As far as I have been able to understand while looking at the code, it very much seems that Joshua pointed out the exact cause for the issue.
munmap'ing a virtual address space region does not evict it from mpool_grdma->pool->lru_list . If a later mmap happens to return the same address (a priori tied to different physical location), the userspace believes this segment is already registered, and eventually the transfer is directed to a bogus location. This also seems to match this old discussion: http://lists.openfabrics.org/pipermail/general/2009-April/058786.html although I didn't read the whole discussion there, it very much seems that the proposal for moving the pinning/caching logic to the kernel did not make it, unfortunately. So are we here in the situation where this "munmap should be intercepted" logic actually proves too fragile ? (in that it's not intercepted in my case). The memory MCA in my configuration is: MCA memory: linux (MCA v2.0, API v2.0, Component v1.8.3) I traced a bit what happens at the mmap call, it seems to go straight to the libc, not via openmpi first. For the time being, I think I'll have to consider any mmap()/munmap() rather unsafe to play with in an openmpi application. E. P.S: a last version of the test case is attached. Le 11 nov. 2014 19:48, "Emmanuel Thomé" <emmanuel.th...@gmail.com> a écrit : > > Thanks a lot for your analysis. This seems consistent with what I can > obtain by playing around with my different test cases. > > It seems that munmap() does *not* unregister the memory chunk from the > cache. I suppose this is the reason for the bug. > > In fact using mmap(..., MAP_ANONYMOUS | MAP_PRIVATE) and munmap() as > substitutes for malloc()/free() trigger the same problem. > > It looks to me that there is an oversight in the OPAL hooks around the > memory functions, then. Do you agree ? > > E. > > On Tue, Nov 11, 2014 at 3:17 PM, Joshua Ladd <jladd.m...@gmail.com> wrote: > > I was able to reproduce your issue and I think I understand the problem a > > bit better at least. This demonstrates exactly what I was pointing to: > > > > It looks like when the test switches over from eager RDMA (I'll explain in a > > second), to doing a rendezvous protocol working entirely in user buffer > > space things go bad. > > > > If you're input is smaller than some threshold, the eager RDMA limit, then > > the contents of your user buffer are copied into OMPI/OpenIB BTL scratch > > buffers called "eager fragments". This pool of resources is preregistered, > > pinned, and have had their rkeys exchanged. So, in the eager protocol, your > > data is copied into these "locked and loaded" RDMA frags and the put/get is > > handled internally. When the data is received, it's copied back out into > > your buffer. In your setup, this always works. > > > > $mpirun -np 2 --map-by node --bind-to core -mca pml ob1 -mca > > btl_openib_if_include mlx4_0:1 -mca btl_openib_use_eager_rdma 1 -mca > > btl_openib_eager_limit 512 -mca btl openib,self ./ibtest -s 56 > > per-node buffer has size 448 bytes > > node 0 iteration 0, lead word received from peer is 0x00000401 [ok] > > node 0 iteration 1, lead word received from peer is 0x00000801 [ok] > > node 0 iteration 2, lead word received from peer is 0x00000c01 [ok] > > node 0 iteration 3, lead word received from peer is 0x00001001 [ok] > > > > When you exceed the eager threshold, this always fails on the second > > iteration. To understand this, you need to understand that there is a > > protocol switch where now your user buffer is used for the transfer. Hence, > > the user buffer is registered with the HCA. This operation is an inherently > > high latency operation and is one of the primary motives for doing > > copy-in/copy-out into preregistered buffers for small, latency sensitive > > ops. For bandwidth bound transfers, the cost to register can be amortized > > over the whole transfer, but it still affects the total bandwidth. In the > > case of a rendezvous protocol where the user buffer is registered, there is > > an optimization mostly used to help improve the numbers in a bandwidth > > benchmark called a registration cache. With registration caching the user > > buffer is registered once and the mkey put into a cache and the memory is > > kept pinned until the system provides some notification via either memory > > hooks in p2p malloc, or ummunotify that the buffer has been freed and this > > signals that the mkey can be evicted from the cache. On subsequent > > send/recv operations from the same user buffer address, OpenIB BTL will find > > the address in the registration cache and take the cached mkey and avoid > > paying the cost of the memory registration the memory registration and start > > the data transfer. > > > > What I noticed is when the rendezvous protocol kicks in, it always fails on > > the second iteration. > > > > $mpirun -np 2 --map-by node --bind-to core -mca pml ob1 -mca > > btl_openib_if_include mlx4_0:1 -mca btl_openib_use_eager_rdma 1 -mca > > btl_openib_eager_limit 128 -mca btl openib,self ./ibtest -s 56 > > per-node buffer has size 448 bytes > > node 0 iteration 0, lead word received from peer is 0x00000401 [ok] > > node 0 iteration 1, lead word received from peer is 0x00000000 [NOK] > > -------------------------------------------------------------------------- > > > > So, I suspected it has something to do with the way the virtual address is > > being handled in this case. To test that theory, I just completely disabled > > the registration cache by setting -mca mpi_leave_pinned 0 and things start > > to work: > > > > $mpirun -np 2 --map-by node --bind-to core -mca pml ob1 -mca > > btl_openib_if_include mlx4_0:1 -mca btl_openib_use_eager_rdma 1 -mca > > btl_openib_eager_limit 128 -mca mpi_leave_pinned 0 -mca btl openib,self > > ./ibtest -s 56 > > per-node buffer has size 448 bytes > > node 0 iteration 0, lead word received from peer is 0x00000401 [ok] > > node 0 iteration 1, lead word received from peer is 0x00000801 [ok] > > node 0 iteration 2, lead word received from peer is 0x00000c01 [ok] > > node 0 iteration 3, lead word received from peer is 0x00001001 [ok] > > > > I don't know enough about memory hooks or the registration cache > > implementation to speak with any authority, but it looks like this is where > > the issue resides. As a workaround, can you try your original experiment > > with -mca mpi_leave_pinned 0 and see if you get consistent results. > > > > > > Josh > > > > > > > > > > > > On Tue, Nov 11, 2014 at 7:07 AM, Emmanuel Thomé <emmanuel.th...@gmail.com> > > wrote: > >> > >> Hi again, > >> > >> I've been able to simplify my test case significantly. It now runs > >> with 2 nodes, and only a single MPI_Send / MPI_Recv pair is used. > >> > >> The pattern is as follows. > >> > >> * - ranks 0 and 1 both own a local buffer. > >> * - each fills it with (deterministically known) data. > >> * - rank 0 collects the data from rank 1's local buffer > >> * (whose contents should be no mystery), and writes this to a > >> * file-backed mmaped area. > >> * - rank 0 compares what it receives with what it knows it *should > >> * have* received. > >> > >> The test fails if: > >> > >> * - the openib btl is used among the 2 nodes > >> * - a file-backed mmaped area is used for receiving the data. > >> * - the write is done to a newly created file. > >> * - per-node buffer is large enough. > >> > >> For a per-node buffer size above 12kb (12240 bytes to be exact), my > >> program fails, since the MPI_Recv does not receive the correct data > >> chunk (it just gets zeroes). > >> > >> I attach the simplified test case. I hope someone will be able to > >> reproduce the problem. > >> > >> Best regards, > >> > >> E. > >> > >> > >> On Mon, Nov 10, 2014 at 5:48 PM, Emmanuel Thomé > >> <emmanuel.th...@gmail.com> wrote: > >> > Thanks for your answer. > >> > > >> > On Mon, Nov 10, 2014 at 4:31 PM, Joshua Ladd <jladd.m...@gmail.com> > >> > wrote: > >> >> Just really quick off the top of my head, mmaping relies on the virtual > >> >> memory subsystem, whereas IB RDMA operations rely on physical memory > >> >> being > >> >> pinned (unswappable.) > >> > > >> > Yes. Does that mean that the result of computations should be > >> > undefined if I happen to give a user buffer which corresponds to a > >> > file ? That would be surprising. > >> > > >> >> For a large message transfer, the OpenIB BTL will > >> >> register the user buffer, which will pin the pages and make them > >> >> unswappable. > >> > > >> > Yes. But what are the semantics of pinning the VM area pointed to by > >> > ptr if ptr happens to be mmaped from a file ? > >> > > >> >> If the data being transfered is small, you'll copy-in/out to > >> >> internal bounce buffers and you shouldn't have issues. > >> > > >> > Are you saying that the openib layer does have provision in this case > >> > for letting the RDMA happen with a pinned physical memory range, and > >> > later perform the copy to the file-backed mmaped range ? That would > >> > make perfect sense indeed, although I don't have enough familiarity > >> > with the OMPI code to see where it happens, and more importantly > >> > whether the completion properly waits for this post-RDMA copy to > >> > complete. > >> > > >> > > >> >> 1.If you try to just bcast a few kilobytes of data using this > >> >> technique, do > >> >> you run into issues? > >> > > >> > No. All "simpler" attempts were successful, unfortunately. Can you be > >> > a little bit more precise about what scenario you imagine ? The > >> > setting "all ranks mmap a local file, and rank 0 broadcasts there" is > >> > successful. > >> > > >> >> 2. How large is the data in the collective (input and output), is > >> >> in_place > >> >> used? I'm guess it's large enough that the BTL tries to work with the > >> >> user > >> >> buffer. > >> > > >> > MPI_IN_PLACE is used in reduce_scatter and allgather in the code. > >> > Collectives are with communicators of 2 nodes, and we're talking (for > >> > the smallest failing run) 8kb per node (i.e. 16kb total for an > >> > allgather). > >> > > >> > E. > >> > > >> >> On Mon, Nov 10, 2014 at 9:29 AM, Emmanuel Thomé > >> >> <emmanuel.th...@gmail.com> > >> >> wrote: > >> >>> > >> >>> Hi, > >> >>> > >> >>> I'm stumbling on a problem related to the openib btl in > >> >>> openmpi-1.[78].*, and the (I think legitimate) use of file-backed > >> >>> mmaped areas for receiving data through MPI collective calls. > >> >>> > >> >>> A test case is attached. I've tried to make is reasonably small, > >> >>> although I recognize that it's not extra thin. The test case is a > >> >>> trimmed down version of what I witness in the context of a rather > >> >>> large program, so there is no claim of relevance of the test case > >> >>> itself. It's here just to trigger the desired misbehaviour. The test > >> >>> case contains some detailed information on what is done, and the > >> >>> experiments I did. > >> >>> > >> >>> In a nutshell, the problem is as follows. > >> >>> > >> >>> - I do a computation, which involves MPI_Reduce_scatter and > >> >>> MPI_Allgather. > >> >>> - I save the result to a file (collective operation). > >> >>> > >> >>> *If* I save the file using something such as: > >> >>> fd = open("blah", ... > >> >>> area = mmap(..., fd, ) > >> >>> MPI_Gather(..., area, ...) > >> >>> *AND* the MPI_Reduce_scatter is done with an alternative > >> >>> implementation (which I believe is correct) > >> >>> *AND* communication is done through the openib btl, > >> >>> > >> >>> then the file which gets saved is inconsistent with what is obtained > >> >>> with the normal MPI_Reduce_scatter (alghough memory areas do coincide > >> >>> before the save). > >> >>> > >> >>> I tried to dig a bit in the openib internals, but all I've been able > >> >>> to witness was beyond my expertise (an RDMA read not transferring the > >> >>> expected data, but I'm too uncomfortable with this layer to say > >> >>> anything I'm sure about). > >> >>> > >> >>> Tests have been done with several openmpi versions including 1.8.3, on > >> >>> a debian wheezy (7.5) + OFED 2.3 cluster. > >> >>> > >> >>> It would be great if someone could tell me if he is able to reproduce > >> >>> the bug, or tell me whether something which is done in this test case > >> >>> is illegal in any respect. I'd be glad to provide further information > >> >>> which could be of any help. > >> >>> > >> >>> Best regards, > >> >>> > >> >>> E. Thomé. > >> >>> > >> >>> _______________________________________________ > >> >>> users mailing list > >> >>> us...@open-mpi.org > >> >>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > >> >>> Link to this post: > >> >>> http://www.open-mpi.org/community/lists/users/2014/11/25730.php > >> >> > >> >> > >> >> > >> >> _______________________________________________ > >> >> users mailing list > >> >> us...@open-mpi.org > >> >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > >> >> Link to this post: > >> >> http://www.open-mpi.org/community/lists/users/2014/11/25732.php > >> > >> _______________________________________________ > >> users mailing list > >> us...@open-mpi.org > >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > >> Link to this post: > >> http://www.open-mpi.org/community/lists/users/2014/11/25740.php > > > > > > > > _______________________________________________ > > users mailing list > > us...@open-mpi.org > > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > > Link to this post: > > http://www.open-mpi.org/community/lists/users/2014/11/25743.php
#define _GNU_SOURCE #include <sys/types.h> #include <sys/stat.h> #include <fcntl.h> #include <stdio.h> #include <stdint.h> #include <stdlib.h> #include <string.h> #include <unistd.h> #include <assert.h> #include <mpi.h> #include <sys/mman.h> /* This test file illustrates how in certain circumstances, an mmap area * cannot correctly receive data sent from an MPI_Send call. * * This program wants to run on 2 distinct nodes connected with * infiniband. * * Normal behaviour of the program consists in printing output similar * to: node 0 iteration 0, lead word received from peer is 0x00000401 [ok] node 0 iteration 1, lead word received from peer is 0x00000801 [ok] node 0 iteration 2, lead word received from peer is 0x00000c01 [ok] node 0 iteration 3, lead word received from peer is 0x00001001 [ok] * * Abnormal behaviour is when the job ends with MPI_Abort after printing * a line such as: node 0 iteration 1, lead word received from peer is 0x00000000 [NOK] * * Each iteration of the main loop does the same thing. * - rank 0 allocates a buffer with mmap * - rank 1 sends data there with MPI_Send * - rank 0 verifies that the data has been correctly received. * - rank 0 frees the buffer with munmap * * The final check performed by rank 0 fails if the following conditions * are met: * * - the openib btl is used among the 2 nodes * - allocation is done via mmap/munmap (not via malloc/free) * - the send is large enough. * * The first condition is controlled by the btl mca. * The size of the transfer is controlled by the -s command line * argument */ /* For compiling, one may do: MPI=$HOME/Packages/openmpi-1.8.3 $MPI/bin/mpicc -W -Wall -std=c99 -O0 -g prog5.c * For running, assuming /tmp/hosts contains the list of 2 nodes, and * $SSH is used to connect to these: SSH_AUTH_SOCK= DISPLAY= $MPI/bin/mpiexec -machinefile /tmp/hosts --mca plm_rsh_agent $SSH --mca rmaps_base_mapping_policy node -n 2 ./a.out -s 2048 */ /* * Tested (FAIL means that setting USE_MMAP_FOR_FILE_IO above yields to a * program failure, while we succeed if it is unset). * * IB boards MCX353A-FCBT, fw rev 2.32.5100, MLNX_OFED_LINUX-2.3-1.0.1-debian7.5-x86_64 * openmpi-1.8.4rc1 FAIL (ok with --mca btl ^openib) * openmpi-1.8.3 FAIL (ok with --mca btl ^openib) * * A previous, longer test case also failed with IB boards MHGH29-XTC. */ /* Passing --mca mpi_leave_pinned 0 eliminates the bug */ int main(int argc, char * argv[]) { MPI_Init(&argc, &argv); int size; int rank; int eitems = 1530; /* eitems >= 1530 seem to fail on my cluster */ MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &size); if (size != 2) abort(); int use_mmap = 1; for(argc--, argv++; argc ; ) { if (argc >= 2 && strcmp(argv[0], "-s") == 0) { eitems = atoi(argv[1]); argc -= 2; argv += 2; continue; } if (strcmp(argv[0], "-malloc") == 0) { use_mmap = 0; argc--, argv++; continue; } fprintf(stderr, "Unexpected: %s\n", argv[0]); exit(EXIT_FAILURE); } size_t chunksize = eitems * sizeof(unsigned long); size_t wsiz = ((chunksize - 1) | (sysconf (_SC_PAGESIZE)-1)) + 1; unsigned long * localbuf = malloc(chunksize); for(int iter = 0 ; iter < 4 ; iter++) { unsigned long magic = (1 + iter) << 10; int ok = 1; if (rank == 1) { for(int item = 0 ; item < eitems ; item++) { localbuf[item] = magic + rank; } MPI_Send(localbuf, eitems, MPI_UNSIGNED_LONG, 0, 0, MPI_COMM_WORLD); } else { unsigned long * recvbuf; if (use_mmap) { recvbuf = mmap(NULL, wsiz, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); } else { recvbuf = malloc(wsiz); } MPI_Recv(recvbuf, eitems, MPI_UNSIGNED_LONG, !rank, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); ok = (*recvbuf == magic + !rank); fprintf(stderr, "node %d iteration %d, lead word received from peer is 0x%08lx [%s]\n", rank, iter, *recvbuf, ok?"ok":"NOK"); if (use_mmap) { munmap(recvbuf, wsiz); } else { free(recvbuf); } } /* only rank 0 has performed a new check */ MPI_Bcast(&ok, 1, MPI_INT, 0, MPI_COMM_WORLD); if (!ok) MPI_Abort(MPI_COMM_WORLD, 1); } free(localbuf); MPI_Finalize(); }