Evaluation of Reallocation Heuristics for Moldable Tasks in Computational Grids

Caniou, Y., Charrier, G. and Desprez, F.

Grid services often consist of remote sequential or rigid parallel application executions. However, moldable parallel applications, linear algebra solvers for example, are of great interest but requires dynamic tuning which has mostly to be done interactively if performances are needed. Thus, their grid execution depends on a remote and transparent submission to a possibly different batch scheduler on each site, and means an automatic tuning of the job according to the local load. In this paper we study the benefits of having a middleware able to automatically submit and reallocate requests from one site to another when it is also able to configure the services by tuning their number of processors and their walltime. In this context, we evaluate the benefits of such mechanisms on two multi-cluster Grid setups, where the platform is either composed of several heterogeneous dedicated clusters, or non dedicated ones. Different scenarios are explored using simulations of real cluster traces from different origins. Results show that a simple method is good and often the best. Indeed, it is faster and thus can take more jobs into account while having a small execution time. Moreover, users can expect more jobs finishing sooner and a gain on the average job response time between 10% and 40% in most cases if this reallocation mechanism combined to auto-tuning capabilities is implemented in a Grid framework. The implementation and the maintenance of this heuristic coupled to the migration mechanism in a Grid middleware is also simpler because less transfers are involved.

Cite as: Caniou, Y., Charrier, G. and Desprez, F. (2011). Evaluation of Reallocation Heuristics for Moldable Tasks in Computational Grids. In Proc. Australasian Symposium on Parallel and Distributed Computing (AusPDC 2011) Perth, Australia. CRPIT, 118. Jinjun Chen and Rajiv Ranjan Eds., ACS. 15-24

(from crpit.com) (local if available)