On Fri, Jul 3, 2020 at 11:38 AM Karl Lin <karl.lin...@gmail.com> wrote:
> Hi, Barry > > Thanks for the explanation. Following your tip, I have a guess. We use > MatCreateAIJ to create the matrix, I believe this call will preallocate as > well. Before this call we figure out the number of nonzeros per row for all > rows and put those number in an array, say numNonZero. We pass numNonZero > as d_nnz and o_nnz to MatCreateAIJ call, so essentially we preallocate > twice as much as needed. For the process that double the memory footprint > and crashed, there are a lot of values in both the diagonal and > off-diagonal part for the process, so the temporary space gets filled up > for both diagonal and off-diagonal parts of the matrix, also there are > unused temporary space until MatAssembly, so gradually fill up the > preallocated space which doubles the memory footprint. Once MatAssembly is > done, the unused temporary space gets squeezed out, we return the correct > memory footprint of the matrix. But before MatAssembly, large amount of > unused temporary space needs to be kept because of the diagonal and > off-diagonal pattern of the input. Would you say this is a plausible > explanation? thank you. > Yes. We find that it takes a very small amount of time to just loop over the insertion twice, the first time counting the nonzeros. We built something to do this for you: https://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/Mat/MatPreallocatorPreallocate.html Thanks, Matt > Regards, > > Karl > > On Fri, Jul 3, 2020 at 9:50 AM Barry Smith <bsm...@petsc.dev> wrote: > >> >> Karl, >> >> If a particular process is receiving values with MatSetValues() that >> belong to a different process it needs to allocate temporary space for >> those values. If there are many values destined for a different process >> this space can be arbitrarily large. The values are not pass to the final >> owning process until the MatAssemblyBegin/End calls. >> >> If you have not preallocated enough room the matrix actually makes a >> complete copy of itself with extra space for additional values, copies the >> values over and then deletes the old matrix this the memory use can double >> when the preallocation is not correct. >> >> >> Barry >> >> >> On Jul 3, 2020, at 9:44 AM, Karl Lin <karl.lin...@gmail.com> wrote: >> >> Yes, I did. The memory check for rss computes the memory footprint of >> column index using size of unsigned long long instead of int. >> >> For Junchao, I wonder if keeping track of which loaded columns are owned >> by the current process and which loaded columns are not owned also needs >> some memory storage. Just a wild thought. >> >> On Thu, Jul 2, 2020 at 11:40 PM Ernesto Prudencio <epruden...@slb.com> >> wrote: >> >>> Karl, >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> * Are you taking into account that every “integer” index might be 64 >>> bits instead of 32 bits, depending on the PETSc configuration / compilation >>> choices for PetscInt? Ernesto. From: petsc-users >>> [mailto:petsc-users-boun...@mcs.anl.gov <petsc-users-boun...@mcs.anl.gov>] >>> On Behalf Of Junchao Zhang Sent: Thursday, July 2, 2020 11:21 PM To: Karl >>> Lin <karl.lin...@gmail.com <karl.lin...@gmail.com>> Cc: PETSc users list >>> <petsc-users@mcs.anl.gov <petsc-users@mcs.anl.gov>> Subject: [Ext] Re: >>> [petsc-users] matcreate and assembly issue Is it because indices for the >>> nonzeros also need memory? --Junchao Zhang On Thu, Jul 2, 2020 at 10:04 >>> PM Karl Lin <karl.lin...@gmail.com <karl.lin...@gmail.com>> wrote: Hi, >>> Matthew Thanks for the reply. However, I don't really get why additional >>> malloc would double the memory footprint. If I know there is only 1GB >>> matrix being loaded, there shouldn't be 2GB memory occupied even if Petsc >>> needs to allocate more space. regards, Karl On Thu, Jul 2, 2020 at >>> 8:10 PM Matthew Knepley <knep...@gmail.com <knep...@gmail.com>> wrote: On >>> Thu, Jul 2, 2020 at 7:30 PM Karl Lin <karl.lin...@gmail.com >>> <karl.lin...@gmail.com>> wrote: Hi, Matt Thanks for the tip last time. We >>> just encountered another issue with large data sets. This time the behavior >>> is the opposite from last time. The data is 13.5TB, the total number of >>> matrix columns is 2.4 billion. Our program crashed during matrix loading >>> due to memory overflow in one node. As said before, we have a little memory >>> check during loading the matrix to keep track of rss. The printout of rss >>> in the log shows normal increase in many nodes, i.e., if we load in a >>> portion of the matrix that is 1GB, after MatSetValues for that portion, rss >>> will increase roughly about 1GB. On the node that has memory overflow, the >>> rss increased by 2GB after only 1GB of matrix is loaded through >>> MatSetValues. We are very puzzled by this. What could make the memory >>> footprint twice as much as needed? Thanks in advance for any insight. The >>> only way I can imagine this happening is that you have not preallocated >>> correctly, so that some values are causing additional mallocs. Thanks, >>> Matt Regards, Karl On Thu, Jun 11, 2020 at 12:00 PM Matthew >>> Knepley <knep...@gmail.com <knep...@gmail.com>> wrote: On Thu, Jun 11, 2020 >>> at 12:52 PM Karl Lin <karl.lin...@gmail.com <karl.lin...@gmail.com>> wrote: >>> Hi, Matthew Thanks for the suggestion, just did another run and here are >>> some detailed stack traces, maybe will provide some more insight: *** >>> Process received signal *** Signal: Aborted (6) Signal code: (-6) >>> /lib64/libpthread.so.0(+0xf5f0)[0x2b56c46dc5f0] [ 1] >>> /lib64/libc.so.6(gsignal+0x37)[0x2b56c5486337] [ 2] >>> /lib64/libc.so.6(abort+0x148)[0x2b56c5487a28] [ 3] >>> /libpetsc.so.3.10(PetscTraceBackErrorHandler+0xc4)[0x2b56c1e6a2d4] [ 4] >>> /libpetsc.so.3.10(PetscError+0x1b5)[0x2b56c1e69f65] [ 5] >>> /libpetsc.so.3.10(PetscCommBuildTwoSidedFReq+0x19f0)[0x2b56c1e03cf0] [ 6] >>> /libpetsc.so.3.10(+0x77db17)[0x2b56c2425b17] [ 7] >>> /libpetsc.so.3.10(+0x77a164)[0x2b56c2422164] [ 8] >>> /libpetsc.so.3.10(MatAssemblyBegin_MPIAIJ+0x36)[0x2b56c23912b6] [ 9] >>> /libpetsc.so.3.10(MatAssemblyBegin+0xca)[0x2b56c1feccda] By >>> reconfiguring, you mean recompiling petsc with that option, correct? >>> Reconfiguring. Thanks, Matt Thank you. Karl On Thu, Jun 11, >>> 2020 at 10:56 AM Matthew Knepley <knep...@gmail.com <knep...@gmail.com>> >>> wrote: On Thu, Jun 11, 2020 at 11:51 AM Karl Lin <karl.lin...@gmail.com >>> <karl.lin...@gmail.com>> wrote: Hi, there We have written a program using >>> Petsc to solve large sparse matrix system. It has been working fine for a >>> while. Recently we encountered a problem when the size of the sparse matrix >>> is larger than 10TB. We used several hundred nodes and 2200 processes. The >>> program always crashes during MatAssemblyBegin.Upon a closer look, there >>> seems to be something unusual. We have a little memory check during loading >>> the matrix to keep track of rss. The printout of rss in the log shows >>> normal increase up to rank 2160, i.e., if we load in a portion of matrix >>> that is 1GB, after MatSetValues for that portion, rss will increase roughly >>> about that number. From rank 2161 onwards, the rss in every rank doesn't >>> increase after matrix loaded. Then comes MatAssemblyBegin, the program >>> crashed on rank 2160. Is there a upper limit on the number of processes >>> Petsc can handle? or is there a upper limit in terms of the size of the >>> matrix petsc can handle? Thank you very much for any info. It sounds like >>> you overflowed int somewhere. We try and check for this, but catching every >>> place is hard. Try reconfiguring with --with-64-bit-indices Thanks, >>> Matt Regards, Karl -- What most experimenters take for >>> granted before they begin their experiments is infinitely more interesting >>> than any results to which their experiments lead. -- Norbert Wiener >>> https://www.cse.buffalo.edu/~knepley/ >>> <https://urldefense.com/v3/__http:/www.cse.buffalo.edu/*knepley/__;fg!!Kjv0uj3L4nM6H-I!1KBn92fUc-8pAvJy257WTFoHD80IUf6u5iIhyL_vrliEm3psAK4KAJFCdygnPA$> >>> -- What most experimenters take for granted before they begin their >>> experiments is infinitely more interesting than any results to which their >>> experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/ >>> <https://urldefense.com/v3/__http:/www.cse.buffalo.edu/*knepley/__;fg!!Kjv0uj3L4nM6H-I!1KBn92fUc-8pAvJy257WTFoHD80IUf6u5iIhyL_vrliEm3psAK4KAJFCdygnPA$> >>> -- What most experimenters take for granted before they begin their >>> experiments is infinitely more interesting than any results to which their >>> experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/ >>> <https://urldefense.com/v3/__http:/www.cse.buffalo.edu/*knepley/__;fg!!Kjv0uj3L4nM6H-I!1KBn92fUc-8pAvJy257WTFoHD80IUf6u5iIhyL_vrliEm3psAK4KAJFCdygnPA$> >>> Schlumberger-Private * >>> >> >> -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
