Research On EDF Scheduling Strategy For GPGPUs With Spatial Resource Sharing

With the widespread applying of machine learning and multimedia processing applications,there are much stronger demand for computing.While computing demand is surging,CPU performance advances are slowing in post-Moore law period.GPU's blooming relieves the intension between increasing demand and slow hardware development.Real-time systems such as autonomous driving have also begun to use GPUs as coprocessors to increase computational efficiency.Therefore,real-time task scheduling research on GPUs has attracted industry and academic attention.Although NVIDIA provides interrupt functionality in the Pascal architecture and beyond,it does not support priority-based scheduling.In addition,the traditional EDF scheduling on the GPU does not take into account the internal resource utilization of a single SM in the GPU.Often,high-priority tasks monopolize the GPU for a long time,which affects the performance of the entire system scheduling.The mechanism of running multiple applications in parallel on a single SM core in the GPU provides an opportunity to increase the real-time scheduling efficiency on the GPU.We explored the EDF scheduling strategy based on spatial sharing,and designs an automated spatial sharing task scheduling framework based on EDF.By using the Markov chain offline,we use the application's computational instruction scale and memory access latency to statically calculate the impact factors between the two applications when using different resources,and combine the actual tests to get the worst performing IPC and obtain parallelism.The respective application's estimated IPC for both applications.And use this information in the dynamic running phase to determine whether different applications can occupy different GPU resources,and whether the remaining task amount can be completed before the deadline in the case of occupying so many resources.And in the case of ensuring that the task deadline is completed,the GPU utilization efficiency is improved by parallel execution.The experimental results show that for the various working sets composed of different types of applications,the proposed method has a maximum increase of 24.5%(TX2)and 22.6%(Xavier)compared with the traditional EDF-based preemptive scheduling.At the same time,the scheduling overhead of our framework is lower than 0.5%of the total running time.