Hello,
I would like to ask a question regarding the scratch-management of the ricc2 module (Turbomole 6.0). We have quite large jobs running on a 32 CPU shared-memory architecture with 64GB of ram. After some testing we found that a $maxcor value larger than 1000 does not improve the CPU time, so we decided to invest the remaning 32GB into a ramdisk. Redirecting the $tmpdir to the ram significantly sped up the run by approx. 30%!!
The jobs run nicely, but for some configurations 32GB ramdisk is not sufficient and the ricc2 terminates as soon as the limit is reached. In the reply to the question of kabel999 from March 27, 2009 Christof Haettig explained, that there was no option to reduce the disc space requirements. Are the newer versions of ricc2 able to reduce the required disk space?
As all CPUs are within the same node, I tried to set the $sharedtmpdir in addition to the $tmpdir command. I was hoping that all processes can access the same data, thereby reducing the disk space requirement and (eventually) increase the performance.
The keywords are straightforward:
$tmpdir /ram/ricc2-tmp/
$sharedtmpdir /ram/ricc2-shared/
However, as soon as ricc2 is invoked, a new $SHAREDTMPDIR entry is automatically placed in the control file just before the $end keyword overriding the initial path assignment. The directory /ram/ricc2-shared/ is not created and the shared-scratch is not used. Only the folders /ram/ricc2-tmp-000/ to /ram/ricc2-tmp-031/ are found in the ramdisk and after the 32GB are reached, ricc2 terminates again.
Is it this behaviour of ricc2 to avoid the use $sharedtmpdir intended? Is there another way to reduce the disk space requirement or to motivate ricc2 to use shared scratch?
Best wishes and many thanks,
Thomas
