MPC++ Performance for Commodity Clustering

In order to verify the viability finer-grained parallel language MPC++, which had originally been developed for Myrinet-specific environments, we performed ports on top of different breeds of MPI, to be executed on two networks of large performance/cost difference, as well as porting NPB 2.3 apps to test ease of expressiveness of parallel programs. Results were positive, (a) the port of the NPB 2.3 apps were effortless, (b) small penalty of additional MPI layer was negligible for NPB applications, and (c) for large data sets, MPC++/MPI on the 100Base-T network was competitive to both the C+MPI on Myrinet, and the original implementation of MPC++ on PM/Myrinet.