I fixed this up before the release (6ae96dca2adc), but I didn't address this point. Why is the second MPI_Win_fence called immediately after the MPI_Put in VecScatterBegin. I would expect better latency tolerance if this was called in VecScatterEnd, as in the attached patch. Is this patch somehow incorrect?
Jed -------------- next part -------------- A non-text attachment was scrubbed... Name: window.patch Type: text/x-patch Size: 1168 bytes Desc: not available URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20100327/13888dd4/attachment.bin>
