Power Profiling of GeMTC Many Task Computing

workloads to run on hardware accelerators allowing for ad-vantages that come from the many-core architecture. However, presently GeMTC is only written to take advantage of NVIDIA GPUs. Another such hardware accelerator, the Intel Xeon Phi, is also an excellent candidate for MTC workloads. Therefore, the first goal of this project will be to add support to GeMTC to allow it to run on Xeon Phi. While there has been plenty of research on power consumption of hardware accelerators, MTC workloads are a significantly understudied research area. MonEQ, a power profiling library, was primarily developed to measure power consumption of the IBM Blue Gene/Q supercomputer, but has lately evolved to also include profiling of both NVIDIA GPUs as well as the Intel Xeon Phi. As a second goal, this project seeks to profile real MTC workloads running on both NVIDIA GPUs as well as on the Xeon Phi. I. BACKGROUND INFORMATION GeMTC [1], [2] enables Many Task Computing (MTC) workloads to run efficiently on NVIDIA GPUs. The standard CUDA framework can only support 16 kernels, one kernel for each streaming multiprocessor (SM). The biggest downside to this is that each kernel must be started and stopped at the same time, which leads to incredible inefficiencies in heterogeneous workloads which are commonplace for MTC. By working at the warp level, a trade can be made for local memory and concurrency which allows the execution of up to 200 concurrent kernels. Preliminary results show that GeMTC framework is able to achieve a higher level of efficiency for various MTC workloads. MonEQ [3], [4] is a power monitoring library initially designed to measure the power consumption of scientific applications running on the IBM Blue Gene/Q supercomputer. Results show that extremely accurate, precise, and sub-second power data can be obtained with marginal overhead added. The Blue Gene series of supercomputers is homogeneous, however, so the recent evolution of MonEQ has enabled the profiling of applications running on both NVIDIA GPUs and Intel Xeon Phi accelerators (potentially at the same time) using the vendor supplied low-level APIs for data gathering again at sub-second intervals