Discussion Forum

July 6, 2010 11:43pm UTC

endlessly assembling matrices in transient solver

Hello,

after fruitlessly trying almost every possible way to solve a Navier-stokes equation with volumic forces in 3D with very few ram (2GB), and thus putting my software to all the stresses possible, it ended up that now it just refuses to go beyond the first time step in transient analysis, but not because the solver runs out of memory, but because, after solving "successfuly" the first linear system, the solver keeps going by endlessly "assembling matrices", with percentages always in the same set (46-93-97-100 etc.), and this for minutes long, with the time parameter never taking off 0.actually this first happened when, after several trials, I tried UMFPACK again, knowing that it should not succeed and run out of memory around 60%, because that already happened before with same model and same mesh, but I tried it nonetheless, and it surprisingly succeeded going all the way to 100%, and there, it started the endless "matrix assembly". And, I noticed there a weird fact: I always monitor the RAM usage by comsol, and there it appears that UMFPACK allocated itself a lot more ram than usual, which is of course good for me, and is why probably it succeeded to "solve the system", but I still don't understand why it came to this problem of endless assembly. At this run, the "memory allocation factor" was set to 1, and after the bug I lowered it to 0.8, while restarting comsol: there, UMFPACK allocated the same few memory as usual (at least 3 times less), and ran out of it. then I tried Pardiso, and SPOOLES. And SPOOLES succeeded again, even with a restart, but again, I got my endless "assemblies".
Another point: before that I was trying to get something from the GMRES solver: after an incomplete LU, it solves correctly its first 2 systems, then is unable to converge on the third one. At first I thought the third system was for one of the variables, and I immediately thought about the pressure, which was completely unspecified, so I specified it on one vertice. But it did not solve the problem: GMRES is still not converging on the third system. That was why I re-tried the direct solvers, with all the weird things listed above...

Please, have anybody seen any similar behavior, like endlessly assembling matrices in transient navier-stokes analysis in 3D? because I have never seen that before...

Thank you very much for your help, of any kind

Jean-Pierre

Reply  |  Reply with Quote  |  Send private message  |  Report Abuse

August 3, 2011 10:59pm UTC in response to Tero Isotalo

Re: endlessly assembling matrices in transient solver

did you find the solution to this problem?

Thanks.

Reply  |  Reply with Quote  |  Send private message  |  Report Abuse