mukautuva

mukautuva copied to clipboard

fix Op <-> User_function mapping

2

If we are going to store the mapping forever, we need to make Op_free a no-op, to avoid the ABA problem (thanks to Jed Brown for noticing this) if we...

jeffhammond

`cleanup_mapped_request` does a check on `.empty()` for all the relevant maps every call. We can hoist this to optimize the common case. ```c for (int i=0; i

jeffhammond

disconnect reconnect

1

Do I care? No. ``` mpirun.openmpi --oversubscribe -n 3 ./disconnect_reconnect3 . No Errors [nuclear][[13406,1],0][../../../../../../opal/mca/btl/tcp/btl_tcp.c:559:mca_btl_tcp_recv_blocking] recv(17) failed: Connection reset by peer (104) Unexpected output in disconnect_reconnect3: [nuclear][[13406,1],0][../../../../../../opal/mca/btl/tcp/btl_tcp.c:559:mca_btl_tcp_recv_blocking] recv(17) failed: Connection reset...

jeffhammond

comm spawn issues

1

1. It only works with Open-MPI as the default SO. `MPI_LIB` is ignored by spawned processes and thus things don't work. 2. This: ``` mpirun.openmpi --oversubscribe -n 1 ./spawninfo1 .--------------------------------------------------------------------------...

jeffhammond

All of this is either unimplemented or XFAIL. I need to reorganize the `dlsym` stuff to make this work.

jeffhammond

Sigh. I suppose we need one. Someone else can write it.

jeffhammond

We can create a MUK launcher that sets MPI_LIB and the appropriate implementation launcher together, based on MPI_PATH or some other combination of env var or launcher args. Bonus points...

jeffhammond

The following means that one can't just initialize everything during MPI initialization. One needs load these on demand. This requires modest changes to libinit.c but is not going to be...

jeffhammond