Multitasking scheduling method for gpu

Abstract

The invention discloses a multi-task scheduling method for GPU. The method includes: when each clock cycle arrives, the streaming multiprocessor continues to schedule and execute the current thread warp; if the current thread warp cannot be scheduled for execution, then Select the task with the highest priority that can be selected among all the tasks to form a task set; select a task that is scheduled to execute the least number of times from the task set; schedule and execute the oldest thread warp in the selected task. The multi-task scheduling method for GPU of the present invention is executed by preferentially scheduling the same thread warp. When the thread warp is blocked, tasks with higher priority are scheduled and executed according to the priority order of the tasks, and tasks with the same priority are scheduled and executed. Scheduling and executing tasks that are scheduled to execute fewer times can not only consider the scheduling priority of high-priority tasks, but also consider the scheduling fairness among tasks of the same priority, and can better cover up the access of tasks. storage delay.

Description

technical field [0001] The invention relates to the technical field of GPUs, in particular to a multi-task scheduling method for GPUs. Background technique [0002] Graphics Processing Unit (GPU) is a microprocessor used to do image and graphics-related calculations. GPU is widely used in cloud computing platforms and data centers because of its powerful computing capabilities, providing users with the calculation. Compared with a single-task GPU that only runs one task on the GPU, a multi-task GPU can run multiple tasks on the GPU at the same time, which can effectively improve resource utilization. Specifically, a multi-task GPU can simultaneously run computation-intensive and storage-intensive programs on one GPU, and the computation resources and storage resources on the GPU can be fully utilized at the same time. [0003] Currently, there are two ways to support the GPU to run multiple tasks at the same time. The first way is spatial multitasking, and the second way i...

AI Technical Summary

Problems solved by technology

[0006] However, the LRR scheduling algorithm and GTO scheduling algorithm adopted on the existing GPU only consider the fairness and efficiency of scheduling between different warps of the same task in a single-task GPU, and do not take into account the multi-task scenarios. Scheduling fairness among multiple tasks of different priorities, and scheduling differences among multiple tasks with different priorities

Among them, although the GTO scheduling algorithm can better cover the memory access delay of the GPU program and has high performance, it continuously schedules several older thread warps, which obviously ignores the fairness among other tasks at the same priority. performance, and the scheduling priority of high-priority tasks

Although the LRR scheduling algorithm can treat different warps of the same task equally when performing a single task, in a multi-task scenario, due to the difference in the number of warps between different tasks with the same priority, the LRR scheduling algorithm cannot bring Better fairness; moreover, the LRR scheduling algorithm cannot meet the requirements of different priorities among multiple tasks and cannot cover up the memory access delay like the GTO scheduling algorithm

Images