Empirical Performance Evaluation of Message Passing Programs Running in Virtual Machines

The virtualization technology can bring some advantages to HPC applications such as fault tolerant, load balance and simplify the system management etc. However, the performance is still a very important factor under considerations in the HPC community. This paper provides some empirical studies on the performance evaluations for the massive passing programs, which composing large amount of HPC applications, running over the virtual execution environments. NPB Benchmark, a widely used benchmark for benchmarking the scientific computing platforms, is used as the typical workload of message passing programs. The experiments are performed on two types of virtualization platforms, hardware supported virtualization and para-virtualization. Experimental results show that, for message passing HPC programs, a lot of performance improvements are still required on programs running on virtualization platforms especially for hardware supported virtualization platforms.