Presentation is loading. Please wait.

Presentation is loading. Please wait.

October 14, 2005Guassian Performance Workshop at Jackson State University The Mississippi Center for Supercomputing Research Gaussian Performance Workshop.

Similar presentations


Presentation on theme: "October 14, 2005Guassian Performance Workshop at Jackson State University The Mississippi Center for Supercomputing Research Gaussian Performance Workshop."— Presentation transcript:

1 October 14, 2005Guassian Performance Workshop at Jackson State University The Mississippi Center for Supercomputing Research Gaussian Performance Workshop At Jackson State University

2 October 14, 2005Guassian Performance Workshop at Jackson State University Topics Estimating Parallel Efficiency with qstat –f Emails to Researchers Parallel Efficiency Example New g03sub

3 October 14, 2005Guassian Performance Workshop at Jackson State University qstat –f Redwood Example qstat -f 7508 | grep resources resources_used.cpupercent = 601 resources_used.cput = 96:32:39 resources_used.mem = 421584kb resources_used.ncpus = 2 resources_used.vmem = 846304kb resources_used.walltime = 53:15:14 96 / 53 = …. Speedup = 1.81 1.81 / 2 = … Efficiency = 90.5%

4 October 14, 2005Guassian Performance Workshop at Jackson State University qstat –f Sweetgum Example sweetgum> qstat -f 466 | grep resources resources_used.cpupercent = 98 resources_used.cput = 133:48:31 resources_used.mem = 360320kb resources_used.ncpus = 4 resources_used.vmem = 1082416kb resources_used.walltime = 136:04:22 133 / 136 = …. Speedup = 0.98 0.98 / 4 = … Efficiency = 24.5%

5 October 14, 2005Guassian Performance Workshop at Jackson State University Emails to Researchers http://www.mcsr.olemiss.edu/admin/pbsjobs/pbsjobs.php -Job id 466 -Job id 7508

6 October 14, 2005Guassian Performance Workshop at Jackson State University Mimosa Parallel Efficiency - Can’t use resources_used.cput - Can’t trust resources_used.cpupercent - Run same job on 1, 2, 4 nodes & compare wallclock times - qstat –f | grep exec_host - rsh to nodes and execute “top”

7 October 14, 2005Guassian Performance Workshop at Jackson State University Mimosa Example - Gaussian job took 8 days on 8 nodes - Same job took 11 days on 4 nodes - Same job took 13 days on 1 node - 8 node speedup: 1.5x - 8 node parallel efficiency: 19% - 4 node speedup: 1.3x - 4 node parallel efficiency: 64%

8 October 14, 2005Guassian Performance Workshop at Jackson State University Mimosa Example - User saved 5 days by running on 1 node - If user had only 8 nodes… - And had 8 similar jobs to run… - If he runs them as 1 node jobs… - He can finish all 8 in 13 days. - If he runs then as 8 node jobs… - It will take 64 days to finish them all.

9 October 14, 2005Guassian Performance Workshop at Jackson State University newg03sub - Validates input file memory with PBS - Validates input file %nprocl and %nprocs with PBS - Adds prefix to job name based on calculation type - Names scratch file directories based on jobid - Takes additional input parameter: max disk/scratch file - Example: -/ptmp/jghale/benchmarking/sweetgum/h artree-fock/medium/first/400mb/2proc


Download ppt "October 14, 2005Guassian Performance Workshop at Jackson State University The Mississippi Center for Supercomputing Research Gaussian Performance Workshop."

Similar presentations


Ads by Google