A Competitive Environment for Parallel Applications on Heterogeneous Workstation Clusters

This paper 1 describes a system that facilitates a competitive environment for parallel applications on heterogeneous workstation clusters. The system tries to maximize the speedup of the parallel applications and, at the same time, to allocate workstations ef-ciently and fairly to the applications. It tackles the resource allocation problem in regard to heterogeneity between the workstations and dynamic workload interference between the applications. The applications in the system are constructed in a way that their partitioned execution segments are grouped into grains and the sizes of the grains can be adjusted dynamically. To enable fast workstation allocation, the number of possible grain-to-workstation mappings is restricted. Competition between applications is resolved by auctions held by the distributed scheduler that is running on each cluster. No prior knowledge about the run-time performance of the applications is required for bidding because the values of bids are formulated in terms of the past and current performance information collected at runtime. The design of the auction and bidding mechanisms is the centrepiece of this paper. The eeectiveness of the system is tested and the results of the experiments exhibit the dynamics of competition between parallel applications.