Experimental study of large-scale computing on virtualized resources

Abstract

Parallel applications have a pressing need for the utilization of more and more resources to meet user's performance expectations. Unfortunately, these resources are not necessarily available within one single domain. Grid computing provides a solution to scaling out from a single domain; however, it also brings another problem for some applications: resource heterogeneity. Since some applications require having homogeneous resources for their execution, virtualizing the resources is a noble and viable solution. In this paper, we present two parallel applications, namely WRF and mpiBLAST and report the results of different runs scaling them out from 2 to 128 virtual nodes. Later, we analyze the effects of scaling out based on the application's communication behavior.