Jerome Soumagne
Jerome Soumagne
add psm, psm2 and hwloc variants
When going over the eager message size limit, Mercury automatically uses an overflow mechanism that lets larger messages be transferred using RMA. When using one-way RPCs however, there is currently...
OFI can generate cancel events without users calling cancel resulting in callbacks returning an NA_CANCELED event. This is confusing for users and we should return an appropriate error code instead.
Original message from Chuck: > I have found a bad feature interaction between mercury's atomic queue data structure, NA_NO_BLOCK, and thread CPU binding. > The impact of this issue varies...
Not really high-priority but once in a while we have timeout issues (not at the moment) and it would be fairly easy to measure how much time is spent in...
There are a number of places where operation IDs are allocated upon user request without any hard limit on the number of allocations that can be done. We should think...
It would be useful to add a call that tells HG that no response will be sent (if for example some random error occurred) in the RPC callbacks so that...
Some users have expressed the need for having that on by default. We should measure how much of an impact that could potentially have.
E-mail from Xuezhao: > 1) Is it considerable to add a range list parameter to HG_Bulk_transfer API? (or by adding another HG_Bulk_transferv() API to keep current API not changed) >...
Attached are 4 stack traces of this dump. The config is a little bit different than the previous one in that we have 4 DeltaFS compaction threads instead of 1...