WO1999035572A1

WO1999035572A1 - Multinode computer system configuration

Info

Publication number: WO1999035572A1
Application number: PCT/FR1998/002885
Authority: WO
Inventors: Alain Boudou; Stéphane BUREAU; Angelo Casamatta; René CHEVANCE; Jean Furelaud; Jean Papadopoulo; Angelo Ramolini; Jacques Talbot
Original assignee: Bull S.A.
Priority date: 1997-12-30
Filing date: 1998-12-28
Publication date: 1999-07-15
Also published as: FR2773239A1

Abstract

The invention concerns a system (S) with a number of processors (P) corresponding to the power desired by the client and shared into nodes (N) according to the tasks assigned to the system by the client, so as to optimise the system performance. The number and size of the nodes can also vary according to the tasks assigned to the system and the change in number of processors in the system.

Description

Title: Configuration of a multinodal computer system.

Description. Technical area.

The subject of the invention is a method for configuring a multinodal computer system and a system implementing this method.

State of the art.

Two main types of multiprocessor computer systems are known today. A system of the first type is called symmetric multiprocessor system and, more commonly, SMP system (Symmetric Multi-Processor). It includes a number n of processors connected to a memory shared by the processors and perceived by them as a single memory. The processors and memory are also connected to an input-output subsystem for communication of the system with the outside. This type of system includes a variant in which the processors are physically connected to respective memories which communicate with the other processors to constitute a shared memory. Communication takes place via a rapid interconnection network so that transfers between processors and memories are carried out in a satisfactory manner for the processors. A symmetrical multiprocessor system according to this variant is called a distributed shared memory system or DSM (Distributed Shared Memory) system. DSM systems include CCNUMA (Cache Coherent, Non-Uniform Memory Access) systems. A system of this first type is managed by a single operating system.

A system of the second type is called a multinodal multiprocessor system or a cluster system of processors. Each node includes one or more processors connected to a memory and to a subset input-output. The nodes of the system are managed by operating systems independent of each other, and frequently identical. All the nodes form a cluster, also called by its Anglo-Saxon designation "cluster".

Each of the two types has its own advantages and disadvantages. A system of the first type is powerful and fast, thanks to the management of data processing by a single operating system, to an intimate connection between processors and the memory and to the use of the memory for the communication between processors. Open systems UNIX (registered trademark) have made the effort to adapt to this type of system and continue to evolve in this direction. Such an open system can currently have up to more than sixty-four processors. On the other hand, a system of the first type has the drawback of having an expensive internal interconnection network, since shared memory requires performance which increases significantly faster than the number of processors. In addition, having only one operating system makes it very vulnerable to a failure of this system and does not ensure availability at all times.

A system of the second type offers good availability, since if an operating system fails, other operating systems can manage the failure and maintain service to users. Another benefit is the operating systems themselves. For example, the NT (registered trademark) operating system evolves so as to form a multi-modal system. Currently, it is well suited to a multinodal system, whose nodes have few processors, of the order of four for example. Under these conditions, this operating system can extend to numerous nodes, to currently sixteen by adding appropriate specialized software, known as "clustering", offered by certain manufacturers. To these advantages is added that of requiring an interconnection network which is today considerably cheaper than that of the systems. of the first type. However, systems of the second type have several drawbacks. First, they are less efficient than systems of the first type with equal total number of processors. This is due to the fact that the internodal communications are made by slower interconnections and that the processors communicate by messages. Such a link requires, in addition to the duration of data transmission, the use of protocols for opening and closing the link. Another major drawback possible for certain operating systems is that it cannot currently adapt to nodes having a relatively large number of processors.

A very important problem relates to the evolution of a computer system. Let us suppose for example a client having invested in a system of the second type, suitable for a business management software usually called under the Anglo-Saxon designation ERP (Enterprise Resource Planning). This system usually includes a powerful node, called the database server, and less powerful nodes, called the application servers. For example, the client system includes a database server formed by a node of six processors, expandable to a maximum of eight processors, and three application servers each including a processor and expandable to a maximum of two processors. This system is therefore formed in total of nine processors. Now assume that the client currently requires a system consisting of a database server of ten processors and four application machines of three processors each. To form this system of twenty-two processors, the client can therefore recover from its initial investment only the node of six processors having served as database server, to make it an application server requiring only three processors. Given the strong growth over time in system performance and customer needs, it is understandable that a problem of this kind often arises. The financial consequences are therefore very significant. Another very important problem concerns improving the efficiency of the availability of the system used. The current solution to this problem is cumbersome and expensive. Currently, the availability of the system used is obtained by operating two roughly equivalent machines which distribute the workload so that the failure of one machine is taken care of by the other machine. In this case, the other machine has only about half of the normal system capacity, while the failure of the other machine may only be due to the failure of one of its processors. We understand that it would be much more profitable to be able to configure the system in relation to a possible failure of this system. This is now not at all possible.

The present invention solves all of these problems.

Summary of the invention

A first object of the invention is to be able to flexibly configure a system in order to meet different requests for use of the system, these requests being able to be different because of different customers who have different needs, or because of needs. different over time for the same client.

A second goal is to be able to change the configuration while giving it the best performance for the increased number of its processors and for the tasks it has to perform.

A third goal is to be able to configure the system according to the degree of failure that it may present, in order to ensure high availability of the system.

A fourth goal is to achieve one of the previous three goals at a reduced cost. The subject of the invention is a method of configuring a computer system, comprising determining at least one group of processors to correspond to a desired processing power, characterized in that it consists in determining the number and / or the size of nodes to be formed with the processors of the group according to the tasks assigned to the group, so as to optimize the performance of the group for the execution of the tasks.

The corollary subject of the invention is also a computer system comprising at least one group of processors corresponding to a desired processing power, characterized in that the group is made up of a number of elementary cells each made of a given number of processors , a memory and an input-output interface and connected together to form and adjust the number and / or size of nodes in the group according to tasks assigned to the group, so as to optimize the performance of the group for the execution of tasks.

The characteristics and advantages of the invention appear from the description which follows, given by way of example and made with reference to the accompanying drawings.

In the drawings:

- Figure 1 is a principle view schematically illustrating processors of a multinodal multiprocessor system according to the invention;

- Figure 2 is a graph illustrating an example of implementation of the method of the invention for the configuration of the system shown in Figure 1; and - Figure 3 is a schematic view of an exemplary embodiment of a multinodal multiprocessor system according to the invention.

Detailed description of examples illustrating the invention.

FIG. 1A represents the processors P of a multinodal computer system S configured according to the method illustrated in FIG. 2. The system illustrated initially comprises a predetermined group Gl of twenty-four processors P corresponding to the power desired by a client. We now assume that the client intends this group of processors to execute one or more sets of tasks. For example, the tasks chosen relate to the operation of ERP software. This software requires a database server relating to a first set of tasks, and one or more application servers relating to other respective sets of tasks.

FIG. 2 is a graph having for abscissa P / N representative of the number of processors per node and for ordinates the number of nodes N of the system. This graph illustrates with FIG. 1A an example of the method according to the invention for configuring the system S. The method consists in determining the number and / or the size of nodes N to be formed with the processors P of the group Gl, so as to optimize the performance of the processor group for the execution of the ERP software task sets. In FIG. 2, each node N is followed by a first digit relating to the group to which it belongs and by a second digit in index relating to the number of processors P it contains. With reference to FIGS. 1A and 2, the client having the twenty-four processors P of the group Gl has distributed them into a node of ten processors Nl ₁₀ as the database server, and into a biprocessor node Nl ₂ and three quadriprocessor nodes Nl as application servers. FIG. 1B illustrates an example of evolution of the configuration of the group Gl. According to this example, it is assumed that the client must have, with the same number of P processors from group Gl, a slightly more powerful database server, keep two of the three quad-processor application servers and have two other servers. application instead of the third quad-processor server of the previous configuration. According to the example illustrated in FIGS. 1B and 2, the connection of the twenty-four processors P of the group Gl is modified to have a node of twelve Ni 12 processors as a database server and, as application servers, a node monoprocessor Nli, a triprocessor node NI3 and two quadriprocessor nodes N'l.

FIG. 1C illustrates with FIG. 2 another example of possible evolution of the configuration of the system S according to the invention. In this example, the client still needs to increase the power of the database server and increase the number of application servers. According to an advantage of the invention, the total number of processors can include any additional number desired to develop the system. In the example illustrated in FIGS. 1C and 2, the client having the ERP software simply has to add to the preceding group Gl a number of eight processors P symbolically represented in an additional column. Using the method of the invention, the customer thus has a new group G2 of thirty-two processors and can distribute them at will so as to optimize the performance of the ERP software. The new configuration chosen includes a node of sixteen N2iβ processors as the base server, three biprocessor nodes N2 ₂ , two triprocessor nodes N2 ₃ and a quadriprocessor node N2. However, we understand that it could define any other configuration and that it can evolve into a new configuration.

The customer could also reduce the number of processors for the ERP software and have the other processors to make a configuration variable suitable for other tasks, optionally with additional processors available from the system or added to the system. Following this logic, in the event of a node failure it is now possible to envisage an additional configuration adapted to the failure. For example, if a processor P breaks down in a node N2, it becomes possible to adapt the configuration of the group G2 so as to isolate the failing processor from the nodes of the system. For example, the node N2 ₄ can be configured as a triprocessor node. If this configuration is insufficient for the application for which it is responsible, the node N2 could be reconfigured into a quadriprocessor node by temporarily borrowing a processor from another node of the group or from another group which can support the deletion of a single processor.

FIG. 1D illustrates with FIG. 2 two extreme configurations which can be obtained from the method of the invention from a group G3 of thirty-two processors P intended for various tasks other than the ERP software. These processors can be contained in a single N332 node or in thirty two N3_ monoprocessor nodes. Between these two borderline cases, any configuration is possible. Among the configurations illustrated in FIG. 2 only, the thirty-two processors P can be divided into two nodes of sixteen processors N3i6, or into four octoprocessor nodes N3β, or even into eight quadriprocessor nodes N3 ₄ .

FIG. 3 illustrates an example of implementation of the method of the invention. This implementation consists in forming the number and / or the size of the nodes using elementary cells C each made from a given number of processors. FIG. 3 illustrates a group of eight biprocessor cells interconnected to form three nodes Na, Nb and Ne in a system S according to the invention. Each cell C comprises two processors P connected together as well as to a memory M and to an input-output interface I via a link formed by a bus B in the example illustrated. Each C cell can alone form a node, such as the Na node illustrated. The cell then operates with an autonomous operating system, that is to say independent of all the operating systems which manage the other nodes. The illustrated nodes are delimited by a ghost line. The cells are interconnected with one another by a network K of links L, preferably high speed data transmission links. For reasons of clarity of the drawings, only the links L relating to the cell C constituting the Na node have been shown. Thanks to the intercellular L links, the group of eight system processors S can form an autonomous system whose nodes are each made of a single cell C.

This system can be easily configured according to the needs of each client of such a system. In the example illustrated in FIG. 3, it has been assumed that the client has a system S with sixteen processors and needs a biprocessor node Na, a hexaprocessor node Nb and an octoprocessor node Ne. The invention therefore consists in operating three C cells so as to form the Nb node and four other cells so as to form the Ne node. In practice, as the L links connect each cell to each of the other cells, the assembly of cells to form a node therefore consists simply in including in each cell, preferably in the interface I of the cell as illustrated, a block of U communication, wired or programmable, to identify the other cells and operate the cell with a given number of determined cells. An essential aim of the communication block U consists in enabling shared access to the memories M of the cells of the node by all of the processors of this node and in managing the consistency of these memories. The communication blocks U therefore operate so as to form one or more systems of the first type which has been defined in the introduction to the present application. These are DSM systems as soon as a node has more than one cell. In addition, the communication block U makes it possible to identify the other nodes of the system so as to manage the internodal transmissions according to the second type. According to another approach of the invention, it can be said that a variable configuration can therefore be easily obtained from a system S of the DSM type, by isolating between them sets of processors with their respective memories in order to make them different nodes. For a system of the first type, it suffices to form groups of processors with subsets of the memory of the system, and to make these groups communicate with each other by means of messages. If the groups of processors could not communicate with each other, we would then return to a well-known functionality for partitioning systems, which the invention therefore makes it possible to also treat.

The examples relating to FIGS. 1 and 2 assume the use of monoprocessor cells, which offers the advantage of giving maximum flexibility in the configuration of a system. However, this configuration would require the insertion of a communication block U in each of the cells and would therefore be very expensive. Multiprocessor cells are a compromise between configuration flexibility and system cost. The number of processors per cell also depends on the maximum allowable number of processors in the system. The biprocessor cells of Figure 3 are rather suitable for small systems, while a good current compromise for large systems would be quadriprocessor cells.

The invention also offers the advantage of being able to take advantage of the specific advantages of the operating systems available on the market when choosing the configuration for operating the nodes. Currently, for example, we would choose the NT operating system to manage several nodes with a low number of processors and the UNIX operating system for the other nodes. During a next configuration made a few years later, for example, it would be possible to choose the operating systems best suited to the new configuration. It therefore appears from the preceding examples that the invention has as its general object a method of configuring a computer system consisting in determining at least one group G1-G3 of processors P to correspond to a desired processing power and in determining the number and / or the size of nodes N to be formed with the processors of the group as a function of tasks assigned to the group, so as to optimize the performance of the group for the execution of the tasks.

In short, a group has a number of processors P corresponding to the power desired by the client and is divided into nodes N according to the tasks assigned to the group by the client, so as to optimize the performance of this group. The number and size of nodes can thus vary depending on the tasks assigned to the group and the change in the number of processors in the group.

Consequently, it is possible to deliver to any customer an unconfigured multiprocessor computer system having a sufficient number of processors to meet the needs of this customer. The starting criterion is therefore the power that the system must have to meet the needs of a client, regardless of the nature of these needs. For example, an unconfigured octoprocessor system can serve multiple functions. The customer can then have this system at any time to configure it so as to optimize system performance according to the nature of the needs at the time. When the customer will need more power, for example about half of what he currently has, he will just have to buy the number of processors corresponding to half of the additional power desired and configure the assembly processors it has. This configuration can be an improvement of the existing system, or can be adapted to other tasks. For example, it is possible to transform an existing configuration, adapted to scientific calculations, into a configuration used for example for ERP services. In this case, the processors of the system will be divided into two groups intended for separate tasks, such as the tasks of database server and those of application server which have been illustrated with reference to FIGS. 1A-1C. The notion of group is therefore very flexible and can be all or part of the processors of a system.

We have also seen that an important advantage of the invention consists in varying the number and / or the size of nodes in a group as a function of a number of faulty processors of this group. We can also have other processors available from the system outside the group.

We have also seen that this method consists, for its implementation, in forming the number and / or the size of the nodes in the group by using elementary cells each made of a given number of processors.

For the formation of a node from a cell, it is sufficient that the cell has at least one processor, a memory and an input-output interface. For the formation of a node from at least two cells, a communication block U is included in the cells of the node. In other words, all the cells must include a communication block U if it is desired that each of the cells can form a node with another cell. This block allows shared access to the memories M of the cells of the node by all of the processors of this node and manages the consistency of these memories.

It follows that the invention also relates to a computer system S comprising at least one group G1-G3 of processors P corresponding to a desired processing power, the group being made of a number of elementary cells C each made of a given number of processors P, a memory M and an input-output interface I and connected together to form and adjust the number and / or the size of nodes N in the group as a function of tasks assigned to the group, so as to optimize group performance for the execution of tasks. The input-output interface includes a communication block U allowing shared access to the memories M of the cells of the node by all the processors of this node and the management of the consistency of these memories. Preferably, this block identifies the cells of the node and / or the other nodes, although this identification can be located elsewhere.

Preferably also, the system comprises means M for varying the number and / or the size of nodes in a group as a function of a number of faulty processors of this group. Preferably, these dynamic reconfiguration means M would be included in server or platform management software and placed in a service processor Ps of the system S, as shown in FIG. 3. The service processor of a system is used in particular for system initialization and preferably for fault analysis.

Claims

Claims:

1. A method of configuring a computer system (S), comprising determining at least one group (G1-G3) of processors (P) to correspond to a desired processing power, characterized in that it consists in determine the number and / or size of nodes (N) to be formed with the group processors as a function of tasks assigned to the group, so as to optimize the performance of the group for the execution of the tasks.

2. Method according to claim 1, characterized in that the number of processors of the group being modified, it consists in again determining the number and / or the size of the nodes of the group according to tasks assigned to the group.

3. Method according to claim 2, characterized in that it consists in varying the number and / or the size of nodes in a group as a function of a number of faulty processors of this group.

4. Method according to one of claims 1 to 3, characterized in that it consists in forming the number and / or the size of the nodes in the group using elementary cells (C) each made of a given number of processors.

5. Method according to claim 4, characterized in that the formation of a node from at least two cells consists in inserting into each of the cells of the node a communication block (U) allowing shared access to the memories ( M) of the cells of the node by all the processors of this node and the management of the consistency of these memories.

6. Computer system (S) comprising at least one group

(G1-G3) of processors (P) corresponding to a desired processing power, characterized in that the group is made up of a number of cells elementary (C) each made of a given number of processors (P), a memory (M) and an input-output interface (I) and connected together to form and adjust the number and / or the size of nodes (N) in the group as a function of tasks assigned to the group, so as to optimize the performance of the group for the execution of the tasks.

7. System according to claim 6, characterized in that the input-output interface of a cell intended to form a node with another cell includes a communication block (U) un- oe-de-eommunication— - (U) - allowing shared access to the memories (M) of the cells of the node by all the processors of this node and the management of the consistency of these memories.

8. System according to claim 6 or 7, characterized in that the number of processors of the group being modified, it consists in again determining the number and / or the size of the nodes of the group as a function of tasks assigned to the group.

9. System according to claim 8, characterized in that it comprises means for varying the number and / or the size of nodes in a group as a function of a number of faulty processors of this group.

10. System according to claim 9, characterized in that said means analyze in the system the potential availability of processors outside the group and add to this group at least one of the available processors.