I should clarify something: While opal_hwloc_base_open() is invoked during opal_init(), the global variable opal_hwloc_topology is not filled with topology information during opal_init() due to the overhead of invoking hwloc_topology_load() in many processes simultaneously.
Specifically: the process of loading the topology can be expensive, particularly on large-core-count machines (e.g., SGI machines). Consider 64 (or 128 or 256 or ...) MPI processes all hammering on /proc and/or /sys at the same time. We've had reports on the hwloc users list that this is not a good idea; it *really* slows down the startup process. Additionally, since opal is, by definition, a single process abstraction, it made sense to push the decision of whether/how to load the topology information to a higher layer. On the OMPI trunk, the orted loads the topology information and disseminates it to its local processes during the ESS handshake. To be specific: - the orted and HNP will have non-NULL values in opal_hwloc_topology after the ESS startup - MPI processes will have non-NULL values of opal_hwloc_topology after orte_init() -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/