I'll respond to the non-MPICH aspects of this thread offline...

...

I think the only way for async progress to work well is to have fine-grain locking since of MPI, as is done in PAMID. Any implementation that resorts to fat-locking is probably better off without async progress unless the application is doing something really silly (like never calling MPI on a rank that is the target of MPI RMA).

There's more than 100k cycles between times that I enter the MPI stack. There are only two threads ever that contend for locks (my funneled thread and MPICH's async-progress thread). I'm not convinced that you need super fine-grained locks to make progress during that time period. FWIW, my experience has been that standard Nemesis does a _much_ better job of making asynchronous progress than Cray's implementation. If we could look at the code for Cray's implementation, we might be able to get a better idea of why.

I would be interesting to analyze async progress in detail using MPICH and MVAPICH, since those are OSS. We are never going to get the CrayPICH source (I've asked, NCSA asked) so it's not prudent to debate what we would do with it if we did. We could reimplement the uGNI netmod if we had nothing better to do. Jeff -- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond