erations posted on commB until the typical latency becomes similar to pre-Revoke operations on commA. The collective communication patterns are inherited, without modification, from the Open MPI non-fault tolerant “tuned” module. The Cray optimized MPI can, in some instances, achieve higher performance. For the purpose of our evaluation, the tuned generic implementation, based on
MPI point-to-point message exchanges, is representative of users’ communication patterns commonly found in typical, portable HPC applications.