Exploiting unused cluster resources with virtualization

Cluster applications may have timing constraints. One approach to ensuring their satisfaction is over-provisioning, i.e. to provide more hardware resources than needed for processing a certain peak data rate. This concept is used for the HLT-Chain. This application runs in the ALICE HLT cluster and processes events at runtime of the ALICE experiment. Over-provisioning has a drawback: When physical resources are dimensioned for a peak data rate, then these resources are underutilized at times of decreased data rates. From a perspective of efficiency this is not desirable. Therefore a software framework has been developed which allows to run additional third-party applications in order to exploit temporarily unused cluster resources. To avoid relevant negative impact to the time-critical application the third-party applications are encapsulated in Virtual Machines (VMs) and the resource usage of VMs is dynamically adapted at runtime. The adaption is done both globally by e.g. hot-migrating VMs between nodes, but also locally by modifying the local resource share (e.g. CPU) of a VM. Policies allow to tune the trade-off between benefit of third-party applications (increased cluster usage, computed results) and negative impact to the time-critical application. Experiments show that using the framework in parallel to the HLT-Chain leads to additional computed results, increases the cluster CPU usage from 49% to 79% and causes no relevant impact to the HLT-Chain.