Making Wide-Area, Multi-site MPI Feasible Using Xen VM

Although multi-site MPI execution has been criticized in the past as being ”impractical” due to limitations in network latency and bandwidth, we believe many of the obstacles can be overcome by various means for wider classes of applications than previously believed. One such technique is transparent dynamic migration of MPI, coupled with aggressively performance-oriented overlay networks, assuming availability of gigabits of bandwidth on future WANs. The problem, of course, is to investigate the exact implications to application performances given the arsenal of such techniques, but such work has been quite sparse. Our current work involves using Xen as the underlying virtual machine layer to implement such migration, along with performance-optimizing migration strategies—this particular paper deals with performance evaluations of MPI on Xen VMs including what are the possible performance hindrances, implications of migrations, as well as the effect of variations in latencies and bandwidth parameters as realized by the overlay network using a software network emulator.