SIParCS 2022 - Daniel Hope
Measuring GPU Power Consumption Using NVIDIA Tools
Power consumption measurement is an interesting engineering problem in High Performance Computing. Various factors such as compute time, load of computation, and resource allocation by programmers affect the power consumption of a code. We need to have accurate power consumption tools in order to optimize the resource usage and to reduce the environmental impact of high-performance computing. This summer was devoted to measuring power consumption using two tools that NVIDIA has created. The goal is to see how well they can perform this task – System Management Interface (NVIDIA- SMI) and profiling tool (NV_PROF). A corpus of average computational runs was performed, measuring power consumption of MuRAM and miniweather-OpenACC parallel programs run on a single NVIDIA V100 GPU.
The findings of this research showed the average power consumption and Standard deviation of SMI and NV_PROF, to see how well suited they can be for NCAR Linux-based clusters. This research demonstrates the utility of both programs for power consumption for large scale clusters.
The findings of this research showed power consumption measurements with both SMI and NV_PROF. This siparcs project compares the power readings obtained by both the tools.
Mentors: John Dennis, Cena Miller, Supreeth Suresh