Hi Ralph, I spoke with Jeff Squyres at SC11, and updated him on the status of my OpenMPI port on Hadoop Yarn.
To update everyone, I have OpenMPI examples running on #Yarn, although it requires some code cleanup and refactoring, however that can be done as a later step. Currently, the MPI processes come up, get submitting client's IP and port via environment variables, connect to it, and do a barrier. The result of this barrier is that everyone in MPI_COMM_WORLD gets each other's endpoints. I am aiming to submit the patch to hadoop by the end of this month. I will publish the openmpi patch to github. (As I mentioned to Jeff, OpenMPI requires a CCLA for accepting submissions. That will take some time.) - Milind --- Milind Bhandarkar Greenplum Labs, EMC (Disclaimer: Opinions expressed in this email are those of the author, and do not necessarily represent the views of any organization, past or present, the author might be affiliated with.) > >I'm willing to do the integration work, but wanted to check first to see >if (a) someone in the Hadoop community is already doing so, and (b) if >you would be interested in seeing such a capability and willing to accept >the code contribution? > >Establishing MPI support requires the following steps: > >1. wireup support. MPI processes need to exchange endpoint info (e.g., >for TCP connections, IP address and port) so that each process knows how >to connect to any other process in the application. This is typically >done in a collective "modex" operation. There are several ways of doing >it - if we proceed, I will outline those in a separate email to solicit >your input on the most desirable approach to use. > >2. binding support. One can achieve significant performance improvements >by binding processes to specific cores, sockets, and/or NUMA regions >(regardless of using MPI or not, but certainly important for MPI >applications). This requires not only the binding code, but some logic to >ensure that one doesn't "overload" specific resources. > >3. process mapping. I haven't verified it yet, but I suspect that Hadoop >provides each executing instance with an identifier that is unique within >that job - e.g., we typically assign an integer "rank" that ranges from 0 >to the number of instances being executed. This identifier is critical >for MPI applications, and the relative placement of processes within a >job often dictates overall performance. Thus, we would provide a mapping >capability that allows users to specify patterns of process placement for >their job - e.g., "place one process on each socket on every node". > >I have written the code to implement the above support on a number of >systems, and don't foresee major problems doing it for Hadoop (though I >would welcome a chance to get a brief walk-thru the code from someone). >Please let me know if this would be of interest to the Hadoop community. > >Thanks >Ralph Castain > > >
