On Sun, Mar 17, 2019 at 6:55 PM Jed Brown <j...@jedbrown.org> wrote: > Jeff Hammond <jeff.scie...@gmail.com> writes: > > > When this was written, I was convinced that Dursi was wrong about > > everything because one of the key arguments against MPI was > > fault-intolerance, which I was sure was going to be solved soon. > However, > > LLNL has done everything in their power to torpedo MPI fault-tolerance in > > MPI-4 for the past 3+ years and I am no longer optimistic about MPI's > > ability to grow outside of traditional HPC because of the forum's > inability > > to take fault-tolerance seriously. It's also unclear that we can get by > > without it in a post-exascale world. > > Have you seen any MPI FT proposals that would actually enable a > collective library or application to meaningfully "recover"? > > Seems to me that in-memory checkpointing and process-based FT is more > practical. For example, you could have a Spark or Spark-like system > that manages distributed in-memory data (perhaps standardizing on > Arrow), launches MPI jobs to access that data in-place, and coordinates > the distributed replication so that a fresh MPI job could restart on a > (possibly) different group of nodes after the MPI job crashes. In such > a system, you would need to implement transactional updates to in-memory > checkpoint data, but no special recovery from MPI (just that crashes be > eventually collective, versus deadlock). >
Please look at https://prod-ng.sandia.gov/techlib-noauth/access-control.cgi/2016/1610522.pdf. This is similar to what you are proposing but without the expensive reboot of the MPI runtime. Jeff -- Jeff Hammond jeff.scie...@gmail.com http://jeffhammond.github.io/