Probabilistic Risk Assessment for Resource Provision in Grid

Service Level Agreements (SLAs) are introduced to overcome the shortages of best-effort approach in Grid computing and make Grid computing more attractive for commercial uses. Yet commercial Grid providers are not keen to adopt SLAs, since there is a risk of SLA violation, which will result in a penalty fee. This paper analyses failure data collected from three different Grid sites. We study the statistics of the data including the root cause, the mean time to repair and time between failures. We find that software and hardware failures are the largest contributors, and that the time to repair varies, depending on the root cause, from 13 hours in network errors to around 46 hours in unknown errors. We also find that the repair time is well modelled by a Weibull distribution. From the analysis of the historical data we find that the distribution between failures in a Grid system is well modelled by a Weibull distribution with decreasing hazard rate, and this can be used by a resource provider to predicate the risk of failure.