[pymvpa] Time for single permutation varies between a few seconds and several minutes

[J.A. Etzel]

I don't know if it could be the case here, but I also run large
(non-pymvpa) memory jobs on a large multiple-cluster computer, and the 
run time of equivalent jobs sometimes varies quite a bit. The cluster
administrator said that the cause is the overall machine load: if the
machine is running at near-full capacity it has to slow or pause
operations to avoid overheating, and so my jobs take longer to finish.

Jo



On 4/8/2015 7:55 AM, Yaroslav Halchenko wrote:
>
> On Wed, 08 Apr 2015, Nick Oosterhof wrote:
>
>
>> On 08 Apr 2015, at 14:03, Jan Derrfuss <derrfuss at gmail.com> wrote:
>
>>> I'm stumped. I'm running a permutation for a searchlight analysis
>>> and print the time when the individual permutations start. I
>>> turns out that the onset times are sometimes mere seconds and
>>> sometimes several minutes apart.
>
>> Very quick (and maybe silly) idea: could this be due to disk i/o?
>> When running a multi-core searchlight, data is stored to disc and
>> and then loaded by each child process. One possibility is that
>> after a few children have started, different processes are
>> competing for disk i/o, slowing everything down.
>
> great idea... but unless you know what I don't know -- default
> backend='native' so it doesn't go through files (once again -- we
> are yet to see the code)
>