TW201415409A - System and method of dynamic task allocation - Google Patents

Abstract

The present invention provides a system and method of dynamic task allocation, the method includes: evaluating actual computing ability of GPU and CPU; separating a task into N subtasks; determining a subtask to be a allocating task; computing a first estimated completion time according to the actual computing ability of the GPU, the first estimated completion time indicating that the time needed for the allocating task computing by the GPU; computing a second estimated completion time according to the actual computing ability of the CPU, the second estimated completion time indicating that the time needed for the allocating task computing by the CPU; putting the first and the second estimated completion time in sequence in accordance with the length the time; allocating the task to a task queue according the result of the sequence. The present invention can be used to allocate tasks for GPU and CPU.

Description

Dynamic task allocation system and method

The invention relates to a computer dynamic task allocation system and method.

At present, when a computer based on the CUDA (Compute Unified Device Architecture) architecture performs task assignment, tasks are assigned according to a fixed allocation mode. However, due to the huge computing power difference between different CPUs and different GPUs, and the number of CPUs and the number of display cards that each CPU can control are different, the memory capacity and speed of each display card are different. Obviously, the current CUDA architecture does not take into account this heterogeneous state, and the task allocation method used does not make good use of the CPU and GPU computing processing capabilities.

In view of the above, it is necessary to provide a dynamic task allocation system and method that efficiently allocates tasks to the GPU and CPU for processing.

The dynamic task allocation system includes: an evaluation module for evaluating actual computing power of the GPU and the CPU; and a decomposition module for decomposing the new task into N sub-tasks, wherein N is an integer greater than or equal to 1 a determining module, configured to determine, from the decomposed subtask, an item to be a subtask to be allocated; and a computing module, configured to calculate the subtask to be allocated according to an actual computing capability of the GPU when the subtask to be allocated is executable by the GPU a first expected completion time required by the GPU to execute by the GPU, the first expected completion time being equal to a time required by the GPU to execute the subtask and a pending task in the current task queue of the GPU a sum of time required; the computing module is further configured to calculate, according to an actual computing capability of the CPU, a second estimated completion time required for execution by the CPU for the subtask to be allocated, the second expected completion The time is equal to the sum of the time required by the CPU to execute the subtask and the time required by the CPU to execute the task to be processed in the current task queue; the sorting module is configured to use the subtask to be assigned First Expected completion time of the second sorted according to length; and dispensing module for dispensing according to the ranking result will be assigned sub-task to perform subtask is estimated completion time required for the shortest task queue.

The dynamic task allocation method includes: an evaluation step of evaluating actual computing power of the GPU and the CPU; and an decomposition step of decomposing the new task into N sub-tasks, wherein N is an integer greater than or equal to 1; The subtask obtained by the decomposition determines an item as a subtask to be allocated; the first calculation step, when the subtask to be allocated is executable by the GPU, calculates the subtask required by the GPU according to the actual computing power of the GPU. a first expected completion time, the first estimated completion time being equal to a sum of a time required by the GPU to execute the subtask and a time required by the GPU to perform a task to be processed in the current task queue; a second calculation a step of calculating, according to an actual computing capability of the CPU, a second expected completion time required by the CPU for the subtask to be allocated, the second estimated completion time being equal to a time required by the CPU to execute the subtask a sum of time required for the CPU to execute a task to be processed in its current task queue; a sorting step of following the first and second expected completion times of the subtask to be assigned The length between the sorting; and assigning step assigned to perform subtask shortest time required for completion queue task according to the ranking result subtasks to be dispensed.

Compared with the prior art, the dynamic task allocation system and method can effectively allocate tasks to the GPU and the CPU for processing, and fully utilize the computing processing capability of the GPU and the CPU, thereby improving the task execution efficiency of the computer.

As shown in FIG. 1, it is an operating environment diagram of the dynamic task distribution system of the present invention. In this embodiment, the dynamic task distribution system 10 runs in a computer 100 having a CUDA (Compute Unified Device Architecture) architecture, and the computer 100 includes at least one GPU (Graphic Processing Unit) 20 and at least one CPU (Central Processing Unit). 30 and memory 40. The dynamic task assignment system 10 is stored in the memory 40 for effectively allocating new tasks to the GPU 20 and the CPU 30 for processing when the computer 100 has a new task to process.

In this embodiment, the dynamic task distribution system 10 includes an evaluation module 101, an decomposition module 102, a determination module 103, a calculation module 104, a sequencing module 105, an allocation module 106, an identification module 107, and a determination module. Group 108 (see Figure 2). The module referred to in the present invention is a program segment for performing a specific function, and the functions of each module will be specifically described in the flowchart of FIG.

3 is a flow chart of a preferred embodiment of the dynamic task assignment system of the present invention. In order to clearly illustrate the present invention, the computer 100 includes a GPU 20 and a CPU 30. In other embodiments, the computer 100 may also include a plurality of GPUs 20 and a plurality of CPUs 30.

In step S1, the evaluation module 101 evaluates the actual computing power of the GPU 20 and the CPU 30. It should be noted that this step can be calculated by using the peak value of the industry standard GFLOPS (Floating Point Operations Per Second), or the computing power of the GPU 20 and CPU 30 chip manufacturers. Estimate the law. For example, it is estimated that the computing power of the GPU 20 is 200 GFLOPS, and the computing power of the CPU 30 is 20 GFLOPS.

For the evaluation of the actual computing power of the GPU 20 and the CPU 30, this step can be performed once the computer 100 is booted into the operating system.

In step S2, the decomposition module 102 decomposes the new task into N items of subtasks, and N is an integer greater than or equal to 1. For example, the new task M can be decomposed into two parallel sub-tasks M1, M2 and sub-tasks M3 that need to wait for M1 and M2 to execute after performing, for example, according to the prior art data parallelism and task parallel decomposition principle.

In step S3, the determining module 103 determines an item from the decomposed subtasks as a subtask to be allocated. For example, if the new task M is decomposed into the parallel subtasks M1 and M2 and the subtask M3 that needs to wait for the execution of the M1 and M2, the determination module 103 determines that the subtasks M1 and M2 need to be executed first, and because the subtasks are executed. The M1 and the M2 can be processed in parallel, and the determining module 103 can randomly select one of the subtasks, for example, M1, from the parallel subtasks to perform the allocation, that is, determine the subtask M1 as the subtask to be allocated.

In step S4, the calculation module 104 calculates, according to the estimated actual computing capability of the GPU 20, the first estimated completion time required for the subtask M1 to be allocated to be executed by the GPU 20; the CPU 30 obtained according to the evaluation. The actual computing power calculates the second expected completion time required for the subtask M1 to be allocated to be executed by the CPU 30. The first expected completion time is equal to the sum of the time required for the GPU 20 to execute the subtask to be allocated and the time required by the GPU 20 to process the to-be-processed task in its current task queue, the GPU The pending task in the current task queue is the task that has been assigned to the task queue of the GPU 20 but has not yet been executed; the second estimated completion time is equal to the CPU 30 executing the subtask to be assigned. The sum of the time required and the time required by the CPU 30 to process the task to be processed in its current task queue, the task to be processed in the current task queue of the CPU 30 is the task that has been assigned to the CPU 30 A task that has been queued but not yet executed.

For example, the time required for the GPU 20 to execute the subtask M1 is 0.01 seconds according to the actual computing power budget of the GPU 20, and the GPU 20 performs 11 seconds of the pending task of the current task queue ( Here, the required time can be calculated by running the data prepared in advance when the program is started, and the first expected completion time when the subtask M1 is executed by the GPU 20 is 11.01 seconds.

For example, according to the actual computing power of the CPU 30, it is calculated that the time required for the CPU 30 to execute the subtask M1 is 0.1 second, and the CPU 30 executes the task to be processed of its current task queue for 10 seconds. The second estimated completion time when the subtask M1 is executed by the CPU 30 is 10.1 seconds.

It should be noted that if the computing module 104 is calculating the first expected completion time of a sub-task (ie, the estimated completion time when the GPU 20 is executed), it is found that the sub-task cannot be When the task performed by the GPU 20 is, for example, a logical determination, the computing module 104 can directly set the first expected completion time of the subtask to be equal to infinity, and then calculate the sub-operation according to the actual computing power of the CPU 30. The second expected completion time when the task is executed by the CPU 30.

In step S5, the sorting module 105 sorts the first and second estimated completion times of the subtasks to be allocated.

For example, the first estimated completion time performed by the GPU 20 in the subtask M1 calculated in the step S4 and the second estimated completion time executed by the CPU 30 are sorted in order from short to long, obviously, The second expected completion time (10.1 seconds) when the subtask M1 is executed by the CPU 30 is shorter than the first expected completion time (11.01 seconds) when the subtask M1 is executed by the GPU 20.

In step S6, the distribution module 106 allocates the subtask to be assigned to the task queue with the shortest expected completion time required to execute the subtask.

For example, the subtask M1 is assigned to the task queue of the CPU 30 in accordance with the sort result of the step S5.

It should be noted that, if the first estimated completion time when a subtask is executed under the GPU 20 is equal to the second expected completion time when the CPU 30 is executed, the distribution module 106 will Subtasks are assigned to the task queue of the GPU 20. The reason for this allocation is that the GPU 20 cannot handle the scheduled job tasks of the operating system, so as much as possible, more computing tasks are handed over to the GPU 20 for processing to preserve the CPU 30 general processing power.

In step S7, if the subtask is assigned to the task queue of the GPU 20, the identification module 107 identifies the time required for the GPU 20 to execute the subtask. If the subtask is assigned to the task queue of the CPU 30, the identification module 107 identifies the time at which the CPU 30 executes the subtask. For example, after the subtask M1 is assigned to the CPU 30, the time required for the CPU 30 to process the task whose task name is M1 is 0.1 second. The purpose of this step is to facilitate the step S4 to calculate the time required by the GPU 20 or the CPU 30 to perform the task to be processed of its current task queue when calculating the first and second expected completion times of the subtasks.

In step S8, the determining module 108 determines whether there are other subtasks that have not been allocated yet, and if so, executes step S3, and if not, ends the process. For example, if the subtasks M2 and M3 have not completed the allocation, the process returns to the process step S3, and the determination module 103 determines the next subtask to be assigned. Since M2 and M1 are sub-tasks that can be paralleled, the determining module 103 can determine that the subtask M2 is the next subtask to be assigned. It should be noted that, since the subtask M3 needs to wait for the execution of the M1 and the M2, the determining module 103 needs to determine that the M1 and the M2 are executed before determining that the M3 is the subtask to be allocated.

It should be noted that, if the computer 100 includes multiple GPUs 20 and/or multiple CPUs 30, when the first estimated completion time and the second estimated completion time of the subtasks are calculated in step S4, the subtasks are calculated separately. The estimated completion time required for GPU 20 and CPU 30 to execute.

For example, for example, the computer 100 includes two GPUs 20 with serial numbers I and II, and three CPUs 30 with serial numbers A, B, and C, respectively, and the serial numbers obtained according to step S1 are respectively determined. The first expected completion time and the second expected completion time of the subtasks, such as M1, are calculated for the computing power of the GPU 20 of I, II and the CPU 30 of serial numbers A, B, and C, respectively.

For example, it is calculated that the first estimated completion time (I-GPU 20) when the subtask M1 is executed by the GPU 20 of the serial number I is 8.5 seconds, and the first expected completion time when executed by the GPU 20 of the serial number II (II) - GPU 20) is 9.3 seconds. The second expected completion time (A-CPU 30) when executed by the CPU 30 of sequence number A is 9.1 seconds, and the second expected completion time (B-CPU 30) when executed by the CPU 30 of sequence number B is At 9.2 seconds, the second expected completion time (C-CPU 30) when executed by the CPU 30 of serial number C is 8.6 seconds. Obviously, when the sorting of the expected completion time is performed in step S5, the subtask M1 is the shortest expected completion time when the GPU 20 of the serial number I is executed, and the subtask M1 is assigned to the serial number in step S6. The task of GPU 20 for I is listed.

It should be noted that the above embodiments are only for explaining the technical solutions of the present invention and are not intended to be limiting, and the present invention will be described in detail with reference to the preferred embodiments. Modifications or equivalents are made without departing from the spirit and scope of the invention.

100. . . computer

10. . . Dynamic task allocation system

20. . . GPU

30. . . CPU

40. . . Memory

101. . . Evaluation module

102. . . Decomposition module

103. . . Determine module

104. . . Computing module

105. . . Sorting module

106. . . Distribution module

107. . . Identification module

108. . . Judging module

S1. . . Evaluate the actual computing power of the GPU and CPU

S2. . . Decompose new tasks into N subtasks

S3. . . Determine the subtask to be assigned

S4. . . Calculate the estimated completion time required for the subtask to be allocated to be executed by the GPU and CPU respectively

S5. . . Sorting the expected completion time of the subtask to be assigned

S6. . . Assign the subtask according to the sort result

S7. . . Identifies the time required for the GPU or CPU to perform this subtask

S8. . . Are there other subtasks that have not been assigned yet?

1 is a diagram showing the operating environment of the dynamic task distribution system of the present invention.

2 is a functional block diagram of a dynamic task assignment system of the present invention.

3 is a flow chart of a preferred embodiment of the dynamic task assignment method of the present invention.

S1. . . Evaluate the actual computing power of the GPU and CPU

S2. . . Decompose new tasks into N subtasks

S3. . . Determine the subtask to be assigned

S4. . . Calculate the estimated completion time required for the subtask to be allocated to be executed by the GPU and CPU respectively

S5. . . Sorting the expected completion time of the subtask to be assigned

S6. . . Assign the dice task according to the sort result

S7. . . Identifies the time required by the GPU or CPU to perform this dice task

S8. . . Are there other subtasks that have not been assigned yet?

Claims (10)

A dynamic task allocation system, the system comprising:
Evaluation module for evaluating the actual computing power of the GPU and the CPU;
The decomposition module is configured to decompose the new task into N items of subtasks, where N is an integer greater than or equal to 1;
Determining a module for determining an item from the decomposed subtask as a subtask to be allocated;
a computing module, configured to: when the subtask to be allocated is executable by the GPU, calculate a first estimated completion time required for the subtask to be executed by the GPU according to an actual computing capability of the GPU, where the first estimated completion time is equal to the The sum of time required by the GPU to perform the subtask and the time required by the GPU to perform the task to be processed in its current task queue;
The calculation module is further configured to calculate, according to an actual computing capability of the CPU, a second estimated completion time required for execution by the CPU for the subtask to be allocated, where the second estimated completion time is equal to the CPU executing the The sum of the time required for the item task and the time required by the CPU to execute the task to be processed in its current task queue;
a sorting module, configured to sort the first and second estimated completion times of the subtasks to be allocated according to the length of time; and an allocation module, configured to allocate the subtasks to be assigned to the execution according to the sorting result This subtask requires a task with the shortest expected completion time. The dynamic task assignment system of claim 1, wherein if the first expected completion time and the second estimated completion time of the subtask to be allocated are equal, the distribution module assigns the subtask to the The task queue of the GPU. The dynamic task assignment system of claim 1 or 2, further comprising an identification module, configured to identify the GPU after the subtask to be allocated is assigned to a task queue of the GPU or the CPU Or the time required by the CPU to perform this subtask. The dynamic task assignment system of claim 1 or 2, when the subtask to be allocated cannot be executed by the GPU, the computing module sets the first expected completion of the subtask Time is infinity. For example, in the dynamic task assignment system described in claim 1, the decomposition module decomposes a new task according to a decomposition principle of data parallelism superior to task parallelism. A dynamic task allocation method, the method comprising:
Evaluation steps to evaluate the actual computing power of the GPU and CPU;
Decomposing steps to decompose the new task into N items of subtasks, where N is an integer greater than or equal to 1;
Determining a step of determining an item from the decomposed subtask as a subtask to be assigned;
a first calculating step, when the subtask to be allocated is executable by the GPU, calculating a first estimated completion time required for the subtask to be executed by the GPU according to an actual computing capability of the GPU, where the first estimated completion time is equal to the GPU The sum of the time required to perform the subtask and the time required by the GPU to perform the pending task in its current task queue;
a second calculating step of calculating, according to an actual computing capability of the CPU, a second estimated completion time required for execution by the CPU, the second estimated completion time is equal to the CPU executing the subtask The sum of the time required and the time required by the CPU to execute the task to be processed in its current task queue;
a sorting step of sorting the first and second estimated completion times of the subtasks to be allocated according to the length of time; and assigning, assigning the subtasks to be allocated according to the sorting result to the subtasks required to execute the subtasks The task with the shortest completion time is expected. The dynamic task assignment method of claim 6, wherein if the first estimated completion time and the second estimated completion time of the subtask to be allocated are equal, the subtask is assigned to the The task queue of the GPU. The method for assigning dynamic tasks as described in claim 6 or 7, wherein the method further comprises the step of identifying:
After the subtask to be allocated is allocated to the task queue of the GPU, the time required for the GPU to execute the subtask is identified;
After the subtask to be allocated is assigned to the task queue of the CPU, the time required for the CPU to execute the subtask is identified. The dynamic task assignment method according to claim 6 or 7, wherein when the subtask to be allocated cannot be executed by the GPU, the first step of the subtask is set in the first calculating step. The expected completion time is infinity. For example, in the dynamic task assignment method described in claim 6, the decomposition step decomposes a new task according to a decomposition principle of data parallelism superior to task parallelism.