The problem we'll have with BTLs in opal is going to revolve around that ompi_process_name_t and will occur in a number of places. I've been trying to grok George's statement about accessors and can't figure out a clean way to make that work IF every RTE gets to define the process name a different way.
For example, suppose I define ompi_process_name_t to be a string. I can hash the string down to an opal_identifier_t, but that is a structureless 64-bit value - there is no concept of a jobid or vpid in it. So if you now want to extract a jobid for that identifier, the only way you can do it is to "up-call" back to the RTE to parse it. This means that every RTE would have to initialize OPAL with a registration of its opal_identifier parser function(s), which seems like a really ugly solution. Maybe it is time to shift the process identifier down to the opal layer? If we define opal_identifier_t to include the required jobid/vpid, perhaps adding a void* so someone can put whatever they want in it? Note that I'm not wild about extending the identifier size beyond 64-bits as the memory footprint issue is growing in concern, and I still haven't seen any real use-case proposed for extending it. On May 1, 2014, at 3:41 AM, Jeff Squyres (jsquyres) <jsquy...@cisco.com> wrote: > On Apr 30, 2014, at 10:01 PM, George Bosilca <bosi...@icl.utk.edu> wrote: > >> Why do you need the ompi_process_name_t? Isn’t the opal_identifier_t enough >> to dig for the info of the peer into the opal_db? > > > At the moment, I use the ompi_process_name_t for RML sends/receives in the > usnic BTL. I know this will have to change when the BTLs move down to OPAL > (when is that going to happen, BTW?). So my future use case may be somewhat > moot. > > More detail > =========== > > "Why does the usnic BTL use RML sends/receives?", you ask. > > The reason is rooted in the fact that the usnic BTL uses an unreliable, > connectionless transport under the covert. We had some customers have > network misconfigurations that resulted in usnic traffic not flowing properly > (e.g., MTU mismatches in the network). But since we don't have a > connection-oriented underlying API that will eventually timeout/fail to > connect/etc. when there's a problem with the network configuration, we added > a "connection validation" service in the usnic BTL that fires up in a thread > in the local rank 0 on each server. This thread provides service to all the > MPI processes on its server. > > In short: the service thread sends UDP pings and ACKs to peer service threads > on other servers (upon demand/upon first send between servers) to verify > network connectivity. If the pings eventually fail/timeout (i.e., don't get > ACKs back), the service thread does a show_help and kills the job. > > There's more details, but that's the gist of it. > > This basically gives us the ability to highlight problems in the network and > kill the MPI job rather than spin infinitely while trying to deliver MPI/BTL > messages to a peer that will never get there. > > Since this is really a server-to-server network connectivity issue (vs. an > MPI peer-to-peer connectivity issue), we only need to have one service thread > for a whole server. The other MPI procs on the server use RML to talk to it. > E.g., "Please ping the server where MPI proc X lives," and so on. This > seemed better than having a service thread in each MPI process. > > We've thought a bit about what to do when the BTLs move down to OPAL (since > they won't be able to use RML any more), but don't have a final solution > yet... We do still want to be able to utilize this capability even after the > BTL move. > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/05/14673.php
