Dynamic Data Flow Based Spatial Sorting Method for GPUs: Software Based Autonomous Parallelization

Ankush Rai

doi:10.37591/rtpc.v1i1.144

Abstract

The direct communication between distant parts of a computer system for a mesh or grid based scalable parallel system isn't cost effective. Thus, we have presented and evaluated the performances of spatial sorting method for multi-core GPUs (Graphic Processing Unit) computing. It is a parallel sorting algorithm for autonomous threading of workload which is considerably faster than Thrust Merge sorting, which is known to be the best comparison based sorting algorithm for GPUs. The study shows that the proposed algorithm is more efficient than randomized sample sort, when the memory operations for parallelization of stack units are taken into consideration. Additionally, the proposed algorithm provides an advantage over traditional sorting methods in way that it facilitate dynamic data flow and therefore, its operating time is devoid of the input data dependent fluctuations. Our technique proves its potential extensions of adaptability to more complex parallel algorithms.Keywords: Autonomous parallelization, threading, sorting, GPUs, parallel algorithm, dynamic data flow