Home My Page Projects PaStiX
Summary Activity Forums Lists Docs News Files

Forum: help

Monitor Forum | Start New Thread Start New Thread
RE: Factorization step hangs in larger problems [ Reply ]
By: Nobody on 2015-06-11 15:50
[forum:149635]

examples.tar.bz2 (6) downloads
Hello,

No problem, I've attached both files here. If you have any other issues with the files, or if you would prefer I submit them in a different manner (email, etc.) let me know.

Thanks,
Darin

RE: Factorization step hangs in larger problems [ Reply ]
By: Théophile Terraz on 2015-06-11 15:28
[forum:149633]
Hello,

Attachments were lost because of a bug of the forum.
Could you re-post your step-by-step_dist.c please?
The same problem occured for the Lx=b post.

With my apologies,

Théophile

RE: Factorization step hangs in larger problems [ Reply ]
By: Nobody on 2015-06-04 19:17
[forum:149631]
Hello,

I added a subroutine to the step-by-step_dist example to generate the matrix that was producing the hang. It's a bit crude, partially because I've only ever worked in c++ before, but it does produce the hang. It runs with a mesh size of 350x350 elements (I commented out the part of the code that reads in the options, but it should still work whether it is commented or not), and I've checked to make sure that it consistently produces the hang, usually after about 100 iterations of the factorization. I also tried to see if the laplacian example might produce the hang as well, but it ran out of memory before the matrix became large enough to hang.

As a side note, I also got a hang in the cscd_redispatch function if I increased to a 400x400 mesh. I doubt it's a major concern because that function seems like it might be inefficient to use outside of the example, but I didn't know if it might be related.

Hopefully that is enough to help you out, but if you need more information from me or if the file doesn't make any sense, please let me know.

Thanks,
Darin

RE: Factorization step hangs in larger problems [ Reply ]
By: Théophile Terraz on 2015-06-02 15:03
[forum:149609]
Hello,

It would help us if you could reproduce the hang with the step-by-step_dist.c exemple available in PaStiX.
In the code, comment the 5.1 and the 5.2 steps, give a hight number of factorizations 'nfacts' to match your needs, then run your matrix with this example.

If the hang happens, could you give us your matrix, so we can try to fix it ?

Thanks,

Théophile

Factorization step hangs in larger problems [ Reply ]
By: Nobody on 2015-05-29 16:33
[forum:149608]
Hello,

I'm using Pastix in an optimization code and I have to repeatedly solve a linear system where the matrix is constantly updating. It works great for smaller problems or if I'm only using a small number of processors. However, for large problem sizes, the factorization step will sometimes hang if I use more than a few processors. In this case, by a "large" problem, I mean a matrix dimension of a few hundred thousand with a few million nonzero elements. If I run a smaller problem (dimension in ten thousands and a few hundred thousand nonzero elements), or if I launch less than 6 threads, everything works perfectly.

The hang doesn't occur after a certain number of iterations, it is seemingly random, but it does occur sooner in bigger problems or with more threads. I've played with my own code a bit to see if some other issue is playing into it, but it seems that the issue occurs regardless of what else is going on. Just looping over the lines below will produce the hang, so it seems to me that the issue is somewhere within Pastix.

iparm[IPARM_START_TASK] = API_TASK_NUMFACT;
iparm[IPARM_END_TASK] = API_TASK_NUMFACT;

dpastix(&pastix_data, comm,
ncol, colptr, rows.data(), values.data(), loc2glob,
perm, invp, rhs, 1, iparm, dparm);

(My code is written and compiled in C++, so the rows and values arrays are stored as std::vectors to make things a little more flexible).

I'm sure this probably isn't enough information to identify the problem, so please let me know what other information is relevant. Any help you can offer would be much appreciated.

Thanks,
Darin