Talked to one of our developers (afraid everyone is swamped with SC12 right now), and we have a few suggestions:
1. to reduce OS-jitter, have you tried --bind-to-core or --bind-to-socket? This would avoid context switches throwing you onto different memory segments 2. we recall you had an issue with page faults and we suggested using a different shared memory method - did that help? Are you still seeing page faults? 3. are you on a dedicated segment of the IB fabric? Is there competing traffic on the IB network that might cause the delay? 4. you might run a layer0 diagnostic on the IB fabric to ensure there isn't a problem in that area HTH Ralph On Nov 9, 2012, at 9:36 AM, "Hodge, Gary C" <gary.c.ho...@lmco.com> wrote: > Answering my own question, I have downloaded openMPI 1.6.2 and still get the > delay in the MPI_Send. Previously, I was using openMPI 1.4.1 > > I configured 1.6.2 with –enable-peruse and have implemented PERUSE callbacks. > Here is a trace of gb2 (PROC39) sending messages to ob (PROC28) > > PROC39: 59.2268409810; SEND_REQ_ACTIVATE, count=81536, peer=28 > PROC39: 59.2268649811; SEND_REQ_XFER_BEGIN, count=81536, peer=28 > PROC39: 59.2270019811; SEND_REQ_XFER_CONTINUE, count=27178, peer=28 > PROC39: 59.2270509811; SEND_REQ_XFER_CONTINUE, count=54358, peer=28 > PROC39: 59.2270809811; SEND_REQ_XFER_END, count=81536, peer=28 > PROC39: 59.2270869811; SEND_REQ_COMPLETE, count=81536, peer=28 > PROC39: 59.2270939811; SEND_REQ_NOTIFY, count=81536, peer=28 > PROC28: 59.2271299611; RECV_MSG_ARRIVED, count=0, peer=39 > PROC28: 59.2271389612; RECV_SEARCH_POSTED_Q_BEGIN, count=0, peer=39 > PROC28: 59.2271439611; RECV_REQ_REMOVE_FROM_POSTED_Q, count=172084, peer=39 > PROC28: 59.2271479612; RECV_MSG_MATCH_POSTED_REQ, count=172084, peer=39 > PROC28: 59.2271519612; RECV_SEARCH_POSTED_Q_END, count=0, peer=39 > PROC28: 59.2271559612; RECV_REQ_XFER_BEGIN, count=172084, peer=39 > PROC28: 59.2271609611; RECV_REQ_XFER_CONTINUE, count=27178, peer=39 > PROC28: 59.2271759611; RECV_REQ_XFER_CONTINUE, count=54358, peer=39 > PROC39: 59.2283769811; SEND_REQ_ACTIVATE, count=81536, peer=28 > PROC39: 59.2283959811; SEND_REQ_XFER_BEGIN, count=81536, peer=28 > PROC39: 59.4331379811; SEND_REQ_XFER_CONTINUE, count=27178, peer=28 > PROC28: 59.4331819612; RECV_REQ_XFER_END, count=172084, peer=39 > PROC28: 59.4332009612; RECV_REQ_COMPLETE, count=172084, peer=39 > PROC39: 59.4332049811; SEND_REQ_XFER_CONTINUE, count=54358, peer=28 > PROC28: 59.4332059611; RECV_MSG_ARRIVED, count=0, peer=39 > PROC28: 59.4332109612; RECV_SEARCH_POSTED_Q_BEGIN, count=0, peer=39 > PROC28: 59.4332149612; RECV_MSG_INSERT_IN_UNEX_Q, count=0, peer=39 > PROC28: 59.4332199611; RECV_SEARCH_POSTED_Q_END, count=0, peer=39 > PROC28: 59.4332269612; RECV_REQ_NOTIFY, count=172084, peer=39 > PROC28: 59.4332539611; RECV_REQ_ACTIVATE, count=172084, peer=39 > PROC39: 59.4332569811; SEND_REQ_XFER_END, count=81536, peer=28 > PROC28: 59.4332589612; RECV_SEARCH_UNEX_Q_BEGIN, count=172084, peer=39 > PROC28: 59.4332629612; RECV_REQ_MATCH_UNEX, count=172084, peer=39 > PROC39: 59.4332659811; SEND_REQ_COMPLETE, count=81536, peer=28 > PROC28: 59.4332669611; RECV_MSG_REMOVE_FROM_UNEX_Q, count=0, peer=39 > PROC28: 59.4332719612; RECV_SEARCH_UNEX_Q_END, count=172084, peer=39 > PROC28: 59.4332759612; RECV_REQ_XFER_BEGIN, count=172084, peer=39 > PROC39: 59.4332769811; SEND_REQ_NOTIFY, count=81536, peer=28 > PROC28: 59.4332809611; RECV_REQ_XFER_CONTINUE, count=27178, peer=39 > PROC39: 59.4332849811; ERROR: component gb2 exceeded send time limit by > 0.104915 seconds. > PROC39: 59.4332849811; WARNING: component gb2 increased page faults > R=130898,0; P=130898,0; S=132207,0 > > Notice the 205 ms delay between the hi-lighted SEND_REQ_XFER_BEGIN and the > subsequent SEND_REQ_XFER_CONTINUE > Also notice that there was no such delay in the previous send request. > > The last two lines are my own debug that monitor excessive time spent in an > MPI_Send and monitor increasing number of page faults. > After the delaying MPI_Send, the page faults have increased by 1309 (132207 – > 130898) !!! > > I looked at the PML code (pml_ob1_sendreq.c) and I suspect that something is > going awry in the mca_bml_base_prepare_src function. > I believe that this is the function in which the openib BTL makes calls to > register the user memory. > > This delay in the MPI_Send kills any hope of meeting our real-time > requirements, so any help is greatly appreciated. > > > From: Hodge, Gary C > Sent: Monday, November 05, 2012 12:27 PM > To: us...@open-mpi.org > Subject: gathering problem > > I continue to have a problem where 2 processes are sending to the same > process and one of the sending processes hangs for 150 to 550 ms in the call > to MPI_Send. > > Each process runs on a different node and the receiving process has posted an > MPI_Irecv 17 ms before the hanging send. > The posted receives are for 172K buffers and the sending processes are > sending 81K size messages. > I have set mpi_leave_pinned to 1 and have increased the > btl_openib_receive_queues to …:S,65536,512,256,64 > > How do I trace the various phases of message passing to diagnose where the > send is hanging up? > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users
