Thank you very much Bert for your kind reply.
I will try both your suggestions (running up on just only 8 cores as well as increasing the kernel.shmmax to 24 Gbyte or more).
If nothing changes in a positive way, I will wait for the new release which adresses this issue.
All the best,
Davide
Quote:Bert Apr 17th 2:47 pmDavide,
The code clearly tries to allocate more then 8 Gbyte of shared memory, and in your NWChem input you ask for a maximum up to 24x1Gbyte of shared memory to be available.
One thing you can try is to see if the calculation runs with a maximum of 8 cores per node, which would force the calculation to stay within the shemmax limit you have set.
Alternatively, you could increase the kernel.shmmax to 24 Gbyte. This will probably lead to the code generating multiple shared memory segments of the size ARMCI_DEFAULT_SHMMAX. The development version is addressing this issue, and another issue I will come back to with Mef362.
Mef262,
What are your kernel.shmmax and ARMCI_DEFAULT_SHMMAX settings? In your case, you should have the kernel.shmmax set to over 4 Gbyte and have ARMCI_DEFAULT_SHMMAX=4096.
Both,
We have seen some large memory usage in the TDDFT in NWChem 6.1 versions. We have addressed this in the TDDFT module and the global array parallel infrastructure of the development version that is being prepared for release as we speak.
Bert |