Research For GPU Memory Scheduling Based On Classification Of Application Access Behavior

As general-purpose graphics processing units make advances in high-performance computing,Graphics Processing Unit(GPU)make full use of computing resources by executing multiple applications in parallel.But in the memory system where multiple applications are executed in parallel,negative interactions may cause such applications interfering with one another,introducing serious fairness issues and deteriarating the system performance.Aiming at the memory access interference between applications and the inefficiency of current memory scheduling policy,we propose a new approach—Classification-Driven Memory Scheduling Based on Application Behavior(CDMS-AAB),that reduces inter-application access interference based on awaring the behavior of application,and improves system throughput and fairness.CDMS-AAB classifys the applications that with different memory access behavior,it mainly consisting of two parts: deciding application type and priority-based scheduling.The applicaion categorise decision is carried out by calculating the bandwidth utilization of the application,and is tagged as compute intensive or memory intensive.Priority-based scheduling schedules programs for priority scheduling according to its type.Under different types of applications,prefers requests from computationally intensive applications.Under multiple computationally intensive applications,prefers application with smaller misses per kilo instructions,and always prefers First Ready Round Robin First Come First Serve(FR-RR-FCFS)under multiple memory intensive applications.Finally,CDMS-AAB was compared with other two memory scheduling.The experiments results indicates that CDMS-AAB can reduce interferences between applications,and achieve higher system throughput and fairness.Averaged system throughput and fairness are improves by 10.1% and 11.7% over FR-FCFS and averaged system throughput and fairness are improves by 7.4% and 3.4% over FR-RR-FCFS.