> Most people don't care about cache when binding with MPI, so you may > just ignore the issue and hide the message by setting > HWLOC_HIDE_ERRORS=1 in the environment. It may work fine (assuming > MPIs don't have troubles with asymmetric topologies where there are > some missing L3).
We do see some problems with our MPI (that's what prompted me to install hwloc 1.9 and openmpi 1.8.2rc2 in the first place), but I don't think it's related to this cache problem: when we submit an MPI job to, say, 48 processors, only a handful of random ones work at 100%, the rest are idling (as seen in htop). We are running openmpi via mpi4py to do Markov Chain Monte Carlo (mcmc) fitting. > Otherwise, hwloc can load the topology from XML. So we'll just need to > generate a fixed topology, export it to XML and set an environment > variable to have hwloc load from there. A single file may even be > enough for all similar nodes assuming your MPI and/or applications > don't look in deep details of hwloc topologies. We have 4 identical nodes; one runs as a server and the others boot off the server via NFS. But there are also 6 older nodes (with 4 processors each) that boot off the server in the same way, so for those I imagine that a fixed topology XML won't do much good. That said, we are in the process of phasing those 6 nodes out, and I wouldn't be terribly sad if this gives us a good reason to do so. Thanks a bunch, Andrej
