From Heterogeneous Task Scheduling to Heterogeneous Mixed Parallel Scheduling ⋆ Fr´ ed´ eric Suter 1 , Fr´ ed´ eric Desprez 2 , and Henri Casanova 1,3 1 Dept. of CSE, Univ. of California, San Diego, USA 2 LIP ENS Lyon, UMR CNRS - ENS Lyon - UCB Lyon - INRIA 5668, France 3 San Diego Supercomputer Center, Univ. of California, San Diego, USA Abstract. Mixed-parallelism, the combination of data- and task- parallelism, is a powerful way of increasing the scalability of entire classes of parallel applications on platforms comprising multiple compute clus- ters. While multi-cluster platforms are predominantly heterogeneous, previous work on mixed-parallel application scheduling targets only ho- mogeneous platforms. In this paper we develop a method for extending existing scheduling algorithms for task-parallel applications on heteroge- neous platforms to the mixed-parallel case. 1 Introduction Two kinds of parallelism can be exploited in most scientific applications: data- and task-parallelism. One way to maximize the degree of parallelism of a given application is to combine both kinds of parallelism. This approach is called mixed data and task parallelism or mixed parallelism. In mixed-parallel appli- cations, several data-parallel computations can be executed concurrently in a task-parallel way. This increases scalability as more parallelism can be exploited when the maximal amount of either data- or task-parallelism has been achieved. This capability is a key advantage for today’s parallel computing platforms. Indeed, to face the increasing computation and memory demands of parallel scientific applications, a recent approach has been to aggregate multiple compute clusters either within or across institutions [4]. Typically, clusters of various sizes are used, and different clusters contain nodes with different capabilities depending on the technology available at the time each cluster was assembled. Therefore, the computing environment is at the same time attractive because of the large computing power, and challenging because it is heterogeneous. A number of authors have explored mixed-parallel application scheduling in the context of homogeneous platforms [8–10]. However, heterogeneous platforms have become prevalent and are extremely attractive for deploying applications at unprecedented scales. In this paper we build on existing scheduling algo- rithms for heterogeneous platforms [6, 7, 11, 14] (i.e., specifically designed for task-parallelism) to develop scheduling algorithms for mixed-parallelism on het- erogeneous platforms. ⋆ An extended version of this paper is given by [13]. M. Danelutto, D. Laforenza, M. Vanneschi (Eds.): Euro-Par 2004, LNCS 3149, pp. 230–237, 2004. c  Springer-Verlag Berlin Heidelberg 2004