[petsc-dev] Nonblocking consensus and alternative "freespace" interface

[Barry Smith]

  Jed,

   I think you should namespace the Seg thingy with Petsc and I'd advocate namespacing even all the static functions with PETSc, makes code clearer quicker.

   Barry

On Jan 6, 2013, at 3:08 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:

> I pushed PetscCommBuildTwoSided() which is a scalable interface for setting up two-sided communication given one-sided information.
> 
> https://bitbucket.org/petsc/petsc-dev/commits/721cf8e8a07d6de2abc9b4f621f09a70e0d48012#Lsrc/sys/utils/mpits.cT257
> 
> The default interface uses the O(size) MPI_Reduce, but if you have an MPI-3 implementation (providing MPI_Ibarrier), it uses the nonblocking consensus algorithm from
> 
> http://www.unixer.de/publications/img/hoefler-dsde-protocols.pdf
> 
> You can revert to the old implementation using -build_twosided ALLREDUCE (versus IBARRIER). I'd like to eventually update most uses of PetscGatherNumberOfMessages() and PetscGatherMessageLengths() to use this interface, and to use it in new code whenever possible.
> 
> In the implementation, I wrote a utility similar to PetscFreeSpace, but with what I think is a safer and more generic interface. It's currently called SegArray, but I'd be happy to change the name (maybe StreamBuffer). It's essentially used to buffer a chunked stream (with low-latency incremental allocation) and later return it as a contiguous array. I'd appreciate comments on the API before moving it to include/petsc-private.
> 
> https://bitbucket.org/petsc/petsc-dev/commits/721cf8e8a07d6de2abc9b4f621f09a70e0d48012#Lsrc/sys/utils/mpits.cT42