matlab slowness
predragp at andrew.cmu.edu
predragp at andrew.cmu.edu
Sun Feb 8 19:09:55 EST 2015
> I just submitted the paper I was working on (which used a bunch of
matlab simulations) so I've got some free time to help debug this. Not
really sure where to start though as I know very little about matlab
configurations. I suppose something to verify is whether this is an
issue for people not in Neill's group? Is there an autonlab member who
uses a lot of Matlab and would thus be self-interested? I don't know
Chun-Liang Li but I think he uses matlab a lot...
>
> Best,
> Seth
>
I am forwarding you e-mail to users at autonlab.org hopping that somebody can
help.
Most Kind Regards,
Predrag
> On Mon, Feb 2, 2015 at 4:55 PM, Seth Flaxman <flaxman at gmail.com> wrote:
>> Sure. I compared lot1, low1, and lov3:
>>
>> lot1
>>>> bench(5)
>> ans =
>>
>> 2.4679 0.5507 0.2113 38.6126 0.9343 0.8627
>> 2.0373 0.5239 0.1893 30.0908 1.1267 0.6388
>> 1.9278 0.5322 0.1908 32.5306 0.4670 0.7822
>> 3.1591 0.4540 0.2404 41.5550 1.6787 0.9064
>> 3.2059 0.5684 0.1988 25.0799 0.9904 1.6566
>>
>> low1
>>>> bench(5)
>> ans =
>>
>> 1.0779 0.2725 0.1980 22.7329 0.4850 0.5008
>> 1.3599 0.2884 0.2047 26.0477 0.4822 0.4942
>> 1.2960 0.2476 0.1984 27.1457 0.4907 0.4321
>> 1.2698 0.2802 0.1980 15.6125 0.5065 0.4892
>> 1.0454 0.2650 0.1989 10.4417 0.5208 0.5081
>>
>> lov3
>>>> bench(5)
>> ans =
>>
>> 0.3191 0.1766 0.2183 1.7567 0.4645 0.4663
>> 0.2828 0.1779 0.2291 1.5642 0.4895 0.4726
>> 0.3462 0.1938 0.2316 1.8131 0.4603 0.4967
>> 0.2673 0.1896 0.2086 1.7520 0.4870 0.4616
>> 0.3018 0.1713 0.2180 1.8592 0.4587 0.4514
>>
>>
>> Another clue, perhaps, is that if I start up 5 or 6 instances of matlab
then the 5th or 6th will often crash immediately. Matlab seems to do
very weird things with forking. I'm attaching one of those dumps, which
I just got when I tried to run the above.
>> Seth
>>
>> On Mon, Feb 2, 2015 at 4:32 PM, <predragp at andrew.cmu.edu> wrote:
>>>> Any updates? lov3 and lov4 seem, if anything, to be even slower
especially
>>>> for sparse linear algebra:
>>>>
>>>> lov3:
>>>>>> bench
>>>> 0.6941 0.3787 0.2958 14.4527 0.5514 0.5435
>>>> lov4:
>>>>>> bench
>>>> 1.0360 0.5322 0.2471 16.4350 0.5510 0.6338
>>>>
>>>> Does matlab rely on system packages that are architecture specific, like
>>>> ATLAS?
>>>>
>>>> Best,
>>>> Seth
>>>>
>>>
>>> I have not had chance to work on this stuff but I did little thinking. I
>>> am starting to worry that having SSD HDD in those two computing nodes
what
>>> not such a smart idea. Could you do me a favour and check how the
performance compares to LOW1 which has regular HDD.
>>>
>>> Predrag
>>>
>>>> On Wed, Jan 14, 2015 at 10:31 PM, <predragp at andrew.cmu.edu> wrote:
>>>>
>>>>> > Hah, okay. The fact that it's 0.27 seconds on CMU servers and 33
seconds on low1 seems quite strange.
>>>>> >
>>>>>
>>>>> Just to let you know. I am so busy until the end of next week that
there
>>>>> is not chance I could look into it before than. I apologize for any
inconvenience.
>>>>>
>>>>>
>>>>> Predrag
>>>>>
>>>>> > I found another benchmarking script here:
>>>>> > http://math.ucdenver.edu/~jmandel/matlab/matlab_benchmark.html
>>>>> >
>>>>> > Here's lov3:
>>>>> >
>>>>> > 1. large LU: 0.860
>>>>> > 2. small LU: 0.944
>>>>> > 3. sparse : 0.302
>>>>> >
>>>>> > Here's low1:
>>>>> >
>>>>> > 1. large LU: 1.813
>>>>> > 2. small LU: 0.526
>>>>> > 3. sparse : 2.216
>>>>> >
>>>>> > And here's linux.gp.cs.cmu.edu:
>>>>> >
>>>>> > 1. large LU: 0.569
>>>>> > 2. small LU: 0.491
>>>>> > 3. sparse : 0.176
>>>>> >
>>>>> > Seth
>>>>> >
>>>>> > On Mon, Jan 12, 2015 at 9:29 PM, <predragp at andrew.cmu.edu> wrote:
>>>>> >>> The warning matlab issues when I start it seems worth
>>>>> investigating
>>>>> >>> further:
>>>>> >>>
>>>>> >>> MATLAB is selecting SOFTWARE OPENGL rendering.
>>>>> >>>
>>>>> >>> < M A T L A B (R) >
>>>>> >>> Copyright 1984-2014 The MathWorks, Inc.
>>>>> >>> R2014b (8.4.0.150421) 64-bit (glnxa64)
>>>>> >>> September 15, 2014
>>>>> >>>
>>>>> >>> Warning: MATLAB has disabled some advanced graphics rendering
>>>>> features
>>>>> >>> by switching to software OpenGL. For more information, click <a
href="matlab:opengl('problems')">here</a>.
>>>>> >>>
>>>>> >>> (I'm running matlab at the command-line by ssh'ing into lop1
then low1.)
>>>>> >>>
>>>>> >>> Seth
>>>>> >>>
>>>>> >>
>>>>> >> The fact that we are using Open Graphics Library should not impact
>>>>> your
>>>>> >> code which does non graphics related computations. The fact that
MathWork
>>>>> >> favours closed binary blobs graphics drivers over OpenGL is very
telling.
>>>>> >> The papers produced by binary blobs are immediately rejected from
>>>>> all
>>>>> >> Mathematics Journals because there is no way to properly referee
>>>>> and
>>>>> >> scrutinize somebody work which uses closed source code.
>>>>> >>
>>>>> >>
>>>>> >> Predrag
>>>>> >>
>>>>> >>
>>>>> >>> On Mon, Jan 12, 2015 at 9:05 PM, Seth Flaxman
<flaxman at gmail.com> wrote:
>>>>> >>>> Fair enough. I think the other server comparisons are most
>>>>> relevant,
>>>>> >>>> but I don't have the full specs on these. Here's what I can
>>>>> tell:
>>>>> >>>>
>>>>> >>>> linux.andrew is running RHEL 6.5 / Matlab R2013a:
>>>>> >>>>
>>>>> >>>> [sflaxman at unix5 ~]$ matlab -nodesktop -nosplash -r "bench(1)"
>>>>> >>>> 0.0631 0.0962 0.3625 0.2873 0.4062 0.0177
>>>>> >>>>
>>>>> >>>> linux.gp.cs is running Ubuntu / Matlab R2014a
>>>>> >>>>
>>>>> >>>> 0.0825 0.1056 0.3558 0.2302 0.3217 0.0124
>>>>> >>>>
>>>>> >>>> phd-serv4.heinz.cmu.edu is running Fedora / R2014a
>>>>> >>>>
>>>>> >>>> 0.0671 0.0844 0.2428 0.2703 0.2975 0.0067
>>>>> >>>>
>>>>> >>>> And I just tried low1:
>>>>> >>>>
>>>>> >>>> 1.5470 0.2842 0.2995 33.8112 0.6661 0.5224
>>>>> >>>>
>>>>> >>>> Weird! Column 4 is for Sparse linear algebra...
>>>>> >>>>
>>>>> >>>> Maybe the issue is with R2014b? R2014a isn't still available on
>>>>> any
>>>>> of
>>>>> >>>> the autonlab servers for comparison is it?
>>>>> >>>>
>>>>> >>>> Seth
>>>>> >>>>
>>>>> >>>> On Mon, Jan 12, 2015 at 8:35 PM, <predragp at andrew.cmu.edu>
>>>>> wrote:
>>>>> >>>>> Oh and I forgot to say. If a single core on my server is slower
>>>>> than
>>>>> >>>>> a
>>>>> >>>>> single core on your laptop then I would expect your LAPTOP to
>>>>> execute
>>>>> >>>>> code
>>>>> >>>>> faster. The catch is that server can do massive parallel
>>>>> processing
>>>>> >>>>> and
>>>>> >>>>> that what MATLAB is very good for.
>>>>> >>>>>
>>>>> >>>>> Predrag
>>>>> >>>>>
>>>>> >>>>>
>>>>> >>>>>
>>>>> >>>>>
>>>>> >>>
>>>>> >>
>>>>> >>
>>>>> >
>>>>>
>>>>>
>>>>>
>>>>
>>>
>>>
>
More information about the Autonlab-users
mailing list