Job scheduling optimization method based on Burst Buffer resource reservation

Abstract

The invention belongs to the technical field of super computers, and discloses a job scheduling optimization method based on Burst Buffer resource reservation, a Burst Buffer resource reservation mechanism is introduced, the method specifically comprises the following steps: job submission and resource demand reservation, and the Burst Buffer resource reservation mechanism specifically comprises the following steps: a system performs job scheduling according to the demand of each job for computing resources and Burst Buffer; corresponding Burst Buffer resources are allocated to the Burst buffer resources in advance; according to a time window mechanism of job scheduling, a time window mode is adopted in the scheduling process; the job execution sequence adopts a traditional mode, and the execution sequence is further selected by comparing and optimizing index jobs; performing operation backfilling; the optimization target selects a job execution sequence with the best effect of any one or comprehensive optimization target. According to the method, the job scheduling efficiency and the utilization rate of Burst Buffer resources are improved in the scheduling process.

Description

Technical Field

[0001] The invention belongs to the technical field of supercomputers, and in particular relates to a job scheduling optimization method based on Burst Buffer resource reservation. Background Art

[0002] In high-performance computing (HPC) systems, with the rapid development of computing power, the processing capacity of computing tasks continues to improve. However, I / O operations (such as reading and writing data) often fail to keep pace with the improvement of computing power, resulting in a significant performance gap between computing power and I / O performance. This gap not only affects the execution efficiency of tasks, but also causes the waiting time of jobs to be too long, which in turn leads to inefficient use of resources. In order to alleviate this problem, Burst buffer technology came into being. Burst buffer is a high-speed cache storage device between computing nodes and storage systems. It is designed to solve the bottleneck problem of traditional I / O subsystems. By providing a high-throughput cache area, computing tasks can exchange data more efficiently during execution. Although Burst buffer technology can effectively improve I / O performance, its resource management and scheduling issues are still challenging.

[0003] Traditional job scheduling methods do not fully consider the resource reservation of the burst buffer, which leads to an unbalanced allocation of computing resources and I / O resources, resulting in a large number of resource gaps. These gaps not only waste precious computing resources, but also increase the delay of job backfilling, further affecting the execution efficiency of the job and the overall system performance.

[0004] Existing scheduling methods often rely on fixed orders and simple queuing mechanisms when dealing with task scheduling, and fail to flexibly adapt to resource reservation and dynamic scheduling requirements. Especially in high-performance computing systems, due to the complexity of jobs and the uncertainty of computing requirements, traditional scheduling methods cannot efficiently manage burst buffer resources in a dynamically changing environment. In addition, these methods are usually unable to break through the limitations of scheduling order. For example, backfilling jobs cannot change the execution order and start time of the original scheduled jobs, which leaves a lot of room for optimization in terms of resource usage and job scheduling.

[0005] There are a few studies on Burst buffer scheduling, which mainly add the Burst buffer resource reservation mechanism directly to the traditional scheduling method, take this resource into account in the resource demand, and then execute the job based on the traditional scheduling method, which still cannot effectively solve the problem of resource utilization. In addition, the existing scheduling algorithms also fail to focus on multiple optimization goals, such as the overall system throughput, job execution time, and efficient resource utilization. This makes it impossible for existing research to take into account multiple factors when facing complex job and resource scheduling, thus affecting the comprehensiveness and accuracy of scheduling. Summary of the invention

[0006] The purpose of the present invention is to provide a job scheduling optimization method based on Burst Buffer resource reservation in response to the above-mentioned problems, hoping to improve the order limitation problem of the existing scheduling method when processing task scheduling, and ensure that the execution performance of the job and the resource utilization of Burst Buffer are improved in a variety of job queue environments by focusing on multiple optimization goals.

[0007] The technical solution adopted by the present invention is as follows: a job scheduling optimization method based on Burst Buffer resource reservation, the optimization method introduces a Burst Buffer resource reservation mechanism, specifically including the following: Job submission and resource demand reservation: When submitting a job, the user needs to provide the required parameters for computing resources and Burst Buffer resources. The Burst Buffer resource reservation mechanism specifically includes the system allocating corresponding Burst Buffer resources to the job in advance based on the computing resources and Burst Buffer requirements of each job, and ensuring that these resources can be guaranteed before the job is executed. The time window mechanism of job scheduling, the scheduling process adopts the time window mode; Job scheduling method, the order of job execution is further selected by the job by comparing the optimization indicators; job backfilling; Optimization target selection, based on the system load formed by the job queue resource requirements, select the job execution order that best achieves any or a combination of optimization targets.

[0008] Furthermore, the time window mode means that within a fixed time interval, the jobs in the waiting queue will be fixed and the scheduling optimization will be performed; the size of the time window is set by the administrator.

[0009] It should be noted that the size of the time window is set by the administrator or operator according to specific needs.

[0010] Furthermore, the optimization indicators include: The total waiting time of the queue, which represents the sum of the waiting times of all jobs in the queue; The job deceleration ratio of the queue, which represents the sum of the deceleration ratios of all jobs in the waiting queue; The completion time of the queue, which is the difference between the end time of the last job in the waiting queue and the start time of the first job.

[0011] Furthermore, the reduction ratio calculation formula is:

[0012] in, is the reduction ratio, is the job waiting time, The time the job runs.

[0013] Furthermore, during the job backfilling process, a suitable job will be selected from each subsequent job in a determined order for early backfill scheduling; When a running job is finished and resources are released, if the resource requirement of the first job in the waiting queue cannot be executed, the first job with resource requirement suitable for backfilling will be selected from the subsequent waiting queue according to the determined order.

[0014] It is important to note that backfilling can still be done even if the backfilling job would delay the execution of the first job (the backfilling job takes too long to run).

[0015] Furthermore, for the three optimization indicators, each target is first standardized and dimensioned, and then the triple weight parameters are set , and , respectively corresponding to the weight parameters of each optimization index, find the minimum set of weighted optimization indexes, and the corresponding job execution order is the optimal order; in, The total waiting time of the corresponding queue, The job reduction ratio of the corresponding queue, The completion time of the corresponding queue.

[0016] Furthermore, the weight parameters are selected by the administrator according to specific needs to suit the current load and resource requirements.

[0017] Furthermore, the size of the time window is set according to the actual load of the system and the requirements of the operation.

[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. By introducing the Burst Buffer resource reservation mechanism, time window scheduling method, and backfill mechanism, the job scheduling efficiency and the utilization rate of Burst Buffer resources are improved during the scheduling process; 2. Break the delay limitations of traditional backfill scheduling and provide multiple queue optimization targets for administrators to select and evaluate; ensure that the execution performance of jobs is improved in a variety of job queue environments.

[0019] 3. When each job event occurs, the most appropriate job can be selected for scheduling based on the current system resource status and specific job requirements, thereby achieving efficient use of resources; 4. It can significantly reduce the gaps in computing resources and I / O resources, and significantly reduce the delay of job backfilling, thereby achieving high efficiency in job execution and maximum utilization of system resources. This flexible and efficient scheduling method gives the present invention an important technical advantage in the field of supercomputer job scheduling, and can effectively improve the overall performance of supercomputer systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a schematic diagram of the architecture of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] In high-performance computing (HPC) systems, with the rapid development of computing power, the processing capacity of computing tasks continues to improve. However, I / O operations (such as reading and writing data) often fail to keep pace with the improvement of computing power, resulting in a significant performance gap between computing power and I / O performance. This gap not only affects the execution efficiency of tasks, but also causes the waiting time of jobs to be too long, which in turn leads to inefficient use of resources. In order to alleviate this problem, Burst buffer technology came into being. Burst buffer is a high-speed cache storage device between computing nodes and storage systems. It is designed to solve the bottleneck problem of traditional I / O subsystems. By providing a high-throughput cache area, computing tasks can exchange data more efficiently during execution. Although Burst buffer technology can effectively improve I / O performance, its resource management and scheduling issues are still challenging.

[0024] Currently, most job scheduling methods in supercomputer environments rely on traditional scheduling methods, such as First Come First Served (FCFS) and Shortest Job First (SJF). These methods usually determine the scheduling priority based on the order in which jobs are submitted or the expected execution time of the jobs. However, traditional job scheduling methods fail to take into account the reservation and dynamic scheduling of Burst buffer resources, resulting in low job scheduling efficiency, and may even cause a large number of resource gaps and job delays, affecting the performance of the overall system.

[0025] Currently, the closest implementation scheme to the present invention is the optimization method based on backfill scheduling used in the field of high-performance computing. In these systems, backfill scheduling can be combined with the scheduling method of FCFS or SJF to reduce the waiting time of jobs and improve the efficiency of resource utilization by inserting smaller jobs into the idle time period of long-running jobs. This type of scheduling method relies on the management of job queues and determines the execution order of jobs based on their resource requirements and the submission time of the jobs.

[0026] Many traditional scheduling algorithms are often subject to order constraints when performing backfill scheduling. For example, jobs in backfill scheduling must strictly follow the execution order and start time of the original jobs and cannot be flexibly adjusted. Due to the high coupling between the order of job scheduling and resource allocation, the limitations of traditional scheduling methods have caused obvious constraints on system performance.

[0027] However, backfill scheduling also has obvious shortcomings. First, traditional backfill scheduling cannot handle the reservation of burst buffer resources, resulting in poor coordination between I / O resources and computing resources, and cannot fully utilize the efficient data transmission capabilities provided by the burst buffer. In addition, the job scheduling process only considers the demand for job computing resources. Jobs with high burst buffer requirements cannot be backfilled, resulting in serious job starvation. Secondly, traditional backfill scheduling usually only executes scheduling based on the predetermined order of jobs, and cannot flexibly respond to dynamic changes in the job execution process. For example, in backfill scheduling, once the execution order of jobs is determined, it often cannot be adjusted. This makes it impossible for the system to effectively optimize resource allocation when facing resource shortages or job changes.

[0028] There are a few studies on Burst buffer scheduling, which mainly add the Burst buffer resource reservation mechanism directly to the traditional scheduling method, take this resource into account in the resource demand, and then execute the job based on the traditional scheduling method, which still cannot effectively solve the problem of resource utilization. In addition, the existing scheduling algorithms also fail to focus on multiple optimization goals, such as the overall system throughput, job execution time, and efficient resource utilization. This makes it impossible for existing research to take into account multiple factors when facing complex job and resource scheduling, thus affecting the comprehensiveness and accuracy of scheduling.

[0029] like Figure 1 As shown, the present invention proposes a method for optimizing job scheduling based on Burst Buffer resource reservation. The core idea of ​​the method is to realize the reservation of Burst buffer resources and break the order limitation in traditional scheduling. Specifically, the present invention can select the most suitable job for scheduling according to the current system resource status and the specific needs of the job when each job event occurs, thereby realizing efficient utilization of resources. At the same time, the present invention focuses on multiple optimization goals to ensure that the execution performance of the job is improved in a variety of job queue environments.

[0030] Through this optimization method, the present invention can not only significantly reduce the gaps in computing resources and I / O resources, but also significantly reduce the delay of job backfilling, thereby achieving high efficiency of job execution and maximum utilization of system resources. This flexible and efficient scheduling method enables the present invention to have important technical advantages in the field of supercomputer job scheduling and can effectively improve the overall performance of supercomputer systems.

[0031] like Figure 1 and Figure 2 As shown in FIG, a method for optimizing job scheduling based on Burst Buffer resource reservation is introduced, which is based on the shared Burst Buffer architecture of a supercomputer. Figure 2 As shown in the figure, the scheduling of jobs in the waiting queue is optimized to improve the job scheduling efficiency and the resource utilization of the Burst buffer.

[0032] The optimization method introduces a Burst Buffer resource reservation mechanism, which specifically includes the following: Job submission and resource demand reservation: When submitting a job, the user needs to provide the demand parameters for computing resources and Burst Buffer resources. The Burst Buffer resource reservation mechanism specifically includes the system allocating corresponding Burst Buffer resources to the job in advance based on the computing resources and Burst Buffer requirements of each job. The time window mechanism of job scheduling, the scheduling process adopts the time window mode; Job scheduling method, the order of job execution is further selected by the job by comparing the optimization indicators; job backfilling; Optimization target selection, based on the system load formed by the job queue resource requirements, select the job execution order that best achieves any or a combination of optimization targets.

[0033] The specific technical solutions are as follows: In the scheduling process, considering the limited nature of Burst Buffer resources, a reservation mechanism for Burst Buffer resources is introduced. When each job is submitted, the system will allocate corresponding Burst Buffer resources in advance according to its demand for computing resources and Burst Buffer, and ensure that these resources can be guaranteed before the job is executed. The job scheduling process adopts the time window mode, that is, the waiting jobs are fixed within a fixed time interval, and then the scheduling is optimized, so that a better execution method can be selected within the limited jobs. The size of the time window is generally set by the administrator according to the system status.

[0034] The scheduling method consists of two parts: execution order and job backfilling; 1. When a user submits a job, he needs to provide the resource requirement parameters of the burst buffer so that during the job scheduling process, both computing resources and burst buffer resources can be reserved.

[0035] 2. The job execution order adopts the traditional method, such as first come first served, short job priority, or customized job execution order. After comparing the optimization indicators, the job selects the execution order. Once the execution order of the job is determined, the job execution priority is set according to the execution order.

[0036] 3. During the backfill process, appropriate jobs will be selected from each subsequent job in a determined order for early backfill scheduling. Specifically, when a running job ends and releases resources, if the resource requirement of the first job in the waiting queue cannot be executed, the first job with resource requirements suitable for backfilling will be selected from the subsequent waiting queue according to the determined order. Even if the backfilled job will delay the execution of the first job (the backfilled job runs too long), backfilling can still be achieved. Traditional job scheduling cannot achieve this step because it requires that the backfilled job cannot delay the execution of the first job.

[0037] Among them, the optimization indicators include three items: The total waiting time of the queue, which represents the sum of the waiting times of all jobs in the queue; The job deceleration ratio of a queue represents the sum of the deceleration ratios of all jobs in the waiting queue. By optimizing the deceleration ratio of jobs, the system can ensure the execution efficiency of jobs and reduce the "delay" time of jobs.

[0038] The formula for calculating the reduction ratio is:

[0039] in, is the reduction ratio, is the job waiting time, The time the job runs.

[0040] The completion time of the queue, which is the difference between the end time of the last job in the waiting queue and the start time of the first job.

[0041] The administrator can select the job execution order that best achieves any or a combination of optimization goals based on the job queue resource demand composition and system load. For the three optimization indicators, each goal is first standardized and unified, and then the triple weight parameter ( , , ), corresponding to the weight parameters of each optimization indicator, find the weighted and minimum set of optimization indicators, and the corresponding job execution order is the optimal. The weight parameters are determined by the administrator.

[0042] Example 1 like Figure 1 As shown, one embodiment of the present invention is a method for optimizing job scheduling based on Burst Buffer resource reservation, and its technical solution specifically includes: 1. Job submission and resource reservation: (1) Resource requirements when users submit jobs: When users submit jobs, they must provide the required computing resources and Burst Buffer resource requirements. Specifically, the job must not only specify how much computing resources are required, but also the required Burst Buffer capacity (such as storage space size and access speed requirements). The core purpose of this step is to allow the system to know the resource requirements of each job in advance before the job is executed, and to reserve resources during the scheduling process.

[0043] (2) Burst Buffer resource reservation mechanism: Based on the resource requirements submitted by the user, the system will allocate Burst Buffer resources for each job in advance. In this way, before the job starts executing, the system can ensure that the required resources have been allocated in place, avoiding resource shortages or resource waiting. Resource reservation is an optimization for the limited nature of Burst Buffer resources, ensuring that there will be no resource conflicts during scheduling.

[0044] 2. Job scheduling in time windows: Time window mechanism: During the job scheduling process, the system uses the "time window" mode to optimize the scheduling of jobs. Specifically, the system will "freeze" the jobs currently in the waiting queue at a fixed time interval (for example, every certain time), and then schedule them according to the job resource requirements within that period.

[0045] Setting the time window size: The time window size is set by the administrator based on the actual system load and job requirements. The time window size setting will directly affect the scheduling effect. For example, a smaller window can be scheduled more frequently, but may bring higher scheduling overhead; a larger window may miss some optimization opportunities.

[0046] 3. Job scheduling order and execution: Selection of job scheduling order: Within the time window, the system determines the execution order of jobs according to the pre-set scheduling method. You can choose traditional scheduling methods, such as "First Come First Served" (FCFS) or "Shortest Job First" (SJF), or you can choose the optimal scheduling order based on optimization indicators (such as waiting time, reduction ratio, etc.).

[0047] Job execution priority: Once the job execution order is determined, the system will assign a priority to each job based on that order. Jobs with higher priorities will be scheduled for execution first, while jobs with lower priorities will be executed later.

[0048] 4. Backfill mechanism: The backfill mechanism is that when a job is completed and resources are released during the job execution, if the job in the front of the waiting queue cannot be executed immediately, the system will select a job with appropriate resource requirements from the subsequent jobs in the waiting queue for "early backfill". Even if the execution of the backfill job may delay the execution of the first job, the system will still backfill. This can avoid idle job resources and improve system resource utilization.

[0049] Traditional job scheduling usually does not allow backfilling jobs to delay the execution of the first job. In the method of this embodiment, backfilling is still allowed even if the running time of the backfilling job is longer, which can effectively fill the idle period of resources. When selecting a backfilling job, the system will select the first suitable job from the subsequent jobs for backfilling according to the current resource demand and the order of the waiting queue, ensuring efficient use of resources.

[0050] 5. Optimization indicators: The total waiting time of the queue is the sum of the waiting time of all jobs in the queue. The waiting time of each job refers to the time the job is waiting to be executed. Optimizing this indicator can reduce the average waiting time of the job. The job deceleration ratio of the queue is the sum of the deceleration ratios of all jobs in the waiting queue. The deceleration ratio is a measure of the deceleration effect of each job. The formula for calculating the reduction ratio is:

[0051] in, is the reduction ratio, is the job waiting time, The time the job runs.

[0052] By optimizing the reduction ratio of the job, the system can ensure the efficiency of job execution and reduce the "delay" time of the job.

[0053] The completion time of the queue is the difference between the end time of the last job in the waiting queue and the start time of the first job. Optimizing this indicator can help shorten the processing time of the entire queue, thereby improving overall efficiency.

[0054] 6. Comprehensive optimization and weight parameters: Target standardization: When there are multiple optimization indicators, each indicator needs to be standardized first and converted into a unified dimension so that weighted calculation can be performed effectively.

[0055] Setting of weight parameters: The system will assign a weight parameter to each optimization indicator according to the administrator's settings. , and ,These weights indicate the relative importance of each indicator in the optimization goal, and administrators can adjust them according to the specific situation and choose the optimization goal that best suits the current system load and resource requirements.

[0056] Minimize weighted sum: The system calculates the optimal job execution order by taking a weighted sum of each optimization indicator. The goal is to minimize the weighted sum, thereby selecting the optimal job scheduling order and ensuring that the system maximizes job scheduling efficiency and resource utilization while meeting resource reservation requirements.

[0057] Through the above technical solution, the efficiency of job scheduling can be effectively improved, the waiting time can be reduced, and the resource utilization of BurstBuffer can be improved.

[0058] Example 2 like Figure 2As shown, another embodiment of the present invention is that a job scheduling optimization system based on Burst Buffer resource reservation adopts Lustre parallel file system. Lustre adopts distributed lock management mechanism and object storage technology to fragment and disperse large files and store them on multiple OSTs (object storage targets), thereby improving the concurrency of multi-user access and aggregated I / O bandwidth, and is suitable for large-scale file read and write operations. The computing node is separated from the storage node, and high-speed network interconnection is used to achieve efficient data transmission. The job scheduling system not only schedules computing tasks, but also is responsible for the reasonable allocation of I / O resources to ensure that the back-end storage will not become a performance bottleneck when encountering a large number of sudden I / O requests. The shared Burst Buffer node is a high-speed cache device located between the computing node and the back-end storage, which is specifically used to buffer and accelerate a large number of instantaneous I / O operations, can quickly respond to sudden I / O requests, reduce the waiting time of the computing node, and thus improve the overall job execution efficiency.

[0059] Figure 2 CN is the I / O (input / output) interaction channel between the computing node that executes the computing task and the external storage system. It includes two working modes: when the job requires cache optimization, the computing node reads the original data from the parallel file system through the I / O interface, and temporarily stores the high-frequency access intermediate results to the Burst Buffer node with cache characteristics to alleviate the direct access pressure of the parallel file system, and finally writes the final result to the parallel file system through the I / O interface to achieve persistent storage; if the job does not require Burst Buffer acceleration, the computing node directly completes the data reading and writing operations with the parallel file system through I / O.

[0060] Due to its limited resource capacity and high cost, the scheduling system must evaluate the I / O requirements of the task when allocating jobs, and prioritize the allocation of limited high-speed cache resources to tasks that are sensitive to I / O performance or have obvious I / O burst characteristics. During task execution, the resource allocation should be dynamically adjusted according to the real-time load to make full use of limited high-speed storage resources, ensure that each job obtains sufficient bandwidth at critical moments, and achieve efficient collaborative operation of computing and data storage.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing job scheduling based on Burst Buffer resource reservation, characterized in that: The optimization method introduces a Burst Buffer resource reservation mechanism, which specifically includes the following: Job submission and resource demand reservation: When submitting a job, the user needs to provide the demand parameters for computing resources and Burst Buffer resources. The Burst Buffer resource reservation mechanism specifically includes the system allocating corresponding Burst Buffer resources to the job in advance based on the computing resources and Burst Buffer requirements of each job. The time window mechanism of job scheduling, the scheduling process adopts the time window mode; Job scheduling method, the order of job execution is further selected by the job by comparing the optimization indicators; job backfilling; Optimization target selection, based on the system load formed by the job queue resource requirements, select the job execution order that best achieves any or a combination of optimization targets.

2. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 1, characterized in that: The time window mode means that within a fixed time interval, the jobs in the waiting queue will be fixed and the scheduling optimization will be performed; the size of the time window is set by the administrator.

3. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 1, characterized in that: The optimization indicators include: The total waiting time of the queue, which represents the sum of the waiting times of all jobs in the queue; The job deceleration ratio of the queue, which represents the sum of the deceleration ratios of all jobs in the waiting queue; The completion time of the queue, which is the difference between the end time of the last job in the waiting queue and the start time of the first job.

4. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 3, characterized in that: The reduction ratio calculation formula is: in, is the reduction ratio, is the job waiting time, The time the job runs.

5. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 1, characterized in that: During the job backfilling process, a suitable job will be selected from each subsequent job in a determined order for early backfill scheduling; When a running job is finished and resources are released, if the resource requirement of the first job in the waiting queue cannot be executed, the first job with resource requirement suitable for backfilling will be selected from the subsequent waiting queue according to the determined order.

6. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 3, characterized in that: For the three optimization indicators, each target is first standardized and unified, and then the triple weight parameters are set , and , respectively corresponding to the weight parameters of each optimization index, find the minimum set of weighted optimization indexes, and the corresponding job execution order is the optimal order; in, The total waiting time of the corresponding queue, The job reduction ratio of the corresponding queue, The completion time of the corresponding queue.

7. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 6, characterized in that: The weight parameters are selected by the administrator according to specific needs to suit the current load and resource requirements.

8. The method for optimizing job scheduling based on Burst Buffer resource reservation according to claim 2, characterized in that: The size of the time window is set according to the actual load of the system and the requirements of the operation.

Images