Section: Software

MPICH-V

Participant : Thomas Hérault [ correspondant ] .

Currently, MPICH-V proposes 7 protocols: MPICH-V1, MPICH-V2, MPICH-Vcl, MPICH-Pcl and 3 algorithms for MPICH-Vcausal. MPICH-V1 implements an original fault tolerant protocol specifically developed for Desktop Grids relying on uncoordinated checkpoint and remote pessimistic message logging. It uses reliable nodes called Channel Memories to store all in transit messages. MPICH-V2 is designed for homogeneous networks like clusters where the number of reliable component assumed by MPICH-V1 is too high. It reduces the fault tolerance overhead and increases the tolerance to node volatility. This is achieved by implementing a new protocol splitting the message logging into message payload logging and event logging. These two elements are stored separately on the sender node for the message payload and on a reliable event logger for the message events. The third protocol, called MPICH-Vcl, is derived from the Chandy-Lamport global snapshot algorithm. It implements coordinated checkpoint without message logging. This protocol exhibits less overhead than MPICH-V2 for clusters with low fault frequencies. MPICH-Pcl is a blocking implementation of Chandy-Lamport algorithm. It consists in exchanging messages for emptying every communication channel before checkpointing all processes. MPICH-Vcausal concludes the set of message logging protocols, implementing a causal logging. It provides less synchrony than the pessimistic logging protocols, allowing messages to influence the system before the sender can be sure that non deterministic events are logged, to the cost of appending some information to every communication. This sum of information may increase with the time, and different causal protocols, with different cut techniques, have been studied with the MPICH-V project.

The protocols developed during the first phase of the MPICH-V project are now being integrated into the two main open-source distributions of MPI, namely MPICH2 and OpenMPI. During this integration, we focus on keeping the best performances (i.e. introducing the smallest changes in the library communication driver). Eventually, the fault-tolerance properties of these two distributions should be provided by the Grand-Large project.

In addition to fault tolerant properties, MPICH-V:

1. provides a full runtime environment detecting and re-launching MPI processes in case of faults;

2. works on high performance networks such as Myrinet, Infiniband, etc (the performances are still divided by two);

3. allows the migration of a full MPI execution from one cluster to another, even if they are using different high performance networks.

The software, papers and presentations are available at http://mpich-v.lri.fr/