A Genetic Algorithm Based Maximum Power Consumption Test Framework For GPGPU

Benefiting from powerful computation potential, Graphics Processing Units (GPUs) have become a prevalent computing platform for general-purpose parallel computing, the so-called GPGPU computing. While delivering high computation performance, GPGPUs are power hungry, which has seriously hindered the developing of GPGPUs. The maximum power consumption for which GPGPU systems are designed, called the TDP, is one of the most important parameters in the design of thermal module, cooling module and power supply system of GPGPU.The practically attainable maximum power consumption of GPGPU is extremely difficult to determine for its massive parallel cores, complex memory hierarchy and sophisticated inter-connection. System architects often use hand-crafted power viruses to stress the power of GPGPU systems. However, these stressmarks are very tedious to generate and require significant domain knowledge. But perhaps more importantly, one cannot be sure about the fact that the generated power virus is actually a maximum case. What’s worse, these stressmarks can’t adapt to different GPGPUs.To address these problems, this paper proposes a genetic algorithm based maximum power consumption test framework, which maximizes the power consumption of GPGPUs automatically by using genetic algorithm. We firstly build an abstract power workload model through a random forest algorithm based analytical model. Then we design a code generation algorithm to generate the synthetic power virus exploration space according to the parameters of abstract workload model. Finally, we use GA tools to search the space and find the optimal synthetic power virus.For two different GPGPUs, we show the efficacy of our framework, by comparing the power consumption with about40real benchmarks of GPGPU. The results show that our generated power stressmarks consume19%and14%more power consumption than that of maximum power of all the used benchmarks for GTX480and Quradro FX5600, respectively. Compared with the actual TDP, our framework can achieve high accuracy with an error about6%and3%for two GPGPUs, respectively. Our proposed framework can not only generate the optimal power virus quickly and accurately, but also evaluate the stability and efficiency of new GPGPUs.