Statistical Modeling of Resource Availability in Desktop Grids

Abstract

Desktop grids are compute platforms that aggregate and harvest the idle CPU cycles of individually owned personal computers and workstations. A challenge for using these platforms is that the compute resources are volatile. Due to this volatility the vast majority of desktop grid applications are embarrassingly parallel and high-throughput. Deeper understanding of the nature of resource availability is needed to enable the use of desktop grids for a broader class of applications. In this document we further this understanding thanks to statistical analysis of availability traces collected on real-world desktop grid platforms.