[torqueusers] timing question
Greenseid, Joseph M (IS)
Joseph.Greenseid at ngc.com
Fri Apr 16 06:58:37 MDT 2010
First off, let me say I'm not sure if this is a Torque or Maui question.
I am setting up a system that is going to be running lots of small jobs (each job will likely be 8 or 16 cores, < 7 seconds in length). The time it takes to complete a job is critical, as there is a requirement for a whole workflow (in which the job execution is the largest step, but not the only one) to take 10 seconds, so we've been doing timing tests. I was using Maui's logs to see when jobs got queued, when they started execution, and when they finished execution. I was then comparing that to the output of running the `time` command against my mpirun command inside the script. On average, it takes about 2 extra seconds on top of the actual command being run for the scheduler/queue to do their thing. I'm guessing at least part of this is that something is polling every second or something like that. Is that something that I might be able to tune down to a smaller value? If that's not it, are there any other suggestions to get this time down as much as possible?
I know under normal circumstances, with jobs that run for hours and days, these amounts of time get lost in the noise, so I'd never thought about squeezing an extra second out of the system in this way before.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the torqueusers