Note: This discussion is about an older version of the COMSOL Multiphysics® software. The information provided may be out of date.
Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.
Batch processing bug : One computer vs Cluster grid
Posted 22 ott 2010, 04:00 GMT-4 Version 4.1 5 Replies
Please login with a confirmed email address before reporting spam
Dear all,
I have been struggling for some time to understand my problem, but I couldn't find an explanation. I tried to use the cluster and batch computing feature in comsol 4.0a. So I modeled a file that was running without problem on my computer. I launched it on the remote server via ssh tunneling and batch :
comsol batch -inputfile my_model.mph
And the solving went well. So, in order to save time, i wanted to use our cluster :
comsol -nn 4 mpd boot -f mpd.hosts
comsol -nn 4 bqtch -inputfile my_model.mph
This model contains a moving structure in fluid (module FSI). It could get inverted mesh from time to time, but in this case I already solved the problem on one computer without any inverted mesh.
PROBLEM: on the cluster, my_model.mph was not solved. The log file contain inverted meshes for 2 nodes (over 4 nodes).
Why was that the case? Why did I observed some inverted meshes on some nodes and not all of them? Why was the problem solved on only one computer and not all of them?
Thank you in advance
I have been struggling for some time to understand my problem, but I couldn't find an explanation. I tried to use the cluster and batch computing feature in comsol 4.0a. So I modeled a file that was running without problem on my computer. I launched it on the remote server via ssh tunneling and batch :
comsol batch -inputfile my_model.mph
And the solving went well. So, in order to save time, i wanted to use our cluster :
comsol -nn 4 mpd boot -f mpd.hosts
comsol -nn 4 bqtch -inputfile my_model.mph
This model contains a moving structure in fluid (module FSI). It could get inverted mesh from time to time, but in this case I already solved the problem on one computer without any inverted mesh.
PROBLEM: on the cluster, my_model.mph was not solved. The log file contain inverted meshes for 2 nodes (over 4 nodes).
Why was that the case? Why did I observed some inverted meshes on some nodes and not all of them? Why was the problem solved on only one computer and not all of them?
Thank you in advance
5 Replies Last Post 17 apr 2011, 01:37 GMT-4
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
well
I dont have an answer for your specific problem but it MIGHT have to do with the parametric solver itself
reason I think that is that I am currently experiencing the following problem
I have a complicated stationary problem multiphysics fully coupled
I want to vary One parameter...
if i ran the parameters one by one putting then by hand in the stationnary model [ and changing nothing else of course} all thecase converge and the solution is correct..
If I ran a matlab loop to vary the parameters ..... I get all my solutions again....
BUT if I want to use comsol parametric solver [ which I used often] it provides me with the first solution but then fail to find the other ones because of convergence issue!!
So, It seems that comsol parametric solver try to be ¨smart" on its one and i am currently struggling to find a way to force him to be stupid...
you might face the same problem where it adjust some default value on its own depending on what you tell him to do;;;
well, shared misery is still misery...
butbased on my experience, I will suggest to build a matlab loop instead of using the builtin parametric solver.
You can control your distribution on cluster node as you wish AND make sure the solver settings are not changed behind your back.... which I suspect is what is happening in my case.. still working on it
for whatever its worth....
JF
I dont have an answer for your specific problem but it MIGHT have to do with the parametric solver itself
reason I think that is that I am currently experiencing the following problem
I have a complicated stationary problem multiphysics fully coupled
I want to vary One parameter...
if i ran the parameters one by one putting then by hand in the stationnary model [ and changing nothing else of course} all thecase converge and the solution is correct..
If I ran a matlab loop to vary the parameters ..... I get all my solutions again....
BUT if I want to use comsol parametric solver [ which I used often] it provides me with the first solution but then fail to find the other ones because of convergence issue!!
So, It seems that comsol parametric solver try to be ¨smart" on its one and i am currently struggling to find a way to force him to be stupid...
you might face the same problem where it adjust some default value on its own depending on what you tell him to do;;;
well, shared misery is still misery...
butbased on my experience, I will suggest to build a matlab loop instead of using the builtin parametric solver.
You can control your distribution on cluster node as you wish AND make sure the solver settings are not changed behind your back.... which I suspect is what is happening in my case.. still working on it
for whatever its worth....
JF
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
your reply is exellent. I also maintain that using the matlab loop instead of the bulit-in parametric sweep.
but also have some question about the matlab loop
when I sweep for a large times, the matlab crashed, returning java.lang.OutOfMemoryError: Java heap space, even i change the java heap size.
Did you guys encounter the same question. if you have, how did you figure it out?
but also have some question about the matlab loop
when I sweep for a large times, the matlab crashed, returning java.lang.OutOfMemoryError: Java heap space, even i change the java heap size.
Did you guys encounter the same question. if you have, how did you figure it out?
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Well, actually my first model was not using parametric sweep (I planned to do the parametric study this week end). The problem appear on a simple time dependent problem...
But thank you for the hint, it will help me to save time.
But thank you for the hint, it will help me to save time.
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
your reply is exellent. I also maintain that using the matlab loop instead of the bulit-in parametric sweep.
but also have some question about the matlab loop
when I sweep for a large times, the matlab crashed, returning java.lang.OutOfMemoryError: Java heap space, even i change the java heap size.
Did you guys encounter the same question. if you have, how did you figure it out?
Hi,
I'm suffering the same problem now. When i do a parametric sweep ,only the first several values will lead to solutions,and then the solver will stop to warn " Error in sweep: Java heap space".
Have you resolved your problem or how did you make it? If i want to use loop in matlab replacing the built-in parametric sweep, what should i do?
Thanks
Best regards
Yulong Yang
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
well
I dont have an answer for your specific problem but it MIGHT have to do with the parametric solver itself
reason I think that is that I am currently experiencing the following problem
I have a complicated stationary problem multiphysics fully coupled
I want to vary One parameter...
if i ran the parameters one by one putting then by hand in the stationnary model [ and changing nothing else of course} all thecase converge and the solution is correct..
If I ran a matlab loop to vary the parameters ..... I get all my solutions again....
BUT if I want to use comsol parametric solver [ which I used often] it provides me with the first solution but then fail to find the other ones because of convergence issue!!
So, It seems that comsol parametric solver try to be ¨smart" on its one and i am currently struggling to find a way to force him to be stupid...
you might face the same problem where it adjust some default value on its own depending on what you tell him to do;;;
well, shared misery is still misery...
butbased on my experience, I will suggest to build a matlab loop instead of using the builtin parametric solver.
You can control your distribution on cluster node as you wish AND make sure the solver settings are not changed behind your back.... which I suspect is what is happening in my case.. still working on it
for whatever its worth....
JF
Hi,
I'm suffering the same problem now. When i do a parametric sweep ,only the first several values will lead to solutions,and then the solver will stop to warn " Error in sweep: Java heap space".
Have you resolved your problem or how did you make it? If i want to use loop in matlab replacing the built-in parametric sweep, what should i do?
Thanks
Best regards
Yulong Yang