Re: [OMPI devel] TCP BTL routability (was: ticket #972)
On Tue, Jul 29, 2008 at 03:25:00PM -0400, Jeff Squyres wrote: > For reference, the FAQ entry is here: > > http://www.open-mpi.org/faq/?category=tcp#tcp-routability > > It looks like we now *always* assume that two TCP peers are routable. As long as they share the same address family (IPv4 or IPv6). > The code in question is in btl_tcp_proc.c with the loop starting at > line 413. Yes. The FAQ is outdated, the new code is very different. We now use graph theory, imagine a bipartite graph where each interface is a vertex. (one peer on the left, the other on the right, no connections inside each peer, only from left to right, hence a bipartite graph). Every edge in this graph is given a weight depending on its quality. The quality is "defined" in btl_tcp_proc.h: enum mca_btl_tcp_connection_quality { CQ_NO_CONNECTION, CQ_PRIVATE_DIFFERENT_NETWORK, CQ_PRIVATE_SAME_NETWORK, CQ_PUBLIC_DIFFERENT_NETWORK, CQ_PUBLIC_SAME_NETWORK }; CQ_NO_CONNECTION (weight 0) is for different address families, so we don't try to connect from IPv6 to IPv4 and vice versa. The more likely a connection is going to be established, the higher the weight. So public addresses on the same network (read: very close, probably sharing the same link) are the best one can get, private addresses on different networks have the lowest probability for a succeeding connection. We then try to find a matching in the graph, this means, no two edges may have a common endpoint on either side, thus avoiding oversubscription. In order to support striping, we look for the largest matching (read: selecting as many edges (links) as possible). In order to ensure connectivity, we then choose from the variety of largest matchings the one with the highest summed weights. These edges denote the addresses with the best probability for a succeeding connection. In terms of graph theory, this is called a maximum cardinality maximum weight matching. You can find the whole background story in Chapter 3: http://cluster.inf-ra.uni-jena.de/~adi/peiselt-thesis.pdf We have also a brief IEEE paper on this: http://www.ieeexplore.ieee.org/xpl/freeabs_all.jsp?isnumber=4476518=4476565=56=46 In other words: #972 is somewhat obsolete, the FAQ entry should surely be removed/updated. I don't know to which extent, but if you want me to write some lines, I could probably invent a not so scientific description. HTH -- Cluster and Metacomputing Working Group Friedrich-Schiller-Universität Jena, Germany private: http://adi.thur.de
[OMPI devel] TCP BTL routability (was: ticket #972)
On Jul 29, 2008, at 3:20 PM, Terry Dontje wrote: So, we've pinged ticket #972 several times to see if the issue it covers has been fixed and have not really gotten a response in the last few months. While talking with Jeff about a recent thread on the users list about this issue he has found the code in btl_tcp_proc.c that determines whether the tcp btl can be used has changed significantly between 1.2 and 1.3. So a question is has these changes changed the rules of whether connections between two nodes should use the tcp btl as described on the FAQ? If so we should update the FAQ. For reference, the FAQ entry is here: http://www.open-mpi.org/faq/?category=tcp#tcp-routability It looks like we now *always* assume that two TCP peers are routable. The code in question is in btl_tcp_proc.c with the loop starting at line 413. -- Jeff Squyres Cisco Systems
Re: [OMPI devel] trunk hangs since r19010
On Jul 29, 2008, at 9:47 AM, Jeff Squyres wrote: Ok. FWIW, Pasha and I think that openib has supported "send-to- self" for a while (we don't know exactly when; but Pasha thinks it is very old code that we don't check for self in add_procs). But it only broke recently. More in the FWIW category -- we just checked, and OMPI v1.2 supported "--mca btl openib" (note the lack of ",self"). So the openib BTL has, indeed, supported send-to-self for quite a while. This should help narrow where to start looking for the problem: changes within the last few weeks. -- Jeff Squyres Cisco Systems
Re: [OMPI devel] trunk hangs since r19010
Ok. FWIW, Pasha and I think that openib has supported "send-to-self" for a while (we don't know exactly when; but Pasha thinks it is very old code that we don't check for self in add_procs). But it only broke recently. On Jul 29, 2008, at 9:31 AM, George Bosilca wrote: I ran few tests and the only combination leading to a deadlock is openib and self. As openib is the only BTL supporting self communications (except self of course), I guess it interfere with self in some more or less strange ways. I didn't had the time to dig deeper yet to see what exactly happens there, I'll schedule this later today. george. On Jul 29, 2008, at 8:52 AM, Pavel Shamis (Pasha) wrote: Jeff Squyres wrote: This used to be true, but I think we changed it a while ago (Pasha: do you remember?) because Mellanox HCAs are capable of send-to-self (process) and there were no code changes necessary to enable it. So it allowed a slightly simpler command line. This was quite a while ago, IIRC. Yep, Correct. FYI. In my MTT testing I also see a lot of killed tests. ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel -- Jeff Squyres Cisco Systems
Re: [OMPI devel] Change in hostfile behavior
Lenny's point is true - except for the danger of setting that mca param and its possible impact on ORTE daemons+mpirun - see other note in that regard. However, it would only be useful if the same user was doing it. I believe Tim was concerned about the case where two users are sharing nodes. There is no good solution for that case. Two mpiruns done by different users that share a node and who have no knowledge of the other's actions will cause collision. We should probably warn about that in our FAQ or something since that is a fairly common use-case - only thing I can think of is to recommend people default to running without affinity and only set it when they -know- they have sole use of their nodes. On Jul 29, 2008, at 12:17 AM, Lenny Verkhovsky wrote: for two separate runs we can use slot_list parameter ( opal_paffinity_base_slot_list ) to have paffinity 1: mpirun -mca opal_paffinity_base_slot_list "0-1" 2 :mpirun -mca opal_paffinity_base_slot_list "2-3" On 7/28/08, Ralph Castainwrote: Actually, this is true today regardless of this change. If two separate mpirun invocations share a node and attempt to use paffinity, they will conflict with each other. The problem isn't caused by the hostfile sub-allocation. The problem is that the two mpiruns have no knowledge of each other's actions, and hence assign node ranks to each process independently. Thus, we would have two procs that think they are node rank=0 and should therefore bind to the 0 processor, and so on up the line. Obviously, if you run within one mpirun and have two app_contexts, the hostfile sub-allocation is fine - mpirun will track node rank across the app_contexts. It is only the use of multiple mpiruns that share nodes that causes the problem. Several of us have discussed this problem and have a proposed solution for 1.4. Once we get past 1.3 (someday!), we'll bring it to the group. On Jul 28, 2008, at 10:44 AM, Tim Mattox wrote: My only concern is how will this interact with PLPA. Say two Open MPI jobs each use "half" the cores (slots) on a particular node... how would they be able to bind themselves to a disjoint set of cores? I'm not asking you to solve this Ralph, I'm just pointing it out so we can maybe warn users that if both jobs sharing a node try to use processor affinity, we don't make that magically work well, and that we would expect it to do quite poorly. I could see disabling paffinity and/or warning if it was enabled for one of these "fractional" nodes. On Mon, Jul 28, 2008 at 11:43 AM, Ralph Castain wrote: Per an earlier telecon, I have modified the hostfile behavior slightly to allow hostfiles to subdivide allocations. Briefly: given an allocation, we allow users to specify --hostfile on a per-app_context basis. In this mode, the hostfile info is used to filter the nodes that will be used for that app_context. However, the prior implementation only filtered the nodes themselves - i.e., it was a binary filter that allowed you to include or exclude an entire node. The change now allows you to include a specified #slots for a given node as opposed to -all- slots from that node. You are limited to the #slots included in the original allocation. I just realized that I hadn't output a warning if you attempt to violate this condition - will do so shortly. Rather than just abort if this happens, I set the allocation to that of the original - please let me know if you would prefer it to abort. If you have interest in this behavior, please check it out and let me know if this meets needs. Ralph ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel -- Tim Mattox, Ph.D. - http://homepage.mac.com/tmattox/ tmat...@gmail.com || timat...@open-mpi.org I'm a bright... http://www.the-brights.net/ ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trunk hangs since r19010
Jeff Squyres wrote: This used to be true, but I think we changed it a while ago (Pasha: do you remember?) because Mellanox HCAs are capable of send-to-self (process) and there were no code changes necessary to enable it. So it allowed a slightly simpler command line. This was quite a while ago, IIRC. Yep, Correct. FYI. In my MTT testing I also see a lot of killed tests.
Re: [OMPI devel] Change in hostfile behavior
for two separate runs we can use slot_list parameter ( opal_paffinity_base_slot_list ) to have paffinity 1: mpirun -mca opal_paffinity_base_slot_list "0-1" 2 :mpirun -mca opal_paffinity_base_slot_list "2-3" On 7/28/08, Ralph Castainwrote: > > Actually, this is true today regardless of this change. If two separate > mpirun invocations share a node and attempt to use paffinity, they will > conflict with each other. The problem isn't caused by the hostfile > sub-allocation. The problem is that the two mpiruns have no knowledge of > each other's actions, and hence assign node ranks to each process > independently. > > Thus, we would have two procs that think they are node rank=0 and should > therefore bind to the 0 processor, and so on up the line. > > Obviously, if you run within one mpirun and have two app_contexts, the > hostfile sub-allocation is fine - mpirun will track node rank across the > app_contexts. It is only the use of multiple mpiruns that share nodes that > causes the problem. > > Several of us have discussed this problem and have a proposed solution for > 1.4. Once we get past 1.3 (someday!), we'll bring it to the group. > > > On Jul 28, 2008, at 10:44 AM, Tim Mattox wrote: > > My only concern is how will this interact with PLPA. >> Say two Open MPI jobs each use "half" the cores (slots) on a >> particular node... how would they be able to bind themselves to >> a disjoint set of cores? I'm not asking you to solve this Ralph, I'm >> just pointing it out so we can maybe warn users that if both jobs sharing >> a node try to use processor affinity, we don't make that magically work >> well, >> and that we would expect it to do quite poorly. >> >> I could see disabling paffinity and/or warning if it was enabled for >> one of these "fractional" nodes. >> >> On Mon, Jul 28, 2008 at 11:43 AM, Ralph Castain wrote: >> >>> Per an earlier telecon, I have modified the hostfile behavior slightly to >>> allow hostfiles to subdivide allocations. >>> >>> Briefly: given an allocation, we allow users to specify --hostfile on a >>> per-app_context basis. In this mode, the hostfile info is used to filter >>> the >>> nodes that will be used for that app_context. However, the prior >>> implementation only filtered the nodes themselves - i.e., it was a binary >>> filter that allowed you to include or exclude an entire node. >>> >>> The change now allows you to include a specified #slots for a given node >>> as >>> opposed to -all- slots from that node. You are limited to the #slots >>> included in the original allocation. I just realized that I hadn't output >>> a >>> warning if you attempt to violate this condition - will do so shortly. >>> Rather than just abort if this happens, I set the allocation to that of >>> the >>> original - please let me know if you would prefer it to abort. >>> >>> If you have interest in this behavior, please check it out and let me >>> know >>> if this meets needs. >>> >>> Ralph >>> >>> ___ >>> devel mailing list >>> de...@open-mpi.org >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>> >>> >> >> >> -- >> Tim Mattox, Ph.D. - http://homepage.mac.com/tmattox/ >> tmat...@gmail.com || timat...@open-mpi.org >> I'm a bright... http://www.the-brights.net/ >> ___ >> devel mailing list >> de...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/devel >> > > ___ > devel mailing list > de...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/devel >