On Tue, 2013-06-18 at 07:13 -0400, Hal Rosenstock wrote: > On 6/17/2013 5:38 PM, Albert Chu wrote: > > We've recently noticed that the Node Description for a node can > > mis-mismatch between the output of smpquery and saquery. For example: > > > > # smpquery NodeDesc 427 > > Node Description:.................sierra1932 qib0 > > > > # saquery NodeRecord 427 | grep NodeDesc > > NodeDescription.........QLogic Infiniband HCA > > > > A restart of OpenSM is the current solution to resolve this. > > > > We've noticed it occurring more often on our larger clusters than our > > smaller clusters, leading to a speculation about why it is happening. > > > > The speculation is when a node comes up, there is a window of time in > > which the HCA is up, can be scanned by OpenSM, but not yet have its node > > descriptor set (in RHEL I appears to be set via /etc/init.d/rdma). > > During this window, OpenSM reads/stores the non-desired node descriptor > > (in the above case the non-desired "Qlogic Infiniband HCA"). > > > > When the node descriptor is changed, a trap should be sent to opensm > > indicating the change. Normally OpenSM gets the trap and reads the new > > node descriptor. > > Are you sure the trap is being issued by those devices when the > NodeDescription is changed locally ?
These particular devices do support the trap and tests show they do send traps on changes (i.e. manually changing /sys/class/infiniband/qib0/node_desc). > Also, if so, do these devices implement timeout/retry on sending the > trap (e.g. trying to make sure that they receive trap repress before > giving up on trap) ? This I don't know. I've been trying to figure out if they do and if they do how it might be configurable. Is there a way to figure this out? > > On our large clusters all nodes are typically brought up at the same > > time, so there are probably a ton of node descriptor change traps > > happening at the exact same time. We speculate a number of these are > > dropped/lost, and subsequently OpenSM never realizes that the node > > descriptor has changed. > > Do you see any evidence of that traps are being dropped ? Have you > correlated any VL15Dropped counters in the subnet with this ? Also, > there is a module parameter in MAD kernel module that might help with > any unsolicited MAD bursts. You might try increasing that on your SM > node(s). On our largest clusters we always see a nice chunk of VL15 drops, however we haven't correlated them specifically to a trap. > > I don't know if the speculation sounds reasonable or not. Regardless, > > we're not sure of the best fix. > > > > A trivial fix would be to just make OpenSM re-scan the node descriptor > > of an HCA, perhaps during a heavy sweep. But I don't know if this is > > optimal. It'll introduce more MADs on the wire. However if the present > > solution is to restart OpenSM, we figure this can't be any worse. > > Yes, but to add the additional queries in is O(n) there and has been > resisted in the past. > > > Just wondering what peoples thoughts are of if there's another obvious > > solution we're not seeing. > > I think this issue needs better understanding first. Yeah, just looking for hints/pointers for the time being. Thanks, Al > -- Hal > > > Al > > > -- Albert Chu [email protected] Computer Scientist High Performance Systems Division Lawrence Livermore National Laboratory -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
