A Family of Scheduling Algorithms for Hybrid Parallel Platforms

Abstract

More and more parallel computing platforms are built upon hybrid architectures combining multi-core processors (CPUs) and hardware accelerators like General Purpose Graphics Processing Units (GPGPUs). We present in this paper a new method for scheduling efficiently parallel applications with [Formula: see text] CPUs and [Formula: see text] GPGPUs, where each task of the application can be processed either on an usual core (CPU) or on a GPGPU. We consider the problem of scheduling [Formula: see text] independent tasks with the objective to minimize the time for completing the whole application (makespan). This problem is NP-hard, thus, we present two families of approximation algorithms that can achieve approximation ratios of [Formula: see text] or [Formula: see text] for any integer [Formula: see text] when only one GPGPU is considered, and [Formula: see text] or [Formula: see text] for [Formula: see text] GPGPUs, where [Formula: see text] is an arbitrary small value which corresponds to the target accuracy of a binary search. The proposed method is based on a dual approximation scheme that uses a dynamic programming algorithm. The associated computational costs are for the first (resp. second) family in [Formula: see text] (resp. [Formula: see text]) per step of dual approximation. The greater the value of parameter [Formula: see text], the better the approximation, but the more expensive the computational cost. Finally, we propose a relaxed version of the algorithm which achieves a running time in [Formula: see text] with a constant approximation bound of [Formula: see text]. This last result is compared to the state-of-the-art algorithm HEFT. The proposed solving method is the first general purpose algorithm for scheduling on hybrid machines with a theoretical performance guarantee that can be used for practical purposes.