Speaker:Trenton Sudduth

OpenMP Device Mapping Prediction using GNNs

The increase in scale of high performance computing (HPC) workloads and the need of faster (close to real-time) executions has necessitated the increase in computing power of HPC systems. Parallel computing allows for the workload of a program to be run in parallel either on a CPU with multiple threads or through offloading work to the GPU. This has the potential of better utilizing system resources to improve performance without upgrading to more expensive SoCs. GPUs were first introduced to computer architecture to enhance video and 2D/3D graphics processing capabilities of existing desktops/workstations. Recent innovations and re-purposing of GPUs has led to the rise of GPGPUs which are being frequently used for enhancing the performance of parallel workloads. Deciding which device is the most optimal for execution is difficult to ascertain, as it is hard to predict the performance on the CPU or GPU. We introduce a method to predict the best device for program execution using semantically and structurally aware flow graphs augmented with performance counters as inputs to a graph neural network. We obtained the flow graphs of individual kernels and augmented them with performance counters to input information about how the flow of a kernel translated to the performance of the kernel into our graph neural network model. Our method achieved favorable accuracy results at 87\%. For the kernels that benefited from using the GPU, we measured a speedup of 1.5x in 60\% of the cases.

Committee: Ali Jannesari (major professor), Wensheng Zhang, and Wei Le.

Join on Zoom: Please click this URL to start or join. https://iastate.zoom.us/j/93872427077 Or, go to https://iastate.zoom.us/join and enter meeting ID: 938 7242 7077