FT-MPI, Fault-Tolerant Metacomputing and Generic Name Services: A Case Study

There is a growing interest in deploying MPI over very large numbers of heterogenous, geographically distributed resources. FT-MPI provides the fault-tolerance necessary at this scale, but presents some issues when crossing multiple administrative domains. Using the H2O metacomputing framework, we add cross-administrative domain interoperability and “pluggability” to FT-MPI. The latter feature allows us, using proxies, to transparently replace one vulnerable module – its name service – with fault-tolerant replacements. We present an algorithm for improving performance of operations over the proxies. We evaluate its performance in a comparison using the original name service, OpenLDAP and current Emory research project HDNS.