On 2/8/2013 7:37 AM, Max Fischer wrote:
As TJ mentioned, jobs with mismatched resource estimations are something you may tolerate but should generally not encourage.
Depends on the job: estimates lie all the time, that's one of the reasons overcommit was invented in the first place.
We have one that mmaps ~30GB search space in 2GB chunks and we run them 1/core. Last I looked linux normally mmaps files read-only in shared segments and COWs if needed -- in this case never.
So condor's estimate (and top's report) for these, esp. towards the end of the run, looks like 30.01GB each (times, say, 8 cores). The actual can be as low as 2.01GB and up to 16.01GB _for all 8_.
Dimitri