Am 23.03.2011 um 14:26 schrieb Mark Dixon:

>> When "cpu" and "mem" is already the reserved integral value, you can compute 
>> this "maxvmem" easily affterwards. There it no need to set it IMO.
> 
> They completely different things, aren't they? Isn't maxvmem the peak vmem 
> used during the job, whereas mem ought to be the time integral?

It should be. But as you said and I checked, it's set to the h_vmem request. So 
I wanted to point out, that the current implementation is redundant:

$ qacct -j 1234
...
cpu          120.000      
mem          90.000            
io           0.000             
iow          0.000             
maxvmem      768.000M

92160 / 120 = 768

-- Reuti


> Besides, even if you can calculate it from the other fields, having maxvmem 
> is so much easier to read for the users e.g. to see how much vmem they used, 
> so they know what to ask for in subsequent similar jobs.
> 
> Back in 6.0-land, it also was the most reliable method I could find to see if 
> your job failed due to exceeding any requested h_vmem.
> 
> 
>>> 2) It's interesting to have the actual cpu time in the accounting file, so 
>>> that you can see if cpu time / run time ~= 1.0
>> 
>> Yes, as long as the jobs are not reprioritized with different nice values 
>> and/or the exechosts are oversubscribed, i.e. you have slots=cores.
>> 
>> You would charge for the run time then (in case you do it)?
> 
> We charge for elapsed time, which we get from ru_wallclock.

_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to