interactive vs batch performance

Kevin Ying

2016 3 17

0 个回答

7 次查看（30 天）

0 个投票

I run a single processor matlab code on our HPC batch system and compared the run with my interactive ssh matlab run. The systems used were identical. The same code 'ssh run' is 20%-30% faster than the 'qsub run'. Here is the part of the script run matlab (identical for both runs):

time matlab -nosplash -nodisplay -nodesktop -singleCompThread <<ENDXXX
addpath(genpath('/path/to/matlabcode'));
javaaddpath('/path/to/mysql-connector-java-5.1.35-bin.jar')
[ Diagnostic] = QAQC_RED_1hour
ENDXXX

I checked that there is no cpuset enforced in either case. Both runs open about the same set of files with minor differences between interactive session and a batch session. The only thing looks different is on top output that the 'qsub run' has a VIRT footprint about 20.3g where 'ssh run' has only 4.3g VIRT footprint.

PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                                                                              
29592 kying     20   0 20.3g 442m 108m S 41.5  0.3  26:58.30 MATLAB                                                                                                                
63728 kying     20   0 4305m 401m 108m S 34.6  0.3   7:29.35 MATLAB

Does matlab behave differently for batch vs interactive? Why batch is persistently slower than ssh interactive session? What could be the cause of it?