Design And Implementation Of Workload Performance And Energy Optimization Based On CPU-GPU Heterogeneous System

With the rapid development of technologies such as cloud computing,big data,and artificial intelligence,CPU-GPU heterogeneous high-performance computing methods are increasingly used by data centers,because it has higher parallelism,peak calculations and flexibility.However,its peak power consumption also increased significantly.In a heterogeneous computing environment,when a single workload is running,maintaining the highest GPU core frequency may cause a large amount of energy waste when meeting the performance requirements;when a mixed workload is running,if the mixed workload compete for GPU resources,it will result in loss of performance.Therefore,it is necessary to properly schedule the workload to improve the overall efficiency and improve the energy efficiency without violating the performance constraints.To solve the above problems,first,through actual measurement methods,14 representative loads are run on a specific hardware platform to find the optimal frequency for a single workload operation;then DVFS is used to adjust the frequency to obtain the performance loss and energy consumption of the workload.Through experimental analysis,it is found that the actual scheduling effect obviously depends on the workload type,but the traditional memory/computation-bounded binary classification method is not suitable for such heterogeneous platforms.Based on this,classification training of actual test results is performed through key parameters(such as memory dependency stall,GPU utilization,GPU bandwidth,and workload execution time,etc.),thereby classifying the load into six categories.Compared with the traditional load classification effect,the load of each category exhibits similar performance and power consumption characteristics when using DVFS,so as to obtain a more accurate performance energy efficiency model.Based on this model,it is possible to effectively schedule the mixed workload and thereby increase the overall energy efficiency of the mixed workload.The experimental results show that the energy efficiency optimization program based on the new workload classification reduces the energy consumption of the system and does not result in significant performance degradation.DVFS scheduling for a single wrokload,the workload has an average performance loss of only 9.58%,saving 29.20% of the energy consumption;scheduling of mixed loads,a reasonable mixed workload combination scheme will make the performance increase of 19.54% to 22.24%.