Checkpointing as a Service in Heterogeneous Cloud Environments

International audienceA non-invasive, cloud-agnostic approach is demonstratedfor extending existing cloud platforms to includecheckpoint-restart capability. Most cloud platforms currently relyon each application to provide its own fault tolerance. A uniformmechanism within the cloud itself serves two purposes: (a) directsupport for long-running jobs, which would otherwise require acustom fault-tolerant mechanism for each application; and (b) theadministrative capability to manage an over-subscribed cloudby temporarily swapping out jobs when higher priority jobsarrive. An advantage of this uniform approach is that it alsosupports parallel and distributed computations, over both TCPand InfiniBand, thus allowing traditional HPC applications totake advantage of an existing cloud infrastructure. Additionally,an integrated health-monitoring mechanism detects when longrunningjobs either fail or incur exceptionally low performance,perhaps due to resource starvation, and proactively suspends thejob. The cloud-agnostic feature is demonstrated by applying theimplementation to two very different cloud platforms: Snoozeand OpenStack. The use of a cloud-agnostic architecture alsoenables, for the first time, migration of applications from onecloud platform to another