Performance Considerations for Containers on San Diego Supercomputer Center's Comet
TimeTuesday, July 306:30pm - 8:30pm
LocationCrystal Foyer and Crystal B
DescriptionContainers are an increasingly popular addition to the modern High Performance Computing workflow and are widely used on XSEDE computing resources. They allow users of traditional HPC resources more freedom with OS environments and software stacks while operating in non-privileged userspace. A common concern of researchers is getting the most computation out of their allocations. This poster explores the comparative performance analysis of common HPC application benchmarks (e.g. AMBER, GROMACS) and popular Machine Learning software (e.g. TensorFlow, PyTorch) in containers and on bare metal through regular queues. In addition, performance differences between runs via precompiled containers (such as from Singularity or Docker hub) vs compiling with Comet hardware specific optimization flags are analyzed. Often publicly available precompiled images may not have specific optimization flags for the CPUs and GPUs on a particular machine. Others like NVIDIA's NGC may handle GPU architecture optimizations but not have extensive CPU specific optimizations. In this study special attention is paid to using different libraries (with and without architecture specific optimization) and compilers to understand performance impacts and develop best practices. Methods for validation of these results are additionally presented.