On Jun 5, 2014, at 7:09 AM, Ralph Castain <r...@open-mpi.org> wrote:
> Okay, before you go chasing this, let me explain that we already try to > address this issue in the TCP oob. When we need to connect to someone, we do > the following: > > 1. if we have a direct connection available, we hand the message to the > software module assigned to that NIC > > 2. if none of the available NICs match the target's subnet, then we assign > the message to the software module for the first NIC in the system - i.e., > the one with the lowest kernel index - and let it try to send the message. We > expect the OS to know how to route the connection. > > 3. if that fails for some reason, then we'll try assign it to the software > module for the next NIC in the system, continuing down this path until every > module has had a chance to try. Actually, this isn't quite correct. The NIC we assigned it to will cycle across all of the known connection addresses for the intended target, trying each in turn. If *none* of those successfully connect, then the module declares that it is unable to make the connection. At that point, we let the next software module try. This has always bothered me a bit as I don't see how it can succeed if the first one failed - the OS is going to decide which NIC to send the connection request across anyway. All we are doing is assigning the thread that will make the connection request. So long as that thread tries all the connection addresses, it shouldn't matter which thread makes the attempt. Point being: we can probably just let the one thread make the attempt and give up if it fails on all known addresses for the target. We can then bounce it up to the OOB framework and let someone else try with a different transport, should one be available for that target. This would simplify the logic. > > 4. if no TCP module can send it, we bump it back up to the OOB framework to > see if another component can send it. At the moment, we don't have one, but > that will shortly change. > > My intention is to be a little more intelligent on step #2. At the very > least, I'd like to see us find the closest subnet match - just check tuples > to see who has the most matches. So if the target is on 10.1.2.3 and I have > two NICs 10.2.3.x and 192.168.2.y, then I should pick the first one since it > at least matches something. > > If your IP experts have a better solution, please pass it along! What is > causing the problem here is that the message comes in on one NIC that doesn't > have a direct connection to the target, and the "hop" mechanism isn't working > correctly (kicks into an infinite loop). > > > > On Jun 5, 2014, at 4:27 AM, Jeff Squyres (jsquyres) <jsquy...@cisco.com> > wrote: > >> That raises a larger issue -- what about Ethernet-only clusters that span >> multiple IP/L3 subnets? This is a scenario that Cisco definitely wants to >> enable/support. >> >> The usnic BTL, for example, can handle this scenario. We hadn't previously >> considered the TCP oob component effects in this scenario -- oops. >> >> Hmm. >> >> The usnic BTL both does lazy connections (so to speak...) and uses a >> connectivity checker to ensure that it can actually communicate with each >> peer. In this way, OMPI has a way of knowing whether process A can >> communicate with process B, even if A and B have effectively unrelated IP >> addresses (i.e., they're not on the same IP subnet). >> >> I don't think the TCP oob will be able to use this same kind of strategy. >> >> As a simple solution, there could be an TCP oob MCA param that says >> "regardless of peer IP address, I can connect to them" (i.e., assume IP >> routing will make everything work out ok). >> >> That doesn't seem like a good overall solution, however -- it doesn't >> necessarily fit in the "it just works out of the box" philosophy that we >> like to have in OMPI. >> >> Let me take this back to some IP experts here and see if someone can come up >> with a better idea. >> >> >> >> On Jun 4, 2014, at 10:09 PM, Ralph Castain <r...@open-mpi.org> wrote: >> >>> Well, the problem is that we can't simply decide that anything called >>> "ib.." is an IB port and should be ignored. There is no naming rule >>> regarding IP interfaces that I've ever heard about that would allow us to >>> make such an assumption, though I admit most people let the system create >>> default names and thus would get something like an "ib..". >>> >>> So we leave it up to the sys admin to configure the system based on their >>> knowledge of what they want to use. On the big clusters at the labs, we >>> commonly put MCA params in the default param file for this purpose as we >>> *don't* want OOB traffic going over the IB fabric. >>> >>> But that's the sys admin's choice, not a requirement. I've seen >>> organizations that do it the other way because their Ethernet is really >>> slow. >>> >>> In this case, the problem is really in the OOB itself. The local proc is >>> connecting to its local daemon via eth0, which is fine. When it sends a >>> message to mpirun on a different proc, that message goes from the app to >>> the daemon via eth0. The daemon looks for mpirun in its contact list, and >>> sees that it has a direct link to mpirun via this nifty "ib0" interface - >>> and so it uses that one to relay the message along. >>> >>> This is where we are hitting the problem - the OOB isn't correctly doing >>> the transfer between those two interfaces like it should. So it is a bug >>> that we need to fix, regardless of any other actions (e.g., if it was an >>> eth1 that was the direct connection, we would still want to transfer the >>> message to the other interface). >>> >>> HTH >>> Ralph >>> >>> On Jun 4, 2014, at 7:32 PM, Gilles Gouaillardet >>> <gilles.gouaillar...@gmail.com> wrote: >>> >>>> Thanks Ralf, >>>> >>>> for the time being, i just found a workaround >>>> --mca oob_tcp_if_include eth0 >>>> >>>> Generally speaking, is openmpi doing the wiser thing ? >>>> here is what i mean : >>>> the cluster i work on (4k+ nodes) each node has two ip interfaces : >>>> * eth0 (gigabit ethernet) : because of the cluster size, several subnets >>>> are used. >>>> * ib0 (IP over IB) : only one subnet >>>> i can easily understand such a large cluster is not so common, but on the >>>> other hand i do not believe the IP configuration (subnetted gigE and >>>> single subnet IPoIB) can be called exotic. >>>> >>>> if nodes from different eth0 subnets are used, and if i understand >>>> correctly your previous replies, orte will "discard" eth0 because nodes >>>> cannot contact each other "directly". >>>> directly means the nodes are not on the same subnet. that being said, they >>>> can communicate via IP thanks to IP routing (i mean IP routing, i do *not* >>>> mean orte routing). >>>> that means orte communications will use IPoIB which might not be the best >>>> thing to do since establishing an IPoIB connection can be long (especially >>>> at scale *and* if the arp table is not populated) >>>> >>>> is my understanding correct so far ? >>>> >>>> bottom line, i would have expected openmpi uses eth0 regardless IP routing >>>> is required, and ib0 is simply not used (or eventually used as a fallback >>>> option) >>>> >>>> this leads to my next question : is the current default ok ? if not should >>>> we change it and how ? >>>> /* >>>> imho : >>>> - IP routing is not always a bad/slow thing >>>> - gigE can sometimes be better than IPoIB) >>>> */ >>>> >>>> i am fine if at the end : >>>> - this issue is fixed >>>> - we decide it is up to the sysadmin to make --mca oob_tcp_if_include eth0 >>>> the default if this is really thought to be best for the cluster. (and i >>>> can try to draft a faq if needed) >>>> >>>> Cheers, >>>> >>>> Gilles >>>> >>>> On Wed, Jun 4, 2014 at 11:50 PM, Ralph Castain <r...@open-mpi.org> wrote: >>>> >>>> I'll work on it - may take a day or two to really fix. Only impacts >>>> systems with mismatched interfaces, which is why we aren't generally >>>> seeing it. >>>> >>>> _______________________________________________ >>>> 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/2014/06/14972.php >>> >>> _______________________________________________ >>> 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/2014/06/14973.php >> >> >> -- >> 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/2014/06/14976.php >
