Critical path and area based scheduling of parallel task graphs on heterogeneous platforms

Abstract

While most parallel task graphs scheduling research has been done in the context of single homogeneous clusters, heterogeneous platforms have become prevalent and are extremely attractive for deploying applications at unprecedented scales. In this paper we address the need for scheduling techniques for parallel task applications for heterogeneous clusters of clusters by proposing a method to adapt existing parallel task graphs scheduling heuristics that have proved to be efficient on homogeneous environments. The contributions of this paper are: (i) a novel "virtual" cluster methodology for handling platform heterogeneity; (ii) a novel task placement step, designed to determine whether the placement step of heuristics for homogeneous platforms is adapted to the heterogeneous case; (iii) an empirical evaluation in a wide range of platform and application scenarios. This study shows that the proposed heuristics achieve better performance than the original when platform are heterogeneous and we discuss a number of trends apparent in our results.

A Comparison of Scheduling Approaches for Mixed-Parallel Applications on Heterogeneous Platforms

“…Two algorithms have been recently proposed to schedule mixed-parallel applications on heterogeneous platforms: HCPA [9] and M-HEFT [2]. Since we improve on and compare both algorithms, we describe each in detail hereafter.…”

“…Then, processor allocations on the reference cluster are "translated" to allocations on the original clusters. Such a translation is only possible with an analytical model for T (t, n) as a function of n for all task t. The authors in [9] use a popular model for the execution time of many dataparallel applications based on Amdahl's law, meaning that a fraction α of the sequential task execution time cannot be parallelized. A direct application of Amdahl's law shows that, given an allocation on the reference cluster, one can compute an equivalent allocation on any cluster C k , i.e., an allocation that leads to the same task execution time.…”

“…We first give a short overview of CPA before describing the salient features of HCPA. (Full details on both algorithms can be found in [10] and [9].) CPA attempts to achieve the best tradeoff between the length of the critical path, T CP , and the average processor utilization, T A .…”

“…• We improve the HCPA two-phase algorithm [9] in two ways. First we implement a more efficient stopping criterion for the first phase.…”

“…The first approach consists in adapting the aforementioned two-phase algorithms for mixed-parallel applications on homogeneous platforms and making them amenable to heterogeneous platforms [9]. The second approach consists in adapting list scheduling algorithms that were specifically designed for executing task-parallel applications on heterogeneous platforms and making them amenable to mixed parallelism [2].…”