Hi Sherry, Thanks for your message. I have used superlu_dist default options. I did not realize that I was doing serial symbolic factorization. That is probably the cause of my problem. Each node on Garnet has 60GB usable memory and I can run with 1,2,4,8,16 or 32 core per node.
So I should use: -mat_superlu_dist_r 20 -mat_superlu_dist_c 32 How do you specify the parallel symbolic factorization option? is it -mat_superlu_dist_matinput 1 Thanks, Anthony On Tue, Jul 7, 2015 at 3:08 PM, Xiaoye S. Li <[email protected]> wrote: > For superlu_dist failure, this occurs during symbolic factorization. > Since you are using serial symbolic factorization, it requires the entire > graph of A to be available in the memory of one MPI task. How much memory > do you have for each MPI task? > > It won't help even if you use more processes. You should try to use > parallel symbolic factorization option. > > Another point. You set up process grid as: > Process grid nprow 32 x npcol 20 > For better performance, you show swap the grid dimension. That is, it's > better to use 20 x 32, never gives nprow larger than npcol. > > > Sherry > > > On Tue, Jul 7, 2015 at 1:27 PM, Barry Smith <[email protected]> wrote: > >> >> I would suggest running a sequence of problems, 101 by 101 111 by 111 >> etc and get the memory usage in each case (when you run out of memory you >> can get NO useful information out about memory needs). You can then plot >> memory usage as a function of problem size to get a handle on how much >> memory it is using. You can also run on more and more processes (which >> have a total of more memory) to see how large a problem you may be able to >> reach. >> >> MUMPS also has an "out of core" version (which we have never used) >> that could in theory anyways let you get to large problems if you have lots >> of disk space, but you are on your own figuring out how to use it. >> >> Barry >> >> > On Jul 7, 2015, at 2:37 PM, Anthony Paul Haas <[email protected]> >> wrote: >> > >> > Hi Jose, >> > >> > In my code, I use once PETSc to solve a linear system to get the >> baseflow (without using SLEPc) and then I use SLEPc to do the stability >> analysis of that baseflow. This is why, there are some SLEPc options that >> are not used in test.out-superlu_dist-151x151 (when I am solving for the >> baseflow with PETSc only). I have attached a 101x101 case for which I get >> the eigenvalues. That case works fine. However If i increase to 151x151, I >> get the error that you can see in test.out-superlu_dist-151x151 (similar >> error with mumps: see test.out-mumps-151x151 line 2918 ). If you look a the >> very end of the files test.out-superlu_dist-151x151 and >> test.out-mumps-151x151, you will see that the last info message printed is: >> > >> > On Processor (after EPSSetFromOptions) 0 memory: >> 0.65073152000E+08 =====> (see line 807 of module_petsc.F90) >> > >> > This means that the memory error probably occurs in the call to >> EPSSolve (see module_petsc.F90 line 810). I would like to evaluate how much >> memory is required by the most memory intensive operation within EPSSolve. >> Since I am solving a generalized EVP, I would imagine that it would be the >> LU decomposition. But is there an accurate way of doing it? >> > >> > Before starting with iterative solvers, I would like to exploit as much >> as I can direct solvers. I tried GMRES with default preconditioner at some >> point but I had convergence problem. What solver/preconditioner would you >> recommend for a generalized non-Hermitian (EPS_GNHEP) EVP? >> > >> > Thanks, >> > >> > Anthony >> > >> > On Tue, Jul 7, 2015 at 12:17 AM, Jose E. Roman <[email protected]> >> wrote: >> > >> > El 07/07/2015, a las 02:33, Anthony Haas escribió: >> > >> > > Hi, >> > > >> > > I am computing eigenvalues using PETSc/SLEPc and superlu_dist for the >> LU decomposition (my problem is a generalized eigenvalue problem). The code >> runs fine for a grid with 101x101 but when I increase to 151x151, I get the >> following error: >> > > >> > > Can't expand MemType 1: jcol 16104 (and then [NID 00037] 2015-07-06 >> 19:19:17 Apid 31025976: OOM killer terminated this process.) >> > > >> > > It seems to be a memory problem. I monitor the memory usage as far as >> I can and it seems that memory usage is pretty low. The most memory >> intensive part of the program is probably the LU decomposition in the >> context of the generalized EVP. Is there a way to evaluate how much memory >> will be required for that step? I am currently running the debug version of >> the code which I would assume would use more memory? >> > > >> > > I have attached the output of the job. Note that the program uses >> twice PETSc: 1) to solve a linear system for which no problem occurs, and, >> 2) to solve the Generalized EVP with SLEPc, where I get the error. >> > > >> > > Thanks >> > > >> > > Anthony >> > > <test.out-superlu_dist-151x151> >> > >> > In the output you are attaching there are no SLEPc objects in the >> report and SLEPc options are not used. It seems that SLEPc calls are >> skipped? >> > >> > Do you get the same error with MUMPS? Have you tried to solve linear >> systems with a preconditioned iterative solver? >> > >> > Jose >> > >> > >> > >> <module_petsc.F90><test.out-mumps-151x151><test.out_superlu_dist-101x101><test.out-superlu_dist-151x151> >> >> >
