CN118069349A - A variable depth resource management method and system for multiple scenarios - Google Patents

Abstract

The invention discloses a variable depth resource management method and a system for multiple scenes, wherein the method comprises the following steps: receiving high-performance load or high-flux load submitted by a user and judging the load type; if the high-performance load is the high-performance load, performing resource allocation on the high-performance load job task through a resident resource management instance, and controlling the execution of the high-performance load job task, wherein the resident resource management instance is of a fixed depth; if the high-flux load is the high-flux load, the high-flux load manager analyzes and splits the high-flux load into a plurality of subtask sets, after the resource allocation is completed, a temporary resource management father instance and a subtask instance are started according to the number of the subtask sets to control the execution of the job tasks, and the temporary resource management father instance, the subtask instance and the high-flux load manager are temporarily deployed and destroyed along with the high-flux load to realize variable depth. The invention dynamically changes the depth of the resource management system according to the system load type so as to adapt to high-performance calculation and high-flux calculation, thereby providing stable and efficient resource management for the high-performance calculation and the high-flux calculation at the same time.

Description

A variable depth resource management method and system for multiple scenarios

Technical Field

The present invention mainly relates to the technical field of system resource management, and specifically to a variable depth resource management method and system for multiple scenarios.

Background technique

A supercomputer consists of a large number of tightly coupled computing nodes, which are interconnected through a high-speed interconnection network, providing extremely high parallelism for scientific computing. On this basis, high-performance computing (HPC) and high-throughput computing (HTC) are the two main computing modes of supercomputers. High-performance computing is the traditional computing mode of supercomputers, which is mainly aimed at solving large-scale scientific computing and scientific simulation problems, and is widely used in engineering design, weather forecasting, nuclear physics research and other fields. A single high-performance computing load contains one or more high-performance computing tasks. According to the scientific purpose of high-performance computing, there may be certain dependencies between multiple high-performance tasks in the same high-performance computing load. Each high-performance computing task is mapped as a job in the supercomputer. A single high-performance job usually contains a large number of computing processes and requires a large amount of computing and storage resources. High-performance jobs usually distribute computing tasks to multiple computing processes through the message passing paradigm (MPI), and use the powerful computing and communication performance of supercomputers to complete scientific computing. With the increase in the scale of supercomputers, the scale of high-performance computing continues to increase. High-performance computing requires the resource management system to have high scalability and the ability to manage large-scale parallel jobs. Therefore, for high-performance computing loads, the performance indicators of the resource management system mainly include the overall system resource utilization and the average waiting time of jobs at different scales, etc.

High-throughput computing is an emerging computing model, mainly used in fields such as biomedicine, machine learning, and information security. A single high-throughput computing load consists of a large number of loosely coupled and independent high-throughput tasks. Among them, each high-throughput task is usually mapped as a job in a supercomputer, and the jobs can usually be executed in any order. In addition, a single high-throughput task does not require high computing and memory resources, and usually only needs to be mapped to a single computing core to run. The high concurrency of high-throughput computing requires the resource management system to have high throughput (throughput refers to the average amount of data that can be successfully transmitted per unit time on a communication channel), that is, it has the ability to schedule, load and monitor a large number of small tasks for a long time. Therefore, for high-throughput computing loads, the performance indicators of the resource management system mainly include system throughput, task operation efficiency, etc.

On a supercomputer, high-performance computing and high-throughput computing may occur at the same time, and the same computing resources may be used for both computing modes successively. This complex application scenario requires the supercomputer to provide stable and efficient resource management services for both computing modes at the same time. However, existing resource management systems often cannot take into account the management of high-performance computing and high-throughput computing tasks at the same time, resulting in resource waste or slow computing speed when managing at least one of the two computing tasks. Therefore, there is an urgent need to provide a resource management method and system that can effectively manage high-performance computing and high-throughput computing tasks at the same time.

Summary of the invention

The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a variable-depth resource management method and system for multiple scenarios to solve the problem that the existing resource management system cannot take into account the management and execution of high-performance computing and high-throughput computing tasks at the same time due to the fixed depth.

In order to solve the above technical problems, the technical solution proposed by the present invention is:

A variable depth resource management method for multiple scenarios, comprising the following steps:

Receive high-performance loads or high-throughput loads submitted by users and determine the load type;

If the load type is determined to be a high-performance load, resources are allocated to the high-performance load job task by calling a resident resource management instance in the resource management system, and after the resource allocation is completed, the high-performance load job task is controlled and executed by the resident resource management instance, and the resident resource management instance is a fixed depth;

If it is determined that the load type is a high-throughput load, the high-throughput load manager is started through the unified interface in the resource management system, and the high-throughput load manager parses the high-throughput load job task and splits it into multiple sub-task sets, and allocates resources through the resident resource management instance. After the resource allocation is completed, the resident resource management instance is used to start a temporary resource management parent instance and multiple temporary resource management child instances according to the number of sub-task sets of the high-throughput load job task, and the temporary resource management parent instance and each of the temporary resource management child instances control the execution of the high-throughput load job task. The temporary resource management parent instance, the temporary resource management child instance and the high-throughput load manager are temporarily deployed and destroyed with the high-throughput load to achieve variable depth.

As a further improvement of the above technical solution:

If it is determined that the load type is a high-performance load, resources are allocated to the high-performance load job task by calling the resident resource management instance in the resource management system, including:

The unified interface in the resource management system calls the user interface of the resident resource management instance located on the computer login node, and requests resource allocation from the resident resource management instance control node agent process located on the computer control node;

The resident resource management instance control node agent process parses the high-performance load job task to schedule the corresponding system resources to be allocated to the current job task, and returns the resource allocation result to the resident resource management instance user interface on the login node;

The resident resource management instance control node agent process allocates a unique job number to the current high-performance load job task.

If it is determined that the load type is a high-performance load, the high-performance load job task is executed by controlling the resident resource management instance, including:

The resident resource management instance control node agent process sends the relevant information of the high-performance load job task to the resident resource management instance computing node agent process on the computing node corresponding to the current job task through the user interface;

The resident resource management instance computing node agent process loads and runs high-performance load job tasks;

During the operation of high-performance load tasks, the resident resource management instance monitors the operation status and records the real-time monitored status on the resident resource management instance control node agent process. After the current high-performance load task is completed, the resident resource management instance reclaims the resources allocated to the current high-performance load task to complete the operation of the high-performance load.

If it is determined that the load type is a high-throughput load, resource allocation is performed through the resident resource management instance, including:

The high-throughput load manager analyzes the total resource requirements of the current high-throughput load, determines the total number of computing nodes required, and splits the high-throughput tasks in the high-throughput load into multiple subtask sets through the high-throughput load manager, and generates a subtask set loading script for each subtask set;

The high throughput load manager calls the user interface of the resident resource management instance and sends a resource allocation request to the resident resource management instance control node agent process;

After receiving the resource allocation request, the resident resource management instance control node agent process allocates resources and a unique job number located on the computer computing node to the high-throughput load job task, and returns it to the high-throughput load manager.

If it is determined that the load type is a high-throughput load, the temporary resource management parent instance and multiple temporary resource management child instances are started according to the number of subtask sets of the high-throughput load job task by the resident resource management instance, and the temporary resource management parent instance and each of the temporary resource management child instances control the execution of the high-throughput load job task, including:

The resident resource management instance sends the current job related information to each computing node assigned to the current high throughput load through the user interface on the login node;

On each computing node assigned to the current task, the resident resource management instance computing node agent process starts a temporary resource management parent instance, which is a first-level temporary resource management instance and is composed of a temporary resource management instance agent process on each assigned node to manage all resources assigned to the current high-throughput load;

The temporary resource management parent instance starts multiple temporary resource management child instances, each of which manages the computing resources of some computing nodes and runs the corresponding subtask set loading script. When all high-throughput tasks in the subtask set loading script corresponding to the temporary resource management child instance are completed, the temporary resource management child instance exits;

During the operation of the high-throughput load task, the high-throughput load manager continuously monitors the real-time status information of the high-throughput load task according to the unique job number, and stores the real-time status information in the temporary resource management instance global database in the resource management system.

The temporary resource management parent instance starts M temporary resource management child instances, where M is the number of split high-throughput quantum task sets, each temporary resource management child instance includes N/M proxy processes corresponding to N/M computing nodes, and each temporary resource management child instance manages the computing resources of N/M computing nodes, where N is the number of computing nodes allocated to the current high-throughput load. When the corresponding high-throughput task loading script is run, the high-throughput task is scheduled to run on the managed computing resources according to the FIFO scheduling strategy.

If it is determined that the load type is a high-throughput load, the temporary resource management parent instance and multiple temporary resource management child instances are started according to the number of subtask sets of the high-throughput load job task by the resident resource management instance, and the temporary resource management parent instance and each of the temporary resource management child instances control the execution of the high-throughput load job task, further comprising:

When the high-throughput load manager monitors that some tasks in the high-throughput load job have run errors, it determines whether the task with the run error is the first run error. If it is the first run error, a new task script is generated, and a new temporary resource management child instance is started under the temporary resource management parent instance, and the task with the run error is submitted to the new temporary resource management child instance for execution; if it is not the first run error, error information is reported;

When the high-throughput load manager monitors that all tasks in the high-throughput load job have been completed correctly, the temporary resource management parent instance is destroyed, the resident resource management instance reclaims the resources allocated to the high-throughput load, and the high-throughput load manager exits to complete the operation of the high-throughput load.

The method further includes querying the operation task information, and the operation task information query includes the steps of:

The resource management system receives the query request and determines the type of the job task according to the job number of the job task to be queried in the query request;

If it is determined that the job task to be queried is a high-performance load, the resource management system calls the user interface of the resident resource management instance located on the computer login node through the unified interface to issue a job task information query request, and the control node agent process of the resident resource management instance located on the computer control node parses the job task information query request, and queries the relevant information of the job task from the local data structure in the resource management system and returns it to the unified interface;

If it is determined that the job task to be queried is a high-throughput load, the resource management system queries the relevant information of the job task according to the job number through the temporary resource management instance global database, and returns it to the unified interface.

The present invention also provides a variable depth resource management system for multiple scenarios, the resource management system comprising:

The resource management system interface module is used to receive high-performance loads or high-throughput loads submitted by users and determine the load type;

A high-performance load job task management module is used to allocate resources to the high-performance load job task by calling a resident resource management instance in the resource management system if the load type is determined to be a high-performance load, and after the resource allocation is completed, control the execution of the high-performance load job task through the resident resource management instance, and the resident resource management instance is a fixed depth;

A high-throughput load job task management module is used to start a high-throughput load manager through a unified interface in the resource management system if the load type is determined to be a high-throughput load. The high-throughput load manager parses the high-throughput load job task and splits it into multiple subtask sets. Resources are allocated through the resident resource management instance. After the resource allocation is completed, a temporary resource management parent instance and multiple temporary resource management child instances are started through the resident resource management instance according to the number of subtask sets of the high-throughput load job task. The temporary resource management parent instance and each of the temporary resource management child instances control the execution of the high-throughput load job task. The temporary resource management parent instance and the temporary resource management child instance high-throughput load managers are temporarily deployed and destroyed with the high-throughput load to achieve variable depth.

The present invention also provides a computer-readable storage medium storing a computer program, and the computer program implements the above method when executed by a processor.

Compared with the prior art, the advantages of the present invention are:

The present invention determines the load type and dynamically changes the depth of the resource management system according to the system load type to adapt to high-performance computing and high-throughput computing. For high-performance loads, resident resource management instances are directly used to provide services for them to ensure good support for high-performance loads. When a high-throughput load containing a large number of high-throughput tasks needs to be executed, temporary resource management parent instances and child instances are started to increase the depth of the resource management system, which can improve the throughput of the resource management system and achieve efficient management of high-throughput loads, thereby enabling the computer to provide stable and efficient resource management for both high-performance computing and high-throughput computing.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is an architecture diagram of a variable depth resource management system for multiple scenarios in an embodiment of the present invention.

FIG2 is a flow chart of the variable depth resource management method for multiple scenarios according to this embodiment.

FIG3 is a flow chart of executing high-performance load and high-throughput load job tasks in the resource management method of this embodiment.

FIG. 4 is a flow chart of querying high-performance load and high-throughput load job tasks in the resource management method of this embodiment.

Detailed ways

In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

The resource management system is an important system software of high-performance computer systems, responsible for managing the resources of the entire system and responding to user requests. The performance of the resource management system directly affects key indicators such as system resource utilization, system throughput, job turnaround time and parallel application performance, and is directly related to the overall system performance and user experience. The efficiency of supercomputer use depends to a certain extent on the function and operation quality of the resource management system. Supercomputers mainly contain three main types of nodes, including login nodes, control nodes and computing nodes. Login nodes are used for users to log in to the system and interact with the system's job scheduler. Control nodes are used to coordinate and manage various resources and services in the computer network. Computing nodes are used to perform computing tasks.

A unit in a resource management system that has resource management and job scheduling functions is called a resource management instance. There may be a nested relationship between resource management instances, that is, a resource management instance can assign a subset of the resources it manages to a new resource management instance. Therefore, a resource management system can consist of one resource management instance or multiple resource management instances that may have a nested relationship.

The depth of a resource management system refers to the maximum number of nested relationships between resource management instances. The depth of a resource management system directly affects the performance of the resource management system. Specifically, the depth of a resource management system determines the number of nested layers of resource management instances. The greater the number of nested layers, the more units with resource management capabilities there are in the system, the higher the degree of distribution of the resource management system, the more jobs that can be scheduled simultaneously, and the greater the throughput. On the other hand, the greater the number of nested layers, the higher the overhead of maintaining resource management instances, and the greater the overhead of resource coordination and load balancing in resource management instances at each layer.

Currently, commonly used resource management systems use resident resource management instances with a fixed depth of 1. They have a global view of system resources and jobs and a rich set of resource allocation and job scheduling strategies. They can cope well with multi-user environments and manage system resources and jobs efficiently. Therefore, this resident resource management instance with a fixed depth of 1 has good support for large-scale parallel computing loads and is suitable for high-performance computing. However, due to the fixed depth, this resource management system cannot increase resource management units through instance nesting, is usually limited by throughput, and is not suitable for high-throughput computing.

In order to solve the problem that the depth of the traditional resource management system is fixed and not conducive to high-throughput computing, this embodiment constructs a variable-depth resource management system for multiple scenarios as shown in Figure 1. On the login node, control node and computing node of the computer, the variable-depth resource management system for multiple scenarios is provided with a resident resource management instance, a temporary resource management instance, a high-throughput load manager and a temporary resource management instance global database.

The resident resource management instance is deployed in the system for a long time. As the first-level instance of the resource management system, it manages all resources in the system. The resident resource management instance starts the control node agent process on the control node and starts the compute node agent process on all compute nodes. The user interface provided by the resident resource management instance can be directly called by the user on the login node. The resident resource management system has a global view of resources and jobs, and its rich allocation and scheduling strategies can provide complete job management services for traditional high-performance loads.

The temporary resource management instance is started by the resident resource management instance and is temporarily deployed and destroyed with the high-throughput load. One high-throughput load corresponds to one temporary resource management instance. For high-throughput loads, the resident resource management instance only provides resource allocation services for it, and manages a large number of high-throughput tasks on the allocated resources by starting the temporary resource management instance. When multiple high-throughput load tasks are executed simultaneously in the resource management system, multiple temporary resource management instances are running at the same time. The agent processes of the temporary resource management instance all run on the computing nodes, which are called the computing node agent processes of the temporary resource management instance. The temporary resource management instance provides job management services for the high-throughput load on the resources allocated by the resident resource management instance.

The high-throughput load manager is deployed on the login node as a runtime tool that is temporarily deployed and destroyed with the high-throughput load. One high-throughput load corresponds to one high-throughput load manager. When multiple high-throughput load tasks are executed simultaneously in the resource management system, multiple high-throughput load managers run simultaneously.

The global database of the temporary resource management instance is permanently deployed on the login node to store the real-time status of all high-throughput tasks in each high-throughput load, as well as the information corresponding to each high-performance load.

According to the architecture of the variable depth resource management system for multiple scenarios provided in this embodiment, this embodiment provides a variable depth resource management method for multiple scenarios as shown in FIG2 , and the steps include:

Step S01, receiving a high-performance load or a high-throughput load submitted by a user and determining the load type;

Step S02: if the load type is determined to be a high-performance load, resources are allocated to the high-performance load job task by calling the resident resource management instance in the resource management system, and after the resource allocation is completed, the high-performance load job task is controlled by the resident resource management instance, and the resident resource management instance has a fixed depth;

Step S03, if it is determined that the load type is a high-throughput load, the high-throughput load manager is started through the unified interface in the resource management system, and the high-throughput load manager parses the high-throughput load job task and splits it into multiple sub-task sets, and allocates resources through the resident resource management instance. After the resource allocation is completed, the resident resource management instance is used to start a temporary resource management parent instance and multiple temporary resource management child instances according to the number of sub-task sets of the high-throughput load job task, and the temporary resource management parent instance and each temporary resource management child instance control the execution of the high-throughput load job task, and the temporary resource management parent instance, the temporary resource management child instance and the high-throughput load manager are temporarily deployed and destroyed with the high-throughput load to achieve variable depth.

Specifically, in step S01, the user calls the unified interface of the variable-depth resource management system for multiple scenarios on the system login node to submit a load script. The unified interface of the resource management system checks the load script submitted by the user and determines the corresponding type.

It can be understood that this embodiment can dynamically change the depth of the resource management system according to the system load type to adapt to high-performance computing and high-throughput computing by judging the load type. That is, for traditional high-performance loads, resident resource management instances are directly used to provide services for them to ensure good support for high-performance loads; when it is necessary to execute high-throughput loads containing a large number of high-throughput tasks, temporary resource management parent instances and child instances are started to increase the depth of the resource management system to improve the throughput of the resource management system and achieve efficient management of high-throughput loads, so that the computer can provide stable and efficient resource management for both high-performance computing and high-throughput computing.

In this embodiment, in step S02, if it is determined that the load type is a high-performance load, resources are allocated to the high-performance load job task by calling the resident resource management instance in the resource management system, including:

Step S201, the unified interface in the resource management system calls the user interface of the resident resource management instance located on the computer login node, and requests resource allocation from the resident resource management instance control node agent process located on the computer control node;

Step S202, the resident resource management instance control node agent process parses the high-performance load job task to schedule the corresponding system resources to be allocated to the current job task, and returns the resource allocation result to the resident resource management instance user interface on the login node;

Step S203: the resident resource management instance control node agent process allocates a unique job number to the current high-performance load job task.

In this embodiment, if it is determined that the load type job task is a high-performance load, the high-performance load job task is executed by controlling the resident resource management instance, including:

Step S211, the resident resource management instance control node agent process sends the relevant information of the high-performance load job task (including the job script and resource allocation information) to the resident resource management instance computing node agent process on the computing node corresponding to the current job task through the user interface;

Step S212, the resident resource management instance computing node agent process loads and runs the high-performance load job task;

Step S213, during the running of the high-performance load job task, the resident resource management instance monitors the running status and records the real-time monitored status on the resident resource management instance control node agent process. After the current high-performance load job task is completed, the resident resource management instance reclaims the resources allocated to the current high-performance load job task to complete the running of the high-performance load.

As shown in FIG3 , in a specific application embodiment, the detailed steps when the operation task is a high-performance load type are as follows:

Step 1: The unified interface of the variable-depth resource management system for multiple scenarios directly calls the resident resource management instance user interface to request resource allocation from the resident resource management instance proxy process on the control node;

Step 2: The resident resource management instance control node agent process parses the high-performance load script, schedules appropriate system resources to be allocated to the job task, assigns a unique job number to the high-performance load job task, and returns the resource allocation result to the resident resource management instance user interface on the login node. At this point, the high-performance load runs in the system as a specific high-performance job.

Step 3, the resident resource management instance control node agent process sends the relevant information of the high-performance load job task (including the job script and resource allocation information) to the resident resource management instance computing node agent process on the computing node assigned to the job task through the user interface;

Step 4: The resident resource management instance computing node agent process loads the job task and starts the user process. At this time, the high-performance load job task submitted by the user starts running;

Step 5: The resident resource management instance is responsible for monitoring the running status of the high-performance load job task, recording the real-time status of the job task on the resident resource management instance control node agent process, and recovering the computing resources allocated to the high-performance load job task after the high-performance load job task is completed. At this point, the high-performance load is completed.

It can be understood that the depth of the resource management system refers to the maximum number of nested relationships between resource management instances. When this embodiment determines that the job task is a high-performance load type, based on the characteristics of high-performance computing large-scale parallel computing loads, system resources and job tasks can be efficiently managed through fixed-depth resident resource management instances, effectively saving system resources and reducing costs without affecting computing performance.

In this embodiment, if it is determined that the load type is a high-throughput load, resource allocation is performed through the resident resource management instance, including:

Step S301, the high-throughput load manager analyzes the total resource requirements of the current high-throughput load, determines the total number of computing nodes required, and splits the high-throughput tasks in the high-throughput load into multiple subtask sets through the high-throughput load manager, and generates a subtask set loading script for each subtask set; wherein the number of high-throughput tasks contained in each subtask set should be as similar as possible;

Step S302, the high throughput load manager calls the user interface of the resident resource management instance to send a resource allocation request to the resident resource management instance control node agent process;

Step S303, after receiving the resource allocation request, the resident resource management instance control node agent process allocates resources and a unique job number located on the computer computing node to the high throughput load job task, and returns it to the high throughput load manager.

In this embodiment, if it is determined that the load type job task is a high-throughput load, a temporary resource management parent instance and multiple temporary resource management child instances are started according to the number of subtask sets of the high-throughput load job task through the resident resource management instance, and the temporary resource management parent instance and each temporary resource management child instance control the execution of the high-throughput load job task, including:

Step S311, the resident resource management instance sends the current job related information to each computing node assigned to the current high throughput load through the user interface on the login node;

Step S312, on each computing node assigned to the current task, the resident resource management instance computing node agent process starts the temporary resource management parent instance, the temporary resource management parent instance is a first-level temporary resource management instance, and is composed of a temporary resource management instance agent process on each assigned node, so as to manage all resources assigned to the current high-throughput load;

Step S313, the temporary resource management parent instance starts multiple temporary resource management child instances, each of which manages the computing resources of some computing nodes and runs the corresponding subtask set loading script. When all high-throughput tasks in the subtask set loading script corresponding to the temporary resource management child instance are completed, the temporary resource management child instance exits;

Step S314: During the execution of the high-throughput load task, the high-throughput load manager continuously monitors the real-time status information of the high-throughput load task according to the unique job number, and stores the real-time status information in the temporary resource management instance global database in the resource management system.

In this embodiment, the temporary resource management parent instance starts M temporary resource management child instances, where M is the number of split high-throughput quantum task sets, each temporary resource management child instance contains N/M proxy processes corresponding to N/M computing nodes, and each temporary resource management child instance manages the computing resources of N/M computing nodes, where N is the number of computing nodes allocated to the current high-throughput load. When the corresponding high-throughput task loading script is run, the high-throughput task is scheduled to run on the managed computing resources according to the FIFO scheduling strategy.

In this embodiment, if it is determined that the load type is a high-throughput load, a temporary resource management parent instance and multiple temporary resource management child instances are started according to the number of subtask sets of the high-throughput load job task through the resident resource management instance, and the temporary resource management parent instance and each temporary resource management child instance control the execution of the high-throughput load job task, and also include:

Step S321, when the high-throughput load manager monitors that some tasks in the high-throughput load job have run errors, it is determined whether the task with the run error is the first run error. If it is the first run error, a new task script is generated, and a new temporary resource management child instance is started under the temporary resource management parent instance, and the task with the run error is submitted to the new temporary resource management child instance for execution; if it is not the first run error, an error message is reported;

Step S322, when the high-throughput load manager monitors that all tasks in the high-throughput load job have been completed correctly, it destroys the temporary resource management parent instance, the resident resource management instance reclaims the resources allocated to the high-throughput load, and the high-throughput load manager exits to complete the operation of the high-throughput load.

As shown in FIG3 , in a specific application embodiment, the detailed steps when the operation task is a high-throughput load type are as follows:

Step 1: A high-throughput load manager is started through a unified interface of a resource management system with variable depth for multiple scenarios;

Step 2: The job parsing/submission component in the high-throughput workload manager parses the total resource requirements of the high-throughput workload and determines the total number of computing nodes N required. Then, the high-throughput workload manager splits the high-throughput tasks in the high-throughput workload into M subtask scripts (M is configurable) and ensures that the number of high-throughput tasks contained in each script is as close as possible.

Step 3: The job parsing/submission component of the high throughput load manager calls the resident resource management instance user interface to send a resource allocation request to the resident resource management instance control node agent process;

Step 4: The resident resource management instance control node agent process allocates computing resources and a unique job number to the high-throughput load, and returns it to the high-throughput load manager. The job monitoring/resubmission component of the high-throughput load manager starts to continuously monitor the real-time status of the task according to the job number and updates the real-time status information to the temporary resource management instance global database;

Step 5: The resident resource management instance sends the relevant information of the job task to each computing node to which the high-throughput load has been assigned through the user interface on the login node;

Step 6: On each computing node to which the high-throughput load subtask is assigned, the resident resource management instance computing node agent process starts a first-layer temporary resource management instance (referred to as the parent instance for short), which is composed of the temporary resource management instance computing node agent processes on the M assigned nodes and is used to manage all resources assigned to the high-throughput load;

Step 7: The parent instance starts M second-layer temporary resource management instances (referred to as child instances). Each child instance contains N/M temporary resource management instance computing node agent processes corresponding to N/M computing nodes. Each child instance manages the computing resources of N/M computing nodes, runs the high-throughput task loading script corresponding to the computing node, and schedules the high-throughput tasks to run on the managed computing resources according to the FIFO (first come, first served) scheduling policy.

Step 8: When all high-throughput tasks corresponding to the sub-instance are completed, the corresponding sub-instance is destroyed;

Step 9, when the job monitoring/resubmission component in the high-throughput load manager monitors that some tasks in the high-throughput load job have run errors, confirm whether the task with the run error is the first run error: if it is the first run error, generate a new task script, start a new child instance under the parent instance, and submit the high-throughput task with the run error to the child instance again for running; if it is the second run error, report the error information to the user and jump to step 11;

Step 10: When the job monitoring/resubmission component in the high-throughput load manager monitors that all tasks in the high-throughput load job task have been successfully completed, the parent instance is destroyed, and the resident resource management instance reclaims the computing resources allocated to the high-throughput load;

Step 11: The high-throughput load manager is destroyed. At this point, the high-throughput load job task is completed.

It can be understood that when the present embodiment determines that the job task is of the high-throughput load type, it is considered that only relying on the resident resource management instance with a fixed depth of 1 will result in the inability to increase the resource management instance through instance nesting, thereby being limited by throughput and affecting the execution of the high-throughput task. At this time, the resident resource management instance only provides resource allocation for the high-throughput task, and starts the parent instance and the child instance 2-layer temporary resource management instance to manage and execute the high-throughput task, that is, through instance nesting, the depth of the resource management system is correspondingly increased from 1 to 3. By adopting a 3-layer deep resource management system for task allocation and execution, the throughput of the resource management system can be improved, thereby achieving efficient management of the high-throughput load.

In this embodiment, the job information of high-throughput tasks is stored in a temporary resource management instance global database. Since the temporary resource management instance is temporary, when a temporary resource management instance exits, the job information it manages is also deleted. To address this issue, the job monitoring/resubmission component of each high-throughput load manager regularly collects the real-time job information of each high-throughput and updates it to the temporary resource management instance global database; the temporary resource management instance global database runs on the login node for a long time and does not change its state with the startup and exit of the temporary resource management instance. Therefore, the temporary resource management instance global database can timely update and store the high-throughput load information managed by all currently running and historically running temporary resource management instances for a long time for user query, avoiding information loss;

In this embodiment, the job monitoring/resubmission component in the high-throughput load manager monitors whether the high-throughput load job task is successfully completed. When a high-throughput task with running errors is monitored, the job monitoring/resubmission component resubmits the erroneous high-throughput task to the temporary resource management instance for execution, thereby providing a certain fault tolerance for the operation of the high-throughput load.

As shown in FIG4 , in this embodiment, the method further includes querying the operation task information, and the operation task information query includes the steps of:

Step S401, the resource management system receives a query request and determines the type of the job task according to the job number of the job task to be queried in the query request;

Step S402: If it is determined that the job task to be queried is a high-performance load, the resource management system calls the user interface of the resident resource management instance located on the computer login node through the unified interface to issue a job task information query request, and the control node agent process of the resident resource management instance located on the computer control node parses the job task information query request, and queries the relevant information of the job task from the local data structure in the resource management system and returns it to the unified interface;

Step S403: If it is determined that the job task to be queried is a high-throughput load job task, the resource management system queries the relevant information of the job task according to the job number through the temporary resource management instance global database, and returns it to the unified interface.

Specifically, in step S401, the user calls the unified interface of the variable depth resource management system for multiple scenarios on the login node to submit a load query request. Since both the high-performance load and the high-throughput load correspond to a job number, the query command includes the job number to be queried; after step S403, it optionally also includes the step of: the unified interface prints the job information to the screen for user viewing.

It can be understood that setting up a temporary resource management instance global database to store information data of all high-performance tasks and high-throughput tasks, and querying through the unique job number corresponding to each task, can pay more attention to the running status of the job task in a more timely manner, and can also view the corresponding data at any time after the execution is completed in case of need.

This embodiment also provides a variable depth resource management system for multiple scenarios, and the resource management system includes:

The resource management system interface module is used to receive high-performance loads or high-throughput loads submitted by users and determine the load type;

The high-performance load job task management module is used to allocate resources to the high-performance load job task by calling the resident resource management instance in the resource management system if the load type is determined to be a high-performance load, and after the resource allocation is completed, control the execution of the high-performance load job task through the resident resource management instance, and the resident resource management instance has a fixed depth;

The high-throughput load job task management module is used to start the high-throughput load manager through the unified interface in the resource management system if the load type is determined to be a high-throughput load. The high-throughput load manager parses the high-throughput load job task and splits it into multiple sub-task sets. Resources are allocated through the resident resource management instance. After the resource allocation is completed, a temporary resource management parent instance and multiple temporary resource management child instances are started through the resident resource management instance according to the number of sub-task sets of the high-throughput load job task. The temporary resource management parent instance and each temporary resource management child instance control the execution of the high-throughput load job task. The temporary resource management parent instance and the temporary resource management child instance high-throughput load manager are temporarily deployed and destroyed with the high-throughput load to achieve variable depth.

This embodiment also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned variable depth resource management method for multiple scenarios is implemented.

The system and medium of this embodiment correspond to the above method and also have the advantages described in the above method.

The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiment can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable media include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium. The memory is used to store computer programs and/or modules. The processor implements various functions by running or executing computer programs and/or modules stored in the memory, and calling data stored in the memory. The memory may include a high-speed random access memory and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims (10)

1. The variable depth resource management method for multiple scenes is characterized by comprising the following steps:

receiving high-performance load or high-flux load submitted by a user and judging the load type;

If the load type is judged to be high-performance load, carrying out resource allocation on the high-performance load job task by calling a resident resource management instance in a resource management system, and controlling to execute the high-performance load job task by the resident resource management instance after the resource allocation is completed, wherein the resident resource management instance is of a fixed depth;

If the load type is judged to be high-flux load, a high-flux load manager is started through a unified interface in a resource management system, the high-flux load manager analyzes a high-flux load job task and splits the high-flux load job task into a plurality of subtask sets, resource allocation is carried out through the resident resource management instance, after the resource allocation is completed, a temporary resource management father instance and a plurality of temporary resource management child instances are started through the resident resource management instance according to the number of the subtask sets of the high-flux load job task, the high-flux load job task is controlled and executed by the temporary resource management father instance and each temporary resource management child instance, and the temporary resource management father instance, the temporary resource management child instance and the high-flux load manager are temporarily deployed and destroyed along with the high-flux load so as to realize variable depth.

2. The multi-scenario-oriented variable depth resource management method of claim 1, wherein:

And if the load type is judged to be the high-performance load, carrying out resource allocation on the high-performance load job task by calling a resident resource management instance in the resource management system, wherein the method comprises the following steps of:

A unified interface in the resource management system calls a user interface of a resident resource management instance positioned on a computer login node, and a resident resource management instance control node agent process positioned on a computer control node requests resource allocation;

The resident resource management instance control node proxy process analyzes the high-performance load job task to schedule corresponding system resources to be allocated to the current job task, and returns a resource allocation result to a resident resource management instance user interface on the login node;

the resident resource management instance controls the node proxy process to allocate a unique job number for the current high-performance load job task.

3. The multi-scenario oriented variable depth resource management method of claim 2, wherein:

And if the load type is judged to be a high-performance load, controlling the execution of the high-performance load job task through the resident resource management instance, wherein the method comprises the following steps of:

the resident resource management instance control node proxy process sends the related information of the high-performance load job task to the resident resource management instance computing node proxy process on the computing node corresponding to the current job task through the user interface;

the resident resource management instance computing node proxy process loads and runs the high-performance load job task;

in the operation process of the high-performance load job task, the resident resource management instance monitors the operation state, records the state monitored in real time on the resident resource management instance control node proxy process, and recovers the resources allocated to the current high-performance load job task through the resident resource management instance after the operation of the current high-performance job task is finished, so that the operation of the high-performance load is completed.

4. The multi-scenario-oriented variable depth resource management method of claim 1, wherein:

And if the load type is judged to be a high-flux load, performing resource allocation through the resident resource management instance, wherein the method comprises the following steps:

Analyzing the total resource requirement of the current high-flux load by a high-flux load manager, determining the total number of required computing nodes, splitting the high-flux tasks in the high-flux load into a plurality of subtask sets by the high-flux load manager, and generating a subtask set loading script for each subtask set;

the high-throughput load manager calls a user interface of the resident resource management instance, and sends a resource allocation request to a resident resource management instance control node proxy process;

after the resident resource management instance control node proxy process receives the resource allocation request, the resource and the unique job number on the computer computing node are allocated for the high-throughput load job task, and the resource and the unique job number are returned to the high-throughput load manager.

5. The multi-scenario oriented variable depth resource management method of claim 4, wherein:

And if the load type is judged to be high-throughput load, starting a temporary resource management father instance and a plurality of temporary resource management child instances according to the number of the child task sets of the high-throughput load job task through the resident resource management instance, and controlling the execution of the high-throughput load job task through the temporary resource management father instance and each temporary resource management child instance, wherein the method comprises the following steps:

The resident resource management instance sends the current job related information to each computing node allocated to the current high-throughput load through a user interface on the login node;

On each computing node allocated to the current task, starting a temporary resource management father instance by a resident resource management instance computing node proxy process, wherein the temporary resource management father instance is a first layer temporary resource management instance and consists of one temporary resource management instance proxy process on each allocated node for managing all resources allocated to the current high-throughput load;

The temporary resource management father instance starts a plurality of temporary resource management sub-instances, each temporary resource management sub-instance manages the computing resources of a part of computing nodes, and runs a corresponding sub-task set loading script, and when all high-flux tasks in the sub-task set loading script corresponding to the temporary resource management sub-instance are run, the temporary resource management sub-instance exits;

In the running process of the high-flux load job task, the high-flux load manager continuously monitors real-time state information of the high-flux load job task according to the unique job number and stores the real-time state information into a temporary resource management instance global database in the resource management system.

6. The multi-scenario oriented variable depth resource management method of claim 5, wherein: the temporary resource management parent instance starts M temporary resource management child instances, M is the number of split high-throughput child task sets, each temporary resource management child instance comprises N/M proxy processes so as to correspond to N/M computing nodes, each temporary resource management child instance manages the computing resources of the N/M computing nodes, N is the number of computing nodes distributed to the current high-throughput load, and when a corresponding high-throughput task loading script is operated, the high-throughput task is scheduled to the managed computing resources to operate according to a first-in first-out (FIFO) scheduling strategy.

7. The multi-scenario-oriented variable depth resource management method according to any one of claims 1 to 6, wherein:

If the load type is judged to be high-throughput load, a temporary resource management father instance and a plurality of temporary resource management child instances are started according to the number of the child task sets of the high-throughput load job task through the resident resource management instance, and the high-throughput load job task is controlled to be executed by the temporary resource management father instance and each temporary resource management child instance, and the method further comprises the following steps:

When the high-flux load manager monitors that part of tasks in the high-flux load job task run wrong, judging whether the task with the running wrong is the first running error, if so, generating a new task script, starting a new temporary resource management sub-instance under the temporary resource management father instance, and submitting the task with the running wrong to the new temporary resource management sub-instance for running; reporting error information if the first run is not wrong;

when the high-flux load manager monitors that all tasks in the high-flux load job tasks are correctly operated, destroying the temporary resource management parent instance, recovering the resources allocated to the high-flux load by the resident resource management instance, and exiting the high-flux load manager to finish the operation of the high-flux load.

8. The multi-scenario-oriented variable depth resource management method of claim 1, wherein:

the method also comprises a job task information query, wherein the job task information query comprises the following steps:

The resource management system receives the query request and judges the type of the job task according to the job number of the job task to be queried in the query request;

If the job task to be queried is judged to be a high-performance load, the resource management system calls a user interface of a resident resource management instance positioned on a computer login node through a unified interface to send a job task information query request, a control node proxy process of the resident resource management instance positioned on a computer control node analyzes the job task information query request, and related information of the query job task in a local data structure in the resource management system is returned to the unified interface;

if the job task to be queried is judged to be a high-throughput load, the resource management system queries related information of the job task according to the job number through the temporary resource management instance global database and returns the related information to the unified interface.

9. A multi-scenario oriented variable depth resource management system, the resource management system comprising:

The resource management system interface module is used for receiving high-performance load or high-flux load submitted by a user and judging the load type;

The high-performance load job task management module is used for carrying out resource allocation on the high-performance load job task by calling a resident resource management instance in the resource management system if the load type is judged to be a high-performance load, and controlling the execution of the high-performance load job task by the resident resource management instance after the resource allocation is completed, wherein the resident resource management instance is of a fixed depth;

And the high-throughput load job task management module is used for starting a high-throughput load manager through a unified interface in the resource management system if the load type is judged to be high-throughput load, analyzing the high-throughput load job task by the high-throughput load manager and splitting the high-throughput load job task into a plurality of subtask sets, carrying out resource allocation through the resident resource management instance, starting a temporary resource management father instance and a plurality of temporary resource management child instances through the resident resource management instance according to the number of the subtask sets of the high-throughput load job task after the resource allocation is completed, and controlling the temporary resource management father instance and each temporary resource management child instance to execute the high-throughput load job task, wherein the temporary resource management father instance and the temporary resource management child instance high-throughput load manager are temporarily deployed and destroyed along with the high-throughput load so as to realize variable depth.

10. A computer-readable storage medium storing a computer program, characterized by: the computer program, when executed by a processor, implements the method of any one of claims 1 to 8.