Performance, Experimentation

“GPU Computing “utilizes high level language to run sequential part of the code on the CPU as well as speeds up parallel part via running it on GPUs but GPUs are SIMD by default which means they can run only single instruction on multiple data. The introduction of GEMTC framework [1] addresses these limitations by providing an efficient middleware through which tasks are submitted to a common task queue to the device and workers (warp which represent the lowest possible level of control on device) take out the tasks, execute them and put them back on the result queue. This work explores porting and evaluation of real world applications into GEMTC framework. I choose Imogen [2] advanced astrophysical simulation tool and SciColSim [3] which simulates scientific discovery. I was able to port pure fluid kernels from Imogen and expensive functions of SciColSim to GEMTC. The evaluation resulted in performance up to 200 plus tasks/sec for kernel with moderate size data inputs. The results were compared with the CPU equivalent code and GEMTC was able to outperform CPU code for moderate size data inputs