Section: Software
MPICH2-NewMadeleine
Participants : Guillaume Mercier, François Trahay.
MPICH2-NewMadeleine is the successor to the former MPICH-Madeleine implementation which derived from MPICH 1.2.7. MPICH2-NewMadeleine is based on the most recent MPICH2 software (1.2.1) and utilizes the NewMadeleine communication library for all network communication. As far as intranode communication are concerned, the Nemesis communication subsystem is employed [6] . Nemesis is a generic communication subsystem which goal is to address the communication needs of a wide range of programming tools and environments for clusters and parallel architectures. It has been designed to yield very low latency and high bandwidth, especially for intranode communication.
The resulting MPI implementation exhibits excellent performance, especially in the shared-memory case, which is crucial in the case of NUMA clusters. The level of performance is indeed very good and MPICH2-Nemesis compares favorably with other next-generation MPI implementations such as Open MPI or GridMPI. The latencies achieved by MPICH2-NewMadeleine in shared-memory are currently the best among generic MPI implementations and are extremely close to that of highly-tuned vendor-specific ports.
The NewMadeleine communication library has been integrated as a Nemesis network module but some architectural changes have been made to the upper layers, in particular at the ADI3 level (ch3 has thus been modified). Those changes allow the MPICH2 software stack to efficiently take advantage of tag-matching capable interfaces such as NewMadeleine or Myricom MX. Also, all optimization mechanisms implemented in NewMadeleine are made available at the MPI level. For instance, MPICH2, with its NewMadeleine Nemesis module can use natively multirail configurations.
This work takes advantage of the Associate Team program in order to make frequent visits in order to coordinate the work.
MPICH2-NewMadeleine , a joint development between the ANL and the Runtime project-team will soon be available on the Runtime website and is developed and maintained by Darius Buntinas , Guillaume Mercier , François Trahay and David Goodell .
