The following papers should clarify the motivation and the usage mechanism: http://greg.bronevetsky.com/papers/2003PPoPP.pdf http://sc07.supercomputing.org/schedule/pdf/pap224.pdf The first describes a checkpointing protocol that was implemented on top of MPI. The library intercepted all application MPI calls and added appropriate piggybacking and message logging logic to ensure that regardless of when each application process decided to take a checkpoint, it would still be possible to put together a consistent global state on restart. The user did not need to add any code. The second paper describes a library called PnMPI that uses the PMPI interface to make it possible to insert as many PMPI interception layers as one would like, with the possibility of having them interact and use each other's functionality. We expect that the piggybacking interface will primarily be used by tools written at this level, using tools like PnMPI or just plain PMPI. In particular, our focus here is on tool builders, not application programmers and we believe that APIs like this will be useful to all application developers simply because they're important to tool builders whose tools may be used by any developer. In addition to checkpointing, piggybacking is useful for tracing libraries, performance analysis tools and generic libraries that need to lazily propagate information and would prefer not to introduce any new communication for performance reasons. Greg Bronevetsky Post-Doctoral Researcher 1028 Building 451 Lawrence Livermore National Lab (925) 424-5756 [email protected] At 06:36 AM 4/21/2008, Terry Dontje wrote:

So I reread the piggybacking document on wiki. I am not thrilled with the amount of new APIs this would be adding to the standard but can also see the point of the paper. I am curious how the new API is expected to be used? The proposal say's this API is needed for user-level fault tolerance solutions. So do we expect a user to change all application calls to the MPI library to use the PB calls? I wonder if a more general solution that doesn't require a direct change to the API would work.

I wonder if there might be a way one could register piggybacking with a communicator and somehow have the actual piggybacking occur as a callback from an implementations messaging layer?

Just a thought,

--td _______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft