Predictive State Restoration in Desktop Workstation Clusters

Though existing systems for sharing distributed resources in clusters of workstations are generally e ective at recruiting idle resources, these systems often have a disruptive e ect on desktop workstation users. Even when recruiting computing cycles from strictly idle workstations, a by-product of running foreign jobs is that the virtual memory pages of the original user's idle processes are ushed to disk and the workstation's le cache is disrupted. Consequently, users resuming work after an idle period experience delays while the system restores this state. This paper presents novel methods for minimizing the disruptions to desktop workstation users in a cluster environment while still maintaining a high utilization of the idle resources of the cluster. Disruptions to the user are reduced by identifying the memory-resident state of the user's processes when the machine becomes idle and then actively restoring that state before the user returns, using measurements of past activity patterns to predict when that user is likely to return. Trace-driven simulations show that this method can predict a user's arrival up to 43% of the time while still recruiting 83% of a workstation's idle cycles.