[bug fixed v6.6] running abinit-6.4.3 on IBM power6
Posted: Mon Feb 28, 2011 3:41 pm
Hi,
I have compiled abinit-6.4.3 on IBM power6, I have been able to run the test successfully, then I run a G0W0 calculation with the plasmon pole model of Godby and Needs, which finished without any trouble.
Then I tried to run some calculation without any plasmon pole, i.e. using the numerical integration along the deformed path, but I got an overload of the memory (although I used 3GB/proc on 32 procs), which was unexpected, since my requirements on the number of bands, planewaves and k-points is reasonable and I could run it easily on another platform with only 4 procs with 2GB/proc (Intel(R) Xeon(R)).
I noticed that nothing is written in the output file after running during at least 15 minutes, is this normal?
Moreover in the standart output file, there is no complain, the job just stop
Since on the Intel cluster I compile without scalapack, I tried to compile with and without scalapack on the IBM platform, but this makes no difference.
Then I tried to change the fftalg, since the default one is different whether running on the IBM or Intel cluster. On Intel, the default fftalg=112 and fftalg=312 on IBM.
So when I changed from fftalg=312 to fftalg=112, the job run a bit further but then crashed again.
Now I don't know what to do to circumvent this problem.
I have attached my input file, and if someone wants to have a look on my output, here is the links to access them:
https://docs.google.com/leaf?id=0Bz-lOO ... NjQ5&hl=en (run on IBM with 32 procs without scalapack)
https://docs.google.com/leaf?id=0Bz-lOO ... NGRm&hl=en (run on IBM with 16 procs with scalapack)
https://docs.google.com/leaf?id=0Bz-lOO ... YWVk&hl=en (run on Intel with 4 procs with scalapack)
Any help would be greatly appreciated!
Thanks in advance
Best Regards
Pierre-Yves
I have compiled abinit-6.4.3 on IBM power6, I have been able to run the test successfully, then I run a G0W0 calculation with the plasmon pole model of Godby and Needs, which finished without any trouble.
Then I tried to run some calculation without any plasmon pole, i.e. using the numerical integration along the deformed path, but I got an overload of the memory (although I used 3GB/proc on 32 procs), which was unexpected, since my requirements on the number of bands, planewaves and k-points is reasonable and I could run it easily on another platform with only 4 procs with 2GB/proc (Intel(R) Xeon(R)).
I noticed that nothing is written in the output file after running during at least 15 minutes, is this normal?
Moreover in the standart output file, there is no complain, the job just stop
Since on the Intel cluster I compile without scalapack, I tried to compile with and without scalapack on the IBM platform, but this makes no difference.
Then I tried to change the fftalg, since the default one is different whether running on the IBM or Intel cluster. On Intel, the default fftalg=112 and fftalg=312 on IBM.
So when I changed from fftalg=312 to fftalg=112, the job run a bit further but then crashed again.
Now I don't know what to do to circumvent this problem.
I have attached my input file, and if someone wants to have a look on my output, here is the links to access them:
https://docs.google.com/leaf?id=0Bz-lOO ... NjQ5&hl=en (run on IBM with 32 procs without scalapack)
https://docs.google.com/leaf?id=0Bz-lOO ... NGRm&hl=en (run on IBM with 16 procs with scalapack)
https://docs.google.com/leaf?id=0Bz-lOO ... YWVk&hl=en (run on Intel with 4 procs with scalapack)
Any help would be greatly appreciated!
Thanks in advance
Best Regards
Pierre-Yves