alternatively, in such embodiments, the local memory of the NIC 832 may be integrated into one or more components of the compute sled at the board level, socket level, chip level, and/or other levels.
The communication circuit 830 is communicatively coupled to an optical data connector 834. The optical data connector 834 is configured to mate with a corresponding optical data connector of the rack 240 when the compute sled 800 is mounted in the rack 240. Illustratively, the optical data connector 834 includes a plurality of optical fibers which lead from a mating surface of the optical data connector 834 to an optical transceiver 836. The optical transceiver 836 is configured to convert incoming optical signals from the rack-side optical data connector to electrical signals and to convert electrical signals to outgoing optical signals to the rack-side optical data connector. Although shown as forming part of the optical data connector 834 in the illustrative embodiment, the optical transceiver 836 may form a portion of the communication circuit 830 in other embodiments.
In some embodiments, the compute sled 800 may also include an expansion connector 840. In such embodiments, the expansion connector 840 is configured to mate with a corresponding connector of an expansion chassis-less circuit board substrate to provide additional physical resources to the compute sled 800. The additional physical resources may be used, for example, by the processors 820 during operation of the compute sled 800. The expansion chassis-less circuit board substrate may be substantially similar to the chassis-less circuit board substrate 602 discussed above and may include various electrical components mounted thereto. The particular electrical components mounted to the expansion chassis-less circuit board substrate may depend on the intended functionality of the expansion chassis-less circuit board substrate. For example, the expansion chassis-less circuit board substrate may provide additional compute resources, memory resources, and/or storage resources. As such, the additional physical resources of the expansion chassis-less circuit board substrate may include, but is not limited to, processors, memory devices, storage devices, and/or accelerator circuits including, for example, field programmable gate arrays (FPGA), application-specific integrated circuits (ASICs), security co-processors, graphics processing units (GPUs), machine learning circuits, or other specialized processors, controllers, devices, and/or circuits.
Referring now to FIG. 9, an illustrative embodiment of the compute sled 800 is shown. As shown, the processors 820, communication circuit 830, and optical data connector 834 are mounted to the top side 650 of the chassis-less circuit board substrate 602. Any suitable attachment or mounting technology may be used to mount the physical resources of the compute sled 800 to the chassis-less circuit board substrate 602. For example, the various physical resources may be mounted in corresponding sockets (e.g., a processor socket), holders, or brackets. In some cases, some of the electrical components may be directly mounted to the chassis-less circuit board substrate 602 via soldering or similar techniques.
As discussed above, the individual processors 820 and communication circuit 830 are mounted to the top side 650 of the chassis-less circuit board substrate 602 such that no two heat-producing, electrical components shadow each other. In the illustrative embodiment, the processors 820 and communication circuit 830 are mounted in corresponding locations on the top side 650 of the chassis-less circuit board substrate 602 such that no two of those physical resources are linearly in-line with others along the direction of the airflow path 608. It should be appreciated that, although the optical data connector 834 is in-line with the communication circuit 830, the optical data connector 834 produces no or nominal heat during operation.
The memory devices 720 of the compute sled 800 are mounted to the bottom side 750 of the of the chassis-less circuit board substrate 602 as discussed above in regard to the sled 400. Although mounted to the bottom side 750, the memory devices 720 are communicatively coupled to the processors 820 located on the top side 650 via the I/O subsystem 622. Because the chassis-less circuit board substrate 602 is embodied as a double-sided circuit board, the memory devices 720 and the processors 820 may be communicatively coupled by one or more vias, connectors, or other mechanisms extending through the chassis-less circuit board substrate 602. Of course, each processor 820 may be communicatively coupled to a different set of one or more memory devices 720 in some embodiments. Alternatively, in other embodiments, each processor 820 may be communicatively coupled to each memory device 720. In some embodiments, the memory devices 720 may be mounted to one or more memory mezzanines on the bottom side of the chassis-less circuit board substrate 602 and may interconnect with a corresponding processor 820 through a ball-grid array.
Each of the processors 820 includes a heatsink 850 secured thereto. Due to the mounting of the memory devices 720 to the bottom side 750 of the chassis-less circuit board substrate 602 (as well as the vertical spacing of the sleds 400 in the corresponding rack 240), the top side 650 of the chassis-less circuit board substrate 602 includes additional “free” area or space that facilitates the use of heatsinks 850 having a larger size relative to traditional heatsinks used in typical servers. Additionally, due to the improved thermal cooling characteristics of the chassis-less circuit board substrate 602, none of the processor heatsinks 850 include cooling fans attached thereto. That is, each of the heatsinks 850 is embodied as a fan-less heatsinks.
Referring now to FIG. 10, in some embodiments, the sled 400 may be embodied as an accelerator sled 1000. The accelerator sled 1000 is optimized, or otherwise configured, to perform specialized compute tasks, such as machine learning, encryption, hashing, or other computational-intensive task. In some embodiments, for example, a compute sled 800 may offload tasks to the accelerator sled 1000 during operation. The accelerator sled 1000 includes various components similar to components of the sled 400 and/or compute sled 800, which have been identified in FIG. 10 using the same reference numbers. The description of such components provided above in regard to FIGS. 6, 7, and 8 apply to the corresponding components of the accelerator sled 1000 and is not repeated herein for clarity of the description of the accelerator sled 1000.
In the illustrative accelerator sled 1000, the physical resources 620 are embodied as accelerator circuits 1020. Although only two accelerator circuits 1020 are shown in FIG. 10, it should be appreciated that the accelerator sled 1000 may include additional accelerator circuits 1020 in other embodiments. For example, as shown in FIG. 11, the accelerator sled 1000 may include four accelerator circuits 1020 in some embodiments. The accelerator circuits 1020 may be embodied as any type of processor, co-processor, compute circuit, or other device capable of performing compute or processing operations. For example, the accelerator circuits 1020 may be embodied as, for example, field programmable gate arrays (FPGA), application-specific integrated circuits (ASICs), security co-processors, graphics processing units (GPUs), machine learning circuits, or other specialized processors, controllers, devices, and/or circuits.
In some embodiments, the accelerator sled 1000 may also include an accelerator-to-accelerator interconnect 1042. Similar to the resource-to-resource interconnect 624 of the sled 600 discussed above, the accelerator-to-accelerator interconnect 1042 may be embodied as any type of communication interconnect capable of facilitating accelerator-to-accelerator communications. In the illustrative embodiment, the accelerator-to-accelerator interconnect 1042 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the accelerator-to-accelerator interconnect 1042 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications. In some embodiments, the accelerator circuits 1020 may be daisy-chained with a primary accelerator circuit 1020 connected to the NIC 832 and memory 720 through the I/O subsystem 622 and a secondary accelerator circuit 1020 connected to the NIC 832 and memory 720 through a primary accelerator circuit 1020.
Referring now to FIG. 11, an illustrative embodiment of the accelerator sled 1000 is shown. As discussed above, the accelerator circuits 1020, communication circuit 830, and optical data connector 834 are mounted to the top side 650 of the chassis-less circuit board substrate 602. Again, the individual accelerator circuits 1020 and communication circuit 830 are mounted to the top side 650 of the chassis-less circuit board substrate 602 such that no two heat-producing, electrical components shadow each other as discussed above. The memory devices 720 of the accelerator sled 1000 are mounted to the bottom side 750 of the of the chassis-less circuit board substrate 602 as discussed above in regard to the sled 600. Although mounted to the bottom side 750, the memory devices 720 are communicatively coupled to the accelerator circuits 1020 located on the top side 650 via the I/O subsystem 622 (e.g., through vias). Further, each of the accelerator circuits 1020 may include a heatsink 1070 that is larger than a traditional heatsink used in a server. As discussed above with reference to the heatsinks 870, the heatsinks 1070 may be larger than tradition heatsinks because of the “free” area provided by the memory devices 750 being located on the bottom side 750 of the chassis-less circuit board substrate 602 rather than on the top side 650.
Referring now to FIG. 12, in some embodiments, the sled 400 may be embodied as a storage sled 1200. The storage sled 1200 is optimized, or otherwise configured, to store data in a data storage 1250 local to the storage sled 1200. For example, during operation, a compute sled 800 or an accelerator sled 1000 may store and retrieve data from the data storage 1250 of the storage sled 1200. The storage sled 1200 includes various components similar to components of the sled 400 and/or the compute sled 800, which have been identified in FIG. 12 using the same reference numbers. The description of such components provided above in regard to FIGS. 6, 7, and 8 apply to the corresponding components of the storage sled 1200 and is not repeated herein for clarity of the description of the storage sled 1200.
In the illustrative storage sled 1200, the physical resources 620 are embodied as storage controllers 1220. Although only two storage controllers 1220 are shown in FIG. 12, it should be appreciated that the storage sled 1200 may include additional storage controllers 1220 in other embodiments. The storage controllers 1220 may be embodied as any type of processor, controller, or control circuit capable of controlling the storage and retrieval of data into the data storage 1250 based on requests received via the communication circuit 830. In the illustrative embodiment, the storage controllers 1220 are embodied as relatively low-power processors or controllers. For example, in some embodiments, the storage controllers 1220 may be configured to operate at a power rating of about 75 watts.
In some embodiments, the storage sled 1200 may also include a controller-to-controller interconnect 1242. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the controller-to-controller interconnect 1242 may be embodied as any type of communication interconnect capable of facilitating controller-to-controller communications. In the illustrative embodiment, the controller-to-controller interconnect 1242 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the controller-to-controller interconnect 1242 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications.
Referring now to FIG. 13, an illustrative embodiment of the storage sled 1200 is shown. In the illustrative embodiment, the data storage 1250 is embodied as, or otherwise includes, a storage cage 1252 configured to house one or more solid state drives (SSDs) 1254. To do so, the storage cage 1252 includes a number of mounting slots 1256, each of which is configured to receive a corresponding solid state drive 1254. Each of the mounting slots 1256 includes a number of drive guides 1258 that cooperate to define an access opening 1260 of the corresponding mounting slot 1256. The storage cage 1252 is secured to the chassis-less circuit board substrate 602 such that the access openings face away from (i.e., toward the front of) the chassis-less circuit board substrate 602. As such, solid state drives 1254 are accessible while the storage sled 1200 is mounted in a corresponding rack 204. For example, a solid state drive 1254 may be swapped out of a rack 240 (e.g., via a robot) while the storage sled 1200 remains mounted in the corresponding rack 240.
The storage cage 1252 illustratively includes sixteen mounting slots 1256 and is capable of mounting and storing sixteen solid state drives 1254. Of course, the storage cage 1252 may be configured to store additional or fewer solid state drives 1254 in other embodiments. Additionally, in the illustrative embodiment, the solid state drivers are mounted vertically in the storage cage 1252, but may be mounted in the storage cage 1252 in a different orientation in other embodiments. Each solid state drive 1254 may be embodied as any type of data storage device capable of storing long term data. To do so, the solid state drives 1254 may include volatile and non-volatile memory devices discussed above.
As shown in FIG. 13, the storage controllers 1220, the communication circuit 830, and the optical data connector 834 are illustratively mounted to the top side 650 of the chassis-less circuit board substrate 602. Again, as discussed above, any suitable attachment or mounting technology may be used to mount the electrical components of the storage sled 1200 to the chassis-less circuit board substrate 602 including, for example, sockets (e.g., a processor socket), holders, brackets, soldered connections, and/or other mounting or securing techniques.
As discussed above, the individual storage controllers 1220 and the communication circuit 830 are mounted to the top side 650 of the chassis-less circuit board substrate 602 such that no two heat-producing, electrical components shadow each other. For example, the storage controllers 1220 and the communication circuit 830 are mounted in corresponding locations on the top side 650 of the chassis-less circuit board substrate 602 such that no two of those electrical components are linearly in-line with other along the direction of the airflow path 608.
The memory devices 720 of the storage sled 1200 are mounted to the bottom side 750 of the of the chassis-less circuit board substrate 602 as discussed above in regard to the sled 400. Although mounted to the bottom side 750, the memory devices 720 are communicatively coupled to the storage controllers 1220 located on the top side 650 via the I/O subsystem 622. Again, because the chassis-less circuit board substrate 602 is embodied as a double-sided circuit board, the memory devices 720 and the storage controllers 1220 may be communicatively coupled by one or more vias, connectors, or other mechanisms extending through the chassis-less circuit board substrate 602. Each of the storage controllers 1220 includes a heatsink 1270 secured thereto. As discussed above, due to the improved thermal cooling characteristics of the chassis-less circuit board substrate 602 of the storage sled 1200, none of the heatsinks 1270 include cooling fans attached thereto. That is, each of the heatsinks 1270 is embodied as a fan-less heatsink.
Referring now to FIG. 14, in some embodiments, the sled 400 may be embodied as a memory sled 1400. The storage sled 1400 is optimized, or otherwise configured, to provide other sleds 400 (e.g., compute sleds 800, accelerator sleds 1000, etc.) with access to a pool of memory (e.g., in two or more sets 1430, 1432 of memory devices 720) local to the memory sled 1200. For example, during operation, a compute sled 800 or an accelerator sled 1000 may remotely write to and/or read from one or more of the memory sets 1430, 1432 of the memory sled 1200 using a logical address space that maps to physical addresses in the memory sets 1430, 1432. The memory sled 1400 includes various components similar to components of the sled 400 and/or the compute sled 800, which have been identified in FIG. 14 using the same reference numbers. The description of such components provided above in regard to FIGS. 6, 7, and 8 apply to the corresponding components of the memory sled 1400 and is not repeated herein for clarity of the description of the memory sled 1400.
In the illustrative memory sled 1400, the physical resources 620 are embodied as memory controllers 1420. Although only two memory controllers 1420 are shown in FIG. 14, it should be appreciated that the memory sled 1400 may include additional memory controllers 1420 in other embodiments. The memory controllers 1420 may be embodied as any type of processor, controller, or control circuit capable of controlling the writing and reading of data into the memory sets 1430, 1432 based on requests received via the communication circuit 830. In the illustrative embodiment, each storage controller 1220 is connected to a corresponding memory set 1430, 1432 to write to and read from memory devices 720 within the corresponding memory set 1430, 1432 and enforce any permissions (e.g., read, write, etc.) associated with sled 400 that has sent a request to the memory sled 1400 to perform a memory access operation (e.g., read or write).
In some embodiments, the memory sled 1400 may also include a controller-to-controller interconnect 1442. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the controller-to-controller interconnect 1442 may be embodied as any type of communication interconnect capable of facilitating controller-to-controller communications. In the illustrative embodiment, the controller-to-controller interconnect 1442 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the controller-to-controller interconnect 1442 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications. As such, in some embodiments, a memory controller 1420 may access, through the controller-to-controller interconnect 1442, memory that is within the memory set 1432 associated with another memory controller 1420. In some embodiments, a scalable memory controller is made of multiple smaller memory controllers, referred to herein as “chiplets”, on a memory sled (e.g., the memory sled 1400). The chiplets may be interconnected (e.g., using EMIB (Embedded Multi-Die Interconnect Bridge)). The combined chiplet memory controller may scale up to a relatively large number of memory controllers and I/O ports, (e.g., up to 16 memory channels). In some embodiments, the memory controllers 1420 may implement a memory interleave (e.g., one memory address is mapped to the memory set 1430, the next memory address is mapped to the memory set 1432, and the third address is mapped to the memory set 1430, etc.). The interleaving may be managed within the memory controllers 1420, or from CPU sockets (e.g., of the compute sled 800) across network links to the memory sets 1430, 1432, and may improve the latency associated with performing memory access operations as compared to accessing contiguous memory addresses from the same memory device.
Further, in some embodiments, the memory sled 1400 may be connected to one or more other sleds 400 (e.g., in the same rack 240 or an adjacent rack 240) through a waveguide, using the waveguide connector 1480. In the illustrative embodiment, the waveguides are 64 millimeter waveguides that provide 16 Rx (i.e., receive) lanes and 16 Rt (i.e., transmit) lanes. Each lane, in the illustrative embodiment, is either 16 Ghz or 32 Ghz. In other embodiments, the frequencies may be different. Using a waveguide may provide high throughput access to the memory pool (e.g., the memory sets 1430, 1432) to another sled (e.g., a sled 400 in the same rack 240 or an adjacent rack 240 as the memory sled 1400) without adding to the load on the optical data connector 834.
Referring now to FIG. 15, a system for executing one or more workloads (e.g., applications) may be implemented in accordance with the data center 100. In the illustrative embodiment, the system 1510 includes an orchestrator server 1520, which may be embodied as a managed node comprising a compute device (e.g., a compute sled 800) executing management software (e.g., a cloud operating environment, such as OpenStack) that is communicatively coupled to multiple sleds 400 including a large number of compute sleds 1530 (e.g., each similar to the compute sled 800), memory sleds 1540 (e.g., each similar to the memory sled 1400), accelerator sleds 1550 (e.g., each similar to the memory sled 1000), and storage sleds 1560 (e.g., each similar to the storage sled 1200). One or more of the sleds 1530, 1540, 1550, 1560 may be grouped into a managed node 1570, such as by the orchestrator server 1520, to collectively perform a workload (e.g., an application 1532 executed in a virtual machine or in a container). The managed node 1570 may be embodied as an assembly of physical resources 620, such as processors 820, memory resources 720, accelerator circuits 1020, or data storage 1250, from the same or different sleds 400. Further, the managed node may be established, defined, or “spun up” by the orchestrator server 1520 at the time a workload is to be assigned to the managed node or at any other time, and may exist regardless of whether any workloads are presently assigned to the managed node. In the illustrative embodiment, the orchestrator server 1520 may selectively allocate and/or deallocate physical resources 620 from the sleds 400 and/or add or remove one or more sleds 400 from the managed node 1570 as a function of quality of service (QoS) targets (e.g., performance targets associated with a throughput, latency, instructions per second, etc.) associated with a service level agreement for the workload (e.g., the application 1532). In doing so, the orchestrator server 1520 may receive telemetry data indicative of performance conditions (e.g., throughput, latency, instructions per second, etc.) in each sled 400 of the managed node 1570 and compare the telemetry data to the quality of service targets to determine whether the quality of service targets are being satisfied. If the so, the orchestrator server 1520 may additionally determine whether one or more physical resources may be deallocated from the managed node 1570 while still satisfying the QoS targets, thereby freeing up those physical resources for use in another managed node (e.g., to execute a different workload). Alternatively, if the QoS targets are not presently satisfied, the orchestrator server 1520 may determine to dynamically allocate additional physical resources to assist in the execution of the workload (e.g., the application 1532) while the workload is executing
Additionally, in some embodiments, the orchestrator server 1520 may identify trends in the resource utilization of the workload (e.g., the application 1532), such as by identifying phases of execution (e.g., time periods in which different operations, each having different resource utilizations characteristics, are performed) of the workload (e.g., the application 1532) and pre-emptively identifying available resources in the data center 100 and allocating them to the managed node 1570 (e.g., within a predefined time period of the associated phase beginning). In some embodiments, the orchestrator server 1520 may model performance based on various latencies and a distribution scheme to place workloads among compute sleds and other resources (e.g., accelerator sleds, memory sleds, storage sleds) in the data center 100. For example, the orchestrator server 1520 may utilize a model that accounts for the performance of resources on the sleds 400 (e.g., FPGA performance, memory access latency, etc.) and the performance (e.g., congestion, latency, bandwidth) of the path through the network to the resource (e.g., FPGA). As such, the orchestrator server 1520 may determine which resource(s) should be used with which workloads based on the total latency associated with each potential resource available in the data center 100 (e.g., the latency associated with the performance of the resource itself in addition to the latency associated with the path through the network between the compute sled executing the workload and the sled 400 on which the resource is located).
In some embodiments, the orchestrator server 1520 may generate a map of heat generation in the data center 100 using telemetry data (e.g., temperatures, fan speeds, etc.) reported from the sleds 400 and allocate resources to managed nodes as a function of the map of heat generation and predicted heat generation associated with different workloads, to maintain a target temperature and heat distribution in the data center 100. Additionally or alternatively, in some embodiments, the orchestrator server 1520 may organize received telemetry data into a hierarchical model that is indicative of a relationship between the managed nodes (e.g., a spatial relationship such as the physical locations of the resources of the managed nodes within the data center 100 and/or a functional relationship, such as groupings of the managed nodes by the customers the managed nodes provide services for, the types of functions typically performed by the managed nodes, managed nodes that typically share or exchange workloads among each other, etc.). Based on differences in the physical locations and resources in the managed nodes, a given workload may exhibit different resource utilizations (e.g., cause a different internal temperature, use a different percentage of processor or memory capacity) across the resources of different managed nodes. The orchestrator server 1520 may determine the differences based on the telemetry data stored in the hierarchical model and factor the differences into a prediction of future resource utilization of a workload if the workload is reassigned from one managed node to another managed node, to accurately balance resource utilization in the data center 100.
To reduce the computational load on the orchestrator server 1520 and the data transfer load on the network, in some embodiments, the orchestrator server 1520 may send self-test information to the sleds 400 to enable each sled 400 to locally (e.g., on the sled 400) determine whether telemetry data generated by the sled 400 satisfies one or more conditions (e.g., an available capacity that satisfies a predefined threshold, a temperature that satisfies a predefined threshold, etc.). Each sled 400 may then report back a simplified result (e.g., yes or no) to the orchestrator server 1520, which the orchestrator server 1520 may utilize in determining the allocation of resources to managed nodes.
Referring now to FIG. 16, a system 1610 for dynamically adapting a communication protocol to network communications between endpoints may be implemented in accordance with the data centers 100 described above with reference to FIG. 1. In an example embodiment, the system 1610 includes an orchestrator server 1620 communicatively coupled with multiple sleds, including a compute sled 1630 and accelerator sleds 1640, 1650 and 1660.
The compute sled 1630 and accelerator sleds 1640, 1650 and 1660, or portions thereof, may be grouped into a managed node, such as by the orchestrator server 1620. The managed node may collectively execute a workload, such as an application (e.g., application 1634). A managed node may be embodied as an assembly of resources (e.g., physical resources), such as compute resources, memory resources, storage resources, or other resources, from the same or different sleds or racks. As such, it should be appreciated that a sled may include multiple resources and each resource may be dedicated to a different managed node. Further, a managed node may be established, defined, or “spun up” by the orchestrator server 1620 at the time a workload is to be assigned to the managed node or at any other time, and may exist regardless of whether any workloads are presently assigned to the managed node. The system 1610 may be located in a data center and provide storage and compute services (e.g., cloud services) to a client device 1614 that is in communication with the system 1610 through a network 1612. The orchestrator server 1620 may support a cloud operating environment, such as OpenStack, and managed nodes established by the orchestrator server 1620 may execute one or more applications or processes (i.e., workloads), such as in virtual machines or containers, on behalf of a user of the client device 1614.
Illustratively, the compute sled 1630 includes one or more central processing units (CPUs) 1632 (e.g., a processor or other device or circuitry capable of performing a series of operations) that executes a workload (e.g., application 1634). The accelerator sled 1640 includes an accelerator device 1642, the accelerator sled 1650 includes an accelerator device 1652, and the accelerator sled 1660 includes an accelerator device 1662. Each of the accelerator devices 1642, 1652, or 1662 may be embodied as any device or circuitry usable to accelerate the execution of one or more operations. For example, the accelerator devices described herein may be embodied as any device or circuitry (e.g., a specialized processor, a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a graphics processing unit (GPU), reconfigurable hardware, etc.) capable of accelerating execution of a portion of the workload, such as a workload task (e.g., a set of operations within a workload). Further, each of the accelerator devices are configured with accelerated kernels. Illustratively, accelerator device 1642 includes kernels 1644, accelerator device 1652 includes kernels 1654, and accelerator device 1662 includes kernels 1664. Each of the accelerated kernels may be embodied as a set of code or a configuration of a portion of the corresponding accelerator device that causes the respective accelerator device to perform one or more accelerated functions (e.g., cryptographic operations, compression operations, etc.).
Each of the accelerator sleds 1640, 1650, and 1660 provide accelerated functions as a service for workloads processed by the managed node. In particular, each accelerator sled 1640, 1650, and 1660 may process requests from other sleds within the managed node (e.g., the compute sled 1630) to accelerate a function. For instance, FIG. 16 depicts the compute sled 1630 executing application 1634. The application 1634 may include functions to be performed in sequence. The compute sled 1630 may send a request to the accelerator sleds to accelerate the execution of each function, thereby offloading the execution of the function to an accelerator device residing on the accelerator sled. The accelerator sled may, in response to the request, provision a kernel on the accelerator device. For example, the accelerator sled may load a bit stream indicative of the kernel into a slot (e.g., a subset of circuitry or other logic units) of the accelerator device. The application 1634 may include a variety of functions, such as cryptographic operations, machine learning algorithms, and the like, which may be accelerated. The kernel provisioned on the accelerator device may be suited to accelerate the execution of corresponding functions. For example, assume that the underlying function involves matrix multiplication. The kernel provisioned with the accelerator device may be specific to processing matrix multiplication operations. Once the kernel completes acceleration of the function, the kernel may return resulting data to the compute sled 1630. The orchestrator server 1620 may track (e.g., via a database) which kernels are registered to which accelerator sleds and which accelerator devices.
In addition, the system 1610 may expose a kernel-to-kernel communication network that allows any of the kernels 1644, 1654, and 1664 to communicate with one another, e.g., in sending processed workload data downstream to a kernel that processes a subsequent task in the workload. The kernel may establish a network connection via a given network communication protocol, such as TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol). The kernel may encapsulate the workload data in one or more packets (datagrams in UDP) and transmit the packets (datagrams) to the other kernel using the communication protocol.
Generally, network communication protocols may be characterized as reliable and non-reliable. Protocols characterized as reliable (e.g., TCP/IP) ensure that data transmitted by a sender reaches an intended recipient. Such protocols may notify the sender if the transmission fails (e.g., if a packet is dropped). However, reliable protocols typically incur overhead from determining whether a packet was successfully delivered and returning a notification regarding the delivery. As a result, an operational cost in sending data over TCP/IP involves additional latency. By contrast, protocols characterized as non-reliable (e.g., UDP) do not notify the sender if the transmission fails. However, because non-reliable protocols generally do not have error checking and correction mechanisms (otherwise provided by reliable protocols), such protocols incur less overhead and are thus more scalable than reliable protocols. Reliable protocols are often more desirable in instances where the likelihood of packet loss is relatively high, such as in instances where resource and network utilization in a system (e.g., system 1610) is high. Conversely, non-reliable protocols can be used in instances where the likelihood of packet loss is relatively low, such as in instances where resource and network utilization is low.
As further described herein, embodiments of the present disclosure provide techniques for dynamically shifting between reliable to non-reliable protocols (and vice versa) for network communications (e.g., kernel-to-kernel communications) based on observed telemetry in the system 1610. More specifically, an accelerator device in the system 1610 (e.g., accelerator devices 1642, 1652, or 1662) may include logic to receive (or monitor) telemetry data relating to, in part, network utilization for a kernel-to-kernel link. The telemetry data may include characteristics such as latency in communications between the kernels, throughput, present load on the underlying accelerator device(s), and the like. The accelerator device may evaluate the telemetry data against one or more conditions of a policy to determine whether to shift (e.g., change) a present network communication protocol to another network communication protocol. For example, assume that a kernel A is presently communicating data to a kernel B using the UDP protocol. The accelerator device may observe telemetry data indicative of network utilization between kernel A and kernel B exceeds some threshold, which triggers a condition in the policy. Because reliable protocols may be more suited to situations where network utilization is high, the policy may specify to change from UDP to a reliable protocol, such as TCP/IP.
Further, over time, the accelerator device may learn patterns of telemetry data as a function of time to predict instances for changing from one network communication protocol to another. For example, the accelerator device may perform a variety of machine learning techniques using the observed telemetry and temporal data as input to generate prediction data. The prediction data may be indicative of a likelihood that network communications for a given kernel link should be shifted from one kernel to another, based on subsequently observed telemetry data.
Referring now to FIG. 17, an accelerator sled 1700 may be embodied as any type of compute device capable of performing the functions described herein, including monitoring telemetry data associated with network communications between accelerated kernels, determining as a function of the monitored telemetry data that a condition to shift (e.g., change) the network communications from a given communication protocol to another communication protocol is triggered, and changing the network communications to the other communication protocol. The accelerator sled 1700 may be representative of any of the accelerator sleds 1640, 1650, or 1660 depicted in FIG. 16.
As shown in FIG. 17, the accelerator sled 1700 includes a compute engine 1702, an I/O subsystem 1708, communication circuitry 1710, one or more data storage devices 1714, and one or more accelerator devices 1718. Of course, in other embodiments, the accelerator sled 1700 may include other or additional components, such as those commonly found in a computer (e.g., display, peripheral devices, etc.). Additionally, in some embodiments, one or more of the illustrative components may be incorporated in, or otherwise form a portion of, another component.
The compute engine 1702 may be embodied as any type of device or collection of devices capable of performing various compute functions described below. In some embodiments, the compute engine 1702 may be embodied as a single device such as an integrated circuit, an embedded system, a FPGA, a system-on-a-chip (SoC), or other integrated system or device. Additionally, in some embodiments, the compute engine 1702 includes or is embodied as a processor 1704 and a memory 1706. The processor 1704 may be embodied as any type of processor capable of performing the functions described herein. For example, the processor 1704 may be embodied as a single or multi-core processor(s), a microcontroller, or other processor or processing/controlling circuit. In some embodiments, the processor 1704 may be embodied as, include, or be coupled to an FPGA, an ASIC, reconfigurable hardware or hardware circuitry, or other specialized hardware to facilitate performance of the functions described herein.
The memory 1706 may be embodied as any type of volatile (e.g., dynamic random access memory (DRAM), etc.) or non-volatile memory or data storage capable of performing the functions described herein. Volatile memory may be a storage medium that requires power to maintain the state of data stored by the medium. Non-limiting examples of volatile memory may include various types of random access memory (RAM), such as DRAM or static random access memory (SRAM). One particular type of DRAM that may be used in a memory module is synchronous dynamic random access memory (SDRAM). In particular embodiments, DRAM of a memory component may comply with a standard promulgated by JEDEC, such as JESD79F for DDR SDRAM, JESD79-2F for DDR2 SDRAM, JESD79-3F for DDR3 SDRAM, JESD79-4A for DDR4 SDRAM, JESD209 for Low Power DDR (LPDDR), JESD209-2 for LPDDR2, JESD209-3 for LPDDR3, and JESD209-4 for LPDDR4 (these standards are available at www.jedec.org). Such standards (and similar standards) may be referred to as DDR-based standards and communication interfaces of the storage devices that implement such standards may be referred to as DDR-based interfaces.
In one embodiment, the memory device is a block addressable memory device, such as those based on NAND or NOR technologies. A memory device may also include future generation nonvolatile devices, such as a three dimensional crosspoint memory device (e.g., Intel 3D XPoint™ memory), or other byte addressable write-in-place nonvolatile memory devices. In one embodiment, the memory device may be or may include memory devices that use chalcogenide glass, multi-threshold level NAND flash memory, NOR flash memory, single or multi-level Phase Change Memory (PCM), a resistive memory, nanowire memory, ferroelectric transistor random access memory (FeTRAM), anti-ferroelectric memory, magnetoresistive random access memory (MRAM) memory that incorporates memristor technology, resistive memory including the metal oxide base, the oxygen vacancy base and the conductive bridge Random Access Memory (CB-RAM), or spin transfer torque (STT)-MRAM, a spintronic magnetic junction memory based device, a magnetic tunneling junction (MTJ) based device, a DW (Domain Wall) and SOT (Spin Orbit Transfer) based device, a thyristor based memory device, or a combination of any of the above, or other memory. The memory device may refer to the die itself and/or to a packaged memory product.
In some embodiments, 3D crosspoint memory (e.g., Intel 3D XPoint™ memory) may comprise a transistor-less stackable cross point architecture in which memory cells sit at the intersection of word lines and bit lines and are individually addressable and in which bit storage is based on a change in bulk resistance. In some embodiments, all or a portion of the memory 1706 may be integrated into the processor 1704. In operation, the memory 1706 may store various software and data used during operation.
The compute engine 1702 is communicatively coupled with other components of the accelerator sled 1700 via the I/O subsystem 1708, which may be embodied as circuitry and/or components to facilitate input/output operations with the compute engine 1702 (e.g., with the processor 1704 and/or the memory 1706) and other components of the accelerator sled 1700. For example, the I/O subsystem 1708 may be embodied as, or otherwise include, memory controller hubs, input/output control hubs, integrated sensor hubs, firmware devices, communication links (e.g., point-to-point links, bus links, wires, cables, light guides, printed circuit board traces, etc.), and/or other components and subsystems to facilitate the input/output operations. In some embodiments, the I/O subsystem 1708 may form a portion of a system-on-a-chip (SoC) and be incorporated, along with one or more of the processor 1704, the memory 1706, and other components of the accelerator sled 1700, into the compute engine 1702.
The communication circuitry 1710 may be embodied as any communication circuit, device, or collection thereof, capable of enabling communications over the network 1612 between the accelerator sled 1700 and another compute device (e.g., the compute sled 1630, the accelerator sleds 1640, 1650, and 1660, etc.). The communication circuitry 1710 may be configured to use any one or more communication technology (e.g., wired or wireless communications) and associated protocols (e.g., Ethernet, Bluetooth®, Wi-Fi®, WiMAX, etc.) to effect such communication.
The illustrative communication circuitry 1710 includes a network interface controller (NIC) 1712, which may also be referred to as a host fabric interface (HFI). The NIC 1712 may be embodied as one or more add-in-boards, daughter cards, network interface cards, controller chips, chipsets, or other devices that may be used by the accelerator sled 1700 to connect with another compute device (e.g., the orchestrator server 1620, compute sled 1630, the accelerator sleds 1640, 1650, and 1660, etc.). In some embodiments, the NIC 1712 may be embodied as part of a system-on-a-chip (SoC) that includes one or more processors, or included on a multichip package that also contains one or more processors. In some embodiments, the NIC 1712 may include a local processor (not shown) and/or a local memory (not shown) that are both local to the NIC 1712. In such embodiments, the local processor of the NIC 1712 may be capable of performing one or more of the functions of the compute engine 1702 described herein. Additionally or alternatively, in such embodiments, the local memory of the NIC 1712 may be integrated into one or more components of the accelerator sled 1700 at the board level, socket level, chip level, and/or other levels.
The one or more illustrative data storage devices 1714 may be embodied as any type of devices configured for short-term or long-term storage of data such as, for example, memory devices and circuits, memory cards, hard disk drives (HDDs), solid-state drives (SSDs), or other data storage devices. Each data storage device 1714 may include a system partition that stores data and firmware code for the data storage device 1714. Each data storage device 1714 may also include an operating system partition that stores data files and executables for an operating system.
The accelerator devices 1718 can be representative of accelerator devices in the system 1610 depicted in FIG. 16, such as any combination of accelerator devices 1642, 1652, or 1662. The accelerator devices 1718 may form an accelerator subsystem that includes one or more buses or other interfaces between the accelerator devices in the accelerator sled 1800 to enable the accelerator devices to share data. Further, each accelerator device 1718 may send data via the NIC 1712 to other accelerator devices in the system 1610, based on a kernel configuration defined by the orchestrator server 1620. Each accelerator device 1718 may be embodied as any device or circuitry (e.g., a specialized processor, an FPGA, an ASIC, a GPU, reconfigurable hardware, etc.) capable of accelerating the execution of a function.
Additionally or alternatively, the accelerator sled 1700 may include one or more peripheral devices 1716. Such peripheral devices 1716 may include any type of peripheral device commonly found in a compute device such as a display, speakers, a mouse, a keyboard, and/or other input/output devices, interface devices, and/or other peripheral devices.
The orchestrator server 1620, client device 1614, and compute sled 1630 may have components similar to those described relative to FIG. 17. The description of those components of the accelerator sled 1700 is equally applicable to the description of components of those devices and is not repeated herein for clarity of the description. Further, it should be appreciated that any of the client device 1614, the orchestrator server 1620, and the sleds 1630, 1640, 1650, an 1660 may include other components, sub-components, and devices commonly found in a computing device, which are not discussed above in reference to the accelerator sled 1700 and not discussed herein for clarity of the description.
As described above, the client device 1614, the orchestrator server 1620, and the sleds 1630, 1640, 1650, and 1660 are illustratively in communication via the network 1612, which may be embodied as any type of wired or wireless communication network, including global networks (e.g., the Internet), local area networks (LANs) or wide area networks (WANs), cellular networks (e.g., Global System for Mobile Communications (GSM), 3G, Long Term Evolution (LTE), Worldwide Interoperability for Microwave Access (WiMAX), etc.), digital subscriber line (DSL) networks, cable networks (e.g., coaxial networks, fiber networks, etc.), or any combination thereof.
Referring now to FIG. 18, the accelerator sled 1700 may establish an environment 1800 during operation. Of course, any of the accelerator sleds 1640, 1650, 1660 may similarly establish the environment 1800 during operation. Illustratively, the environment 1800 includes a network communicator 1820 and a protocol manager 1830. Each of the components of the environment 1800 may be embodied as hardware, firmware, software, or a combination thereof. As such, in some embodiments, one or more of the components of the environment 1800 may be embodied as circuitry or a collection of electrical devices (e.g., network communicator circuitry 1820, protocol manager circuitry 1830, etc.). It should be appreciated that, in such embodiments, one or more of the network communicator circuitry 1820 or protocol manager circuitry 1830 may form a portion of one or more of the compute engine 1702, the communication circuitry 1710, the I/O subsystem 1708, accelerator devices 1718 and/or other components of the accelerator sled 1700. As shown, the environment 1800 includes a kernel configuration data 1802, which may be embodied as any data indicative of mappings of kernel configurations in the system 1610 relative to a flow of a workload. The kernel configuration data 1802 may also be indicative of network communication protocols used by kernel-to-kernel links in the system 1610. Further, the environment 1800 includes policy data 1804, which may be embodied as any data indicative of protocol change policies including one or more change conditions evaluated as a function of monitored resource and network utilization. Further, the environment 1800 includes telemetry data 1806, which may be embodied as any data indicative of observed performance of the accelerator sled 1700, accelerator devices 1718, and other accelerator sleds and devices of the system 1610 (e.g., power consumption, amount of kernel connections, latency, average period of connections, amount of data transferred per connection, etc.). Further still, the environment 1800 includes prediction data 1808, which may be embodied as any data indicative of a likelihood that a protocol change policy condition is triggered based on subsequently observed telemetry data 1806.
The network communicator 1820, which may be embodied as hardware, firmware, software, virtualized hardware, emulated architecture, and/or a combination thereof as discussed above, is configured to facilitate inbound and outbound network communications (e.g., network traffic, network packets, network flows, etc.) to and from the accelerator sled 1700, respectively. To do so, the network communicator 1820 is configured to receive and process data packets from one system or computing device (e.g., accelerator sleds 1640, 1650, or 1660) and to prepare and send data packets to another computing device or system (e.g., the compute sled 1630, or other accelerator sleds 1640, 1650, or 1660). Accordingly, in some embodiments, at least a portion of the functionality of the network communicator 1820 may be performed by the communication circuitry 1710, and, in the illustrative embodiment, by the NIC 1712.
The protocol manager 1830, which may be embodied as hardware, firmware, software, virtualized hardware, emulated architecture, and/or a combination thereof, is configured to monitor telemetry data 1806 associated with one or more network communications between a kernel and another kernel, where the network communications are established via a given communication protocol. The protocol manager 1830 is also configured to determine, as a function of the monitored telemetry data 1806, that a condition in the policy data 1804 to change the network communications from the present communication protocol to another communication protocol is triggered. The protocol manager 1830 is also configured to change the network communications between the kernels to the other communication protocol. As shown, the protocol manager 1830 includes a monitor component 1832, selector component 1434, configuration component 1836, and a predictor component 1838.
In some embodiments, the monitor component 1832 is configured to obtain telemetry data 1806 associated with a network channel between a given kernel and another kernel. More specifically, a kernel that is configured within a given accelerator device 1818 may be interconnected with another kernel. The other kernel may be configured with the same accelerator device 1818 or another accelerator device in the system 1610. The kernels may be interconnected through a variety of approaches. For example, assume that the other kernel is configured on an accelerator device on the sled 1700. The kernels may communicate with one another via the NIC 1712. As another example, the other kernel might be configured on another accelerator sled in the system 1610. In such a case, the kernels may be interconnected via a switch device in the system 1610 interconnecting that sled with the sled 1300. The monitor component 1832 may obtain telemetry from the NIC 1712 or the switch devices in the system 1610 relating to the link between the kernels. For example, a resource monitor can reside in the NIC 1712 or switch devices and obtain raw metrics regarding performance and utilization and send the metrics to the monitor component 1832. In turn, the monitor component 1832 receives the raw metrics and may normalize the metrics to generate the telemetry data 1806. Normalizing the metrics may involve converting the raw metrics to a value and type that can be further evaluated by the protocol manager 1830. In other embodiments, the monitor component 1832 is also configured to obtain telemetry data 1806 associated with the system 1610, such as a present load on the system 1610, average network utilization between kernel connections, average packet loss in kernel connections, and the like.
In some embodiments, the selector component 1834 is configured to determine, based on an evaluation of the telemetry data 1806, whether to shift (e.g., change) network communications between a kernel configured with an accelerator device 1818 and another kernel to a different protocol. For example, assume network communications between a kernel A and a kernel B are currently performed via a non-reliable protocol, such as UDP. The selector component 1834 may evaluate the telemetry data 1806 relative to policy data 1804 to determine whether one or more conditions for changing to a reliable protocol, such as TCP/IP, is triggered. For example, the policy data 1804 may specify a condition that if network utilization between kernels exceeds a specified threshold, then communications between the kernels should be performed via a reliable protocol to ensure that kernel B receives data regardless of any additional latency resulting from usage of the reliable protocol.
In some embodiments, the configuration component 1836 is configured to change network communications between the kernels to a different protocol if so determined by the selector component 1834. The configuration component 1836 may modify the kernel configuration data 1802 to indicate that, for a link between a given kernel on the accelerator device 1818 and another kernel is to be carried out using the different protocol, as determined by the selector component 1834. The configuration component 1836 may also notify the orchestrator server 1620 of the change in protocol. To do so, the configuration component 1836 may send a message to the orchestrator server 1620 identifying the kernels, the accelerator devices on which the kernels are configured, and the protocol. In response, the orchestrator server 1620 may propagate the update to the other sleds and network devices in the system 1610.
In some embodiments the predictor component 1838 is configured to learn one or more patterns based on the telemetry data 1806 and changing between protocols over time to sooner identify instances where a changing between a given protocol to another would be performed. For example, the predictor component 1838 may perform a variety of machine learning algorithms (e.g., optimization-based machine learning algorithms, prediction algorithms, etc.) and receive, as input, telemetry data 1806 relating to a given kernel-to-kernel network link and also to the system 1610. The predictor component 1838 may also receive timestamp input defining instances and periods where the network link changes from one communication protocol to another. The machine learning algorithm may generate prediction data 1808 as a result. The selector component 1834 may be further configured to evaluate the prediction data 1808 in determining whether to change to a different network communication protocol for a given kernel-to-kernel link For example, the selector component 1834 may retrieve subsequently collected telemetry data 1806 from the monitor component 1832. The selector component 1834 may input the telemetry data 1806 to a machine learning algorithm, which may evaluate the telemetry data 1807 against the prediction data 1808. The result may indicate whether to change the presently configured network communication protocol to another network communication protocol.
Further, the predictor component 1838 may include a variety of prediction algorithms and provide a ranking of the algorithms for selection based on an execution of the telemetry data 1806 on each of the algorithms. For example, the ranking may be based on a percentage that the result of each algorithm converges towards a likely optimal result. Once a selection is provided, the predictor component 1838 may continue to use the selected algorithm in subsequent calculations.
Referring now to FIGS. 19A and 19B, diagrams of an example embodiment of a kernel-to-kernel communication network is shown. In particular, FIGS. 19A and 19B illustrate a kernel-to-kernel communication network in which techniques for dynamically adapting reliable and non-reliable network communication protocols can be implemented. FIG. 19A depicts intercommunications between various kernels in the system 1610. In particular, FIG. 19A includes an accelerator sled 1902 and 1912, which are representative of any of the accelerator sleds 1640, 1650, 1660, or 1700 in the system 1610. The accelerator sled 1902 provides an accelerator device 1903 and a 1907, and the accelerator sled 1912 provides an accelerator device 1913. Further, each of the accelerator devices includes one or more slots for loading one or more accelerated kernels. For example, the accelerator device 1903 includes a slot 1904, the accelerator device 1907 includes a slot 1908, and the accelerator device includes a slot A 1914 and a slot B 1917. Further still, each of the slots is configured with accelerated kernels. Illustratively, slot 1904 is configured with a kernel A 1905 and a kernel B 1906; the slot 1908 is configured with a kernel C 1909 and a kernel D 1910; the slot A 1914 is configured with a kernel E 1915 and a kernel F 1916; and the slot B is configured with a kernel G and a kernel H 1919.
As indicated by the two-way arrows, each of the kernels may communicate with one another via a kernel-to-kernel network exposed by the system 1610. In particular, the system 1610 may provide a kernel network configuration that includes NICs in each of the accelerator sleds 1902 and 1912 as well as network devices interconnecting the accelerator sleds 1902 and 1912 with one another (e.g., a network switch). The orchestrator server 1620 may configure each of the NICs and network devices such that a given kernel sends data to another kernel as a function of a workload flow. The NICs and network devices may form an accelerator subsystem interface that connects components of an accelerator device, including kernels, with one another to form a kernel-to-kernel network. The accelerator subsystem interface may expose a virtual address space that allows kernels to identify and communicate with one another in the network.
FIG. 19B illustrates an abstraction of communication between the kernel A 1905 and kernels B 1906, C 1909, and H 1919. The kernel A 1905 may communicate over a network 1920 (e.g., the accelerator subsystem described above). For example, the kernels B 1906, C 1909, and H 1919 may represent kernels to which the kernel A 1905 transmits processed data downstream as a part of a workload execution. Illustratively, the kernel A 1905 may be interconnected with each of the kernels via TCP/IP or UDP. The network communication protocol used for a given kernel-to-kernel link may be determined as a function of resource and network utilization associated with the link. For example, kernel A 1905 and kernel B 1906 are interconnected via UDP. Using UDP may be indicative of a network link that has relatively low utilization. Under such utilization, packets are less likely to be dropped, and thus packet loss detection mechanisms are less necessary. By contrast, kernel A is interconnected with kernels C 1909 and kernel H 1919 via TCP/IP. Using TCP/IP may be indicative of a network link that has a relatively high network utilization. Under such utilization, packets are more likely to be dropped, and thus the packet loss detection mechanisms provided TCP/IP might be more desirable.
Referring now to FIG. 20, the accelerator sled 1700, in operation, may execute a method 2000 to adapt a communication protocol to network connections between endpoints (e.g., kernel endpoints). Of course, the accelerator sleds 1640, 1650, and 1660 may also perform the method 2000. As shown, the method 2000 begins in block 2002, in which the accelerator sled 1700 establishes network communications between a kernel configured thereon (e.g., in an accelerator device 1718) and another kernel in the system 1610 via a given network communication protocol. For example, the accelerator sled 1700 may do so by evaluating a kernel configuration (e.g., kernel configuration data 1802) to determine a presently specified network communication protocol to use for the network link between the kernel and the other kernel. The kernel configuration data 1802 may specify whether the present protocol is a reliable protocol (e.g., TCP/IP) or a non-reliable protocol (e.g., UDP).
In block 2004, the accelerator sled 1700 monitors telemetry data associated with the established communications. For instance, the accelerator sled 1700 may collect raw metrics from the NIC 1712 and other network components that interconnect the kernels with one another. Once collected, the accelerator sled 1700 may further process the metrics for evaluation. In block 2006, the accelerator sled 1700 determines whether a change condition is triggered. To do so, the accelerator sled 1700 may evaluate the telemetry data relative to policy data for changing between a given network communication protocol and another. For example, a change condition specified in the policy may indicate to change from UDP to TCP/IP if observed network utilization in the telemetry data exceeds a predefined threshold. If a condition in the policy is not triggered, then the method 2000 returns to block 2004, and the accelerator sled 1700 continues to monitor telemetry data.
Otherwise, if a change condition in the policy is triggered, then, in block 2008, the accelerator sled 1700 changes the network communication protocol used in communications between the kernels. In particular, in block 2010, the accelerator sled 1700 evaluates the monitored telemetry data and the presently used protocol for the kernel-to-kernel link relative to the policy. In block 2012, the accelerator sled 1700 determines, as a function of the policy, whether to change to a reliable protocol (e.g., TCP/IP) or a non-reliable protocol (UDP). For example, the policy may specify that if the presently used protocol is a non-reliable protocol and network bandwidth exceeds a specified threshold for a specified duration, then change the network communication protocol to a reliable protocol. As another example, the policy may specify that if the presently used protocol is a reliable protocol and an average packet loss falls below a specified threshold for a specified duration, then change the network communication protocol to a non-reliable protocol.
In block 2014, the accelerator sled 1700 modifies a configuration of the kernel-to-kernel link based on the determination. For example, the accelerator sled 1700 may do so by accessing a locally stored configuration, e.g., the kernel configuration data 1802, and modifying the configuration to indicate the protocol for the kernel link. Further, the accelerator sled 1700 may notify the orchestrator server 1620 to the change in communication protocol for the kernel link As a result, the orchestrator server 1620 may propagate the change to configurations of other accelerator sleds in the system 1610 to preserve integrity. In block 2016, the accelerator sled 1700 establishes subsequent network communications between the kernels using the protocol determined based on the policy.
Referring now to FIG. 21, the accelerator sled 1700, in operation, may apply machine learning techniques to determine whether to change between a non-reliable protocol and a reliable protocol (and vice versa). In block 2018, the accelerator sled 1700 learns one or more change patterns based on the monitored telemetry data. For instance, in block 2020, the accelerator sled 1700 evaluates the monitored telemetry data associated with a kernel-to-kernel link relative to a given point in time (e.g., as indicated by a time stamp) that the accelerator sled 1700 changes from a communication protocol to another. In block 2022, the accelerator sled 1700 identifies the patterns as a function of the evaluated telemetry data and time. For example, the accelerator sled 1700 may identify a tuple of telemetry values at a previous point in time in which the accelerator sled 1700 changes to another network communication protocol. The accelerator sled 1700 may identify additional points in time where the tuple of telemetry values triggers the change to the other protocol. The identified points in time may be indicative of a pattern in correlating to a change to the other protocol.
In block 2024, the accelerator sled 1700 generates prediction data based on the one or more learned patterns. The prediction data indicates a likelihood that a policy condition to change from one protocol to another protocol is triggered based on subsequently observed telemetry data 1806. The prediction data may reduce the amount of telemetry data actually observed before changing to another protocol, and thus improve network utilization. In block 2026, the accelerator sled 1700 uses the prediction data to determine subsequent changes from a network communication protocol to another protocol. For instance, the accelerator sled 1700 may return to the beginning of method 2000 and, in addition to evaluating subsequently monitored telemetry data relative to a policy, the accelerator sled 1700 may further evaluate the telemetry data relative to the prediction data. For example, the accelerator sled 1700 may observe a given tuple of telemetry data at a given point of time in the execution of a workload that would not otherwise trigger a change condition. However, the accelerator sled 1700, may, after an evaluation against prediction, identify the tuple as the beginning of a pattern leading to a change between protocols. Once identified, the accelerator sled 1700 may pre-emptively change the protocol.

EXAMPLES

Illustrative examples of the technologies disclosed herein are provided below. An embodiment of the technologies may include any one or more, and any combination of, the examples described below.
Example 1 includes a sled comprising a compute engine to monitor telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications is established via a first communication protocol, determine, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and change the network communications from the first communication protocol to the second communication protocol.
Example 2 includes the subject matter of Example 1, and wherein to change the network communications from the first communication protocol to the second communication protocol comprises to establish subsequent network communications between the first kernel and the second kernel using the second communication protocol.
Example 3 includes the subject matter of any of Examples 1 and 2, and wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.
Example 4 includes the subject matter of any of Examples 1-3, and wherein the second communication protocol corresponds to one of TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol).
Example 5 includes the subject matter of any of Examples 1-4, and wherein the compute engine is further to learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time.
Example 6 includes the subject matter of any of Examples 1-5, and wherein the compute engine is further to generate prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.
Example 7 includes the subject matter of any of Examples 1-6, and wherein to determine that the condition to change the network communications is triggered is further determined as a function of the prediction data.
Example 8 includes the subject matter of any of Examples 1-7, and wherein to learn one or more change patterns from the monitored telemetry data comprises to evaluate the telemetry data associated with the kernel network connections over time; and identify the change patterns based on the evaluation.
Example 9 includes the subject matter of any of Examples 1-8, and wherein the compute engine is further to generate the prediction data via a machine learning technique.
Example 10 includes the subject matter of any of Examples 1-9, and wherein the compute engine is further to generate the prediction data via one or a plurality of machine learning techniques; and rank the prediction data according to each of the plurality of machine learning techniques.
Example 11 includes the subject matter of any of Examples 1-10, and wherein the telemetry data includes at least one of a packet loss rate, a total amount of the network connections, throughput of the network connections, and latency of the network connections.
Example 12 includes the subject matter of any of Examples 1-11, and wherein the compute engine is further to monitor telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol.
Example 13 includes the subject matter of any of Examples 1-12, and wherein the compute engine is further to change, as a function of the monitored telemetry data, the network communications between the first kernel and the third kernel to the first communication protocol.
Example 14 includes the subject matter of any of Examples 1-13, and wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel exceeds a specified threshold.
Example 15 includes a method comprising monitoring telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications is established via a first communication protocol, determining, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and changing the network communications from the first communication protocol to the second communication protocol.
Example 16 includes the subject matter of Example 15, and wherein changing the network communications from the first communication protocol to the second communication protocol comprises establishing subsequent network communications between the first kernel and the second kernel using the second communication protocol.
Example 17 includes the subject matter of any of Examples 15 and 16, and wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.
Example 18 includes the subject matter of any of Examples 15-17, and wherein the second communication protocol corresponds to one of TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol).
Example 19 includes the subject matter of any of Examples 15-18, and further including learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time.
Example 20 includes the subject matter of any of Examples 15-19, and further including generating prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.
Example 21 includes the subject matter of any of Examples 15-20, and wherein determining that the condition to change the network communications is triggered is further determined as a function of the prediction data.
Example 22 includes the subject matter of any of Examples 15-21, and wherein learning one or more change patterns from the monitored telemetry data comprises evaluating the telemetry data associated with the kernel network connections over time; and identifying the patterns based on the evaluation.
Example 23 includes the subject matter of any of Examples 15-22, and further including generating the prediction data via a machine learning technique.
Example 24 includes the subject matter of any of Examples 15-23, and further including generating the prediction data via one or a plurality of machine learning techniques; and ranking the prediction data according to each of the plurality of machine learning techniques.
Example 25 includes the subject matter of any of Examples 15-24, and wherein the telemetry data includes at least one of a packet loss rate, a total amount of the network connections, throughput of the network connections, and latency of the network connections.
Example 26 includes the subject matter of any of Examples 15-25, and further including monitoring telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol.
Example 27 includes the subject matter of any of Examples 15-26, and further including changing, as a function of the monitored telemetry data, the network communications between the first kernel and the third kernel to the first communication protocol.
Example 28 includes one or more machine-readable storage media comprising a plurality of instructions stored thereon that, in response to being executed, cause a sled to perform the method of any of Examples 15-27.
Example 29 includes a sled comprising means for performing the method of any of Examples 15-27.
Example 30 includes a sled comprising a compute engine to perform the method of any of Examples 15-27.
Example 31 includes a sled, comprising protocol manager circuitry to monitor telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications is established via a first communication protocol, determine, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and change the network communications from the first communication protocol to the second communication protocol.
Example 32 includes the subject matter of Example 31, and wherein to change the network communications from the first communication protocol to the second communication protocol comprises to establish subsequent network communications between the first kernel and the second kernel using the second communication protocol.
Example 33 includes the subject matter of any of Examples 31 and 32, and wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.
Example 34 includes the subject matter of any of Examples 31-33, and wherein the second communication protocol corresponds to one of TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol).
Example 35 includes the subject matter of any of Examples 31-34, and wherein the protocol manager circuitry is further to learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time.
Example 36 includes the subject matter of any of Examples 31-35, and wherein the protocol manager circuitry is further to generate prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.
Example 37 includes the subject matter of any of Examples 31-36, and wherein to determine that the condition to change the network communications is triggered is further determined as a function of the prediction data.
Example 38 includes the subject matter of any of Examples 31-37, and wherein to learn one or more change patterns from the monitored telemetry data comprises to evaluate the telemetry data associated with the kernel network connections over time; and identify the change patterns based on the evaluation.
Example 39 includes the subject matter of any of Examples 31-38, and wherein the protocol manager circuitry is further to generate the prediction data via a machine learning technique.
Example 40 includes the subject matter of any of Examples 31-39, and wherein the protocol manager circuitry is further to generate the prediction data via one or a plurality of machine learning techniques; and rank the prediction data according to each of the plurality of machine learning techniques.
Example 41 includes the subject matter of any of Examples 31-40, and wherein the telemetry data includes at least one of a packet loss rate, a total amount of the network connections, throughput of the network connections, and latency of the network connections.
Example 42 includes the subject matter of any of Examples 31-41, and wherein the protocol manager circuitry is further to monitor telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol.
Example 43 includes the subject matter of any of Examples 31-42, and wherein the protocol manager circuitry is further to change, as a function of the monitored telemetry data, the network communications between the first kernel and the third kernel to the first communication protocol.
Example 44 includes the subject matter of any of Examples 31-43, and wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel exceeds a specified threshold.
Example 45 includes a sled, comprising circuitry for monitoring telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications is established via a first communication protocol, means for determining, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and means for changing the network communications from the first communication protocol to the second communication protocol.
Example 46 includes the subject matter of Example 45, and wherein the means for changing the network communications from the first communication protocol to the second communication protocol comprises circuitry for establishing subsequent network communications between the first kernel and the second kernel using the second communication protocol.
Example 47 includes the subject matter of any of Examples 45 and 46, and wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.
Example 48 includes the subject matter of any of Examples 45-47, and wherein the second communication protocol corresponds to one of TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol).
Example 49 includes the subject matter of any of Examples 45-48, and further including means for learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time.
Example 50 includes the subject matter of any of Examples 45-49, and further including means for generating prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.
Example 51 includes the subject matter of any of Examples 45-50, and wherein the means for determining that the condition to change the network communications is triggered is further determined as a function of the prediction data.
Example 52 includes the subject matter of any of Examples 45-51, and wherein the means for learning one or more change patterns from the monitored telemetry data comprises circuitry for evaluating the telemetry data associated with the kernel network connections over time; and circuitry for identifying the patterns based on the evaluation.
Example 53 includes the subject matter of any of Examples 45-52, and further including means for generating the prediction data via a machine learning technique.
Example 54 includes the subject matter of any of Examples 45-53, and further including means for generating the prediction data via one or a plurality of machine learning techniques; and means for ranking the prediction data according to each of the plurality of machine learning techniques.
Example 55 includes the subject matter of any of Examples 45-54, and wherein the telemetry data includes at least one of a packet loss rate, a total amount of the network connections, throughput of the network connections, and latency of the network connections.
Example 56 includes the subject matter of any of Examples 45-55, and further including circuitry for monitoring telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol.
Example 57 includes the subject matter of any of Examples 45-56, and further including means for changing, as a function of the monitored telemetry data, the network communications between the first kernel and the third kernel to the first communication protocol.

Claims

1. A sled comprising:

one or more accelerator devices; and

a compute engine to:

monitor telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications conform to a first communication protocol,

determine, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and

change the network communications to conform to the second communication protocol.

2. The sled of claim 1, wherein to change the network communications to conform to the second communication protocol comprises to:

establish subsequent network communications between the first kernel and the second kernel using the second communication protocol.

3. The sled of claim 2, wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.

4. The sled of claim 1, wherein the second communication protocol corresponds to one of TCP/IP (Transmission Control Protocol/Internet Protocol) or UDP (User Datagram Protocol).

5. The sled of claim 1, wherein the compute engine is further to:

learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time.

6. The sled of claim 5, wherein the compute engine is further to:

generate prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.

7. The sled of claim 6, wherein to determine that the condition to change the network communications is triggered is further determined as a function of the prediction data.

8. The sled of claim 6, wherein to learn one or more change patterns from the monitored telemetry data comprises to:

evaluate the telemetry data associated with the kernel network connections over time; and

identify the change patterns based on the evaluation.

9. The sled of claim 6, wherein the compute engine is further to:

generate the prediction data via a machine learning technique.

10. The sled of claim 6, wherein the compute engine is further to:

generate the prediction data via one or a plurality of machine learning techniques; and

rank the prediction data according to each of the plurality of machine learning techniques.

11. The sled of claim 1, wherein the telemetry data includes at least one of a packet loss rate, a total amount of the network connections, throughput of the network connections, and latency of the network connections.

12. The sled of claim 1, wherein the compute engine is further to:

monitor telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol.

13. The sled of claim 11, wherein the compute engine is further to:

change, as a function of the monitored telemetry data, the network communications between the first kernel and the third kernel to conform to the first communication protocol.

14. The sled of claim 1, wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel exceeds a specified threshold.

15. One or more machine-readable storage media comprising a plurality of instructions stored thereon that, in response to being executed, cause a sled to:

16. The one or more machine-readable storage media of claim 15, wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.

17. The one or more machine-readable storage media of claim 15, wherein the plurality of instructions further causes the sled to:

learn one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time; and

18. The one or more machine-readable storage media of claim 17, wherein to learn one or more change patterns from the monitored telemetry data comprises to:

identify the patterns based on the evaluation.

19. The one or more machine-readable storage media of claim 17, wherein the plurality of instructions further causes the sled to:

20. The one or more machine-readable storage media of claim 15, wherein the plurality of instructions further causes the sled to:

monitor telemetry data associated with one or more network connections between the first kernel and a third kernel, wherein the one or more network connections between the first kernel and the third kernel are established via the second communication protocol; and

21. A method comprising:

monitoring telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications is established via a first communication protocol,

determining, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and

changing the network communications from the first communication protocol to conform to the second communication protocol.

22. The method of claim 21, wherein changing the network communications from the first communication protocol to the second communication protocol comprises:

establishing subsequent network communications between the first kernel and the second kernel using the second communication protocol.

23. The method of claim 22, wherein the second communication protocol is determined as a function of a policy and the monitored telemetry data, wherein the policy defines a plurality of conditions for changing from a given network communication protocol to another network communication protocol.

24. A sled comprising:

circuitry for monitoring telemetry data associated with one or more network communications between a first kernel of the sled and a second kernel configured on a second sled, wherein the one or more network communications conform to a first communication protocol,

means for determining, as a function of the monitored telemetry data, that a condition to change the network communications from the first communication protocol to a second communication protocol is triggered, and

means for changing the network communications from the first communication protocol to conform to the second communication protocol.

25. The sled of claim 24, further comprising:

means for learning one or more change patterns from the monitored telemetry data, wherein each change pattern defines the telemetry data observed over time; and

means for generating prediction data based on the learned one or more change patterns, wherein the prediction data is indicative of a likelihood that, based on subsequently monitored telemetry data, the network communications are to be shifted from the first communication protocol to the second communication protocol or that the network communications are to be shifted from the second communication protocol to the first communication protocol.

26. The sled of claim 1, wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel falls below a specified threshold.

27. The sled of claim 1, wherein the first communication protocol corresponds to TCP/IP, wherein the second communication protocol corresponds to UDP, and wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel falls below a specified threshold.

28. The sled of claim 1, wherein the first communication protocol corresponds to UDP, wherein the second communication protocol corresponds to TCP/IP, and wherein the condition to change the network communications is an indication that a network utilization between the first kernel and the second kernel exceeds a specified threshold.