Package: gridengine* Version: 6.2u5-1
When a job has finished, the "max vmem" value as it was actually used is reported to the user. Unfortunately, the reported value is never larger than 4G also the application has clearly used more than 4G of Memory. I've been testing this with Octave as well as mathematica. Based on some preliminary tests, it seems that the reported value is the (real value module 2^32). Thus, if more then 4G of vmem are used, the reported maxvmem value is meaningless. attached are: - an Octave sample script, allocating 600M of random double precision numbers resulting in 4.8GB memory need. - the script for submission to the grid engine qsub memtest03.m The e.mail report contains this: Job 4709696 (MEMTEST02) Complete User = schloegl Queue = mai...@nick13.ista.local Host = nick13.ista.local Start Time = 01/16/2013 11:19:11 End Time = 01/16/2013 11:19:36 User Time = 00:00:23 System Time = 00:00:02 Wallclock Time = 00:00:25 CPU = 00:00:25 Max vmem = 710.246M Exit Status = 0 The correct value of "Max vmem" should be 4.47GB or more. But its close to mod(8*6e8,2^32)=505032704 (this is plausible when we assume 200MB overhead by Octave). maxvmem is an important variable of optimizing the use of a cluster. Users can control how much memory is is required, which is then used by the grid engine for scheduleing. For this reason, its important that a meaningful value for maxvmem is reported. Therefore I suggest a severity level "important". Severity: important Alois
% Octave script clear x=randn(6e8,1); size(x) whos exit
memtest03.sh
Description: Bourne shell script