Virtual Machines for Grid Computing

Erik Paulson

Abstract: The emerging computational grid infrastructure will provide users with access to orders of magnitude more computing power than they currently have available. These resources will be heterogeneous in type and implementation and independently controlled and administered. Users of grid resources will need mechanisms to account for these variations that must both be without a high performance overhead while providing resource owners strong safety assurances. We propose using virtual machines, a classic idea from the 1960's and 70's, as this mechanism.

In this work, we briefly explore some of the challenges that grid computing users will be faced with. We compare several different architectures for building virtual machines, and compare performance of several standard benchmarks under these different virtual machines. We examine some unique opportunities the additional layer of abstraction provides, such as checkpointing, split execution, and heterogenous process migration, all with unmodified user executables running on unmodified operating systems. We conclude by examining future work, and propose integrating this work with Condor, a Grid-Enabled High-Throughput Computing system developed at the University of Wisconsin, Madison.

Available as: Postscript or PDF