Re: [AMBER] degraded GPU performance

From: Ross Walker <ross.rosswalker.co.uk>
Date: Tue, 22 Sep 2015 07:41:47 -0700

Hi Alessandro,

Check the fan on the GPU. Has it seized or is it running slowly / smothered in dust? These GPUs can clock down if they hit a temperature limit so it is possible your GPU is massively overheating due to a failed cooling system. The other possibility is it is not receiving enough power through the PCI-E connectors. I'd start by removing the card and cleaning it extensively - put it in another machine if you can (you mentioned a Titan-Black system) and see if it still performs slowly in that environment.

All the best
Ross

> On Sep 22, 2015, at 7:22 AM, Alessandro Contini (alessandro.contini) <alessandro.contini.unimi.it> wrote:
>
> Dear users/developers,
> I'm experiencing a performance issue with pmemd.cuda over a EVGA GTX680 4GB. The GPU worked fine until this summer, then I started experiencing very low performance in running MDs. I tried to recompile Amber (all updated are installed), reinstall drivers and so on, but nothing changed. The strange thing is that the card passes the GPU_Validation_Test (https://dl.dropboxusercontent.com/u/708185/GPU_Validation_Test.tar.gz), obtaining identical results with all 20 tests, taking the same time as it did almost one year ago:
>
> nov 12 2014
> | ns/day = 76.65 seconds/ns = 1127.26
> | -----------------------------------------------------
>
> | Setup CPU time: 0.28 seconds
> | NonSetup CPU time: 1125.99 seconds
> | Total CPU time: 1126.27 seconds 0.31 hours
>
> | Setup wall time: 0 seconds
> | NonSetup wall time: 1128 seconds
> | Total wall time: 1128 seconds 0.31 hours
>
>
> sept 22 2015
> | ns/day = 76.56 seconds/ns = 1128.57
> | -----------------------------------------------------
>
> | Setup CPU time: 0.42 seconds
> | NonSetup CPU time: 1127.14 seconds
> | Total CPU time: 1127.56 seconds 0.31 hours
>
> | Setup wall time: 1 seconds
> | NonSetup wall time: 1129 seconds
> | Total wall time: 1130 seconds 0.31 hours
>
> However, when I try it on my examples, I observe that it has about 1/10 of the performance it use to have:
>
> feb 4 2015 3Z1A_complex_wat.equil3.out
> | ns/day = 83.41 seconds/ns = 1035.90
> | -----------------------------------------------------
>
> | Setup CPU time: 0.33 seconds
> | NonSetup CPU time: 102.64 seconds
> | Total CPU time: 102.97 seconds 0.03 hours
>
> | Setup wall time: 0 seconds
> | NonSetup wall time: 104 seconds
> | Total wall time: 104 seconds 0.03 hours
>
>
> set 22 16:08 3Z1A_complex_wat.equil3.out
> | ns/day = 7.03 seconds/ns = 12296.33
> | -----------------------------------------------------
>
> | Setup CPU time: 0.34 seconds
> | NonSetup CPU time: 1228.39 seconds
> | Total CPU time: 1228.73 seconds 0.34 hours
>
> | Setup wall time: 0 seconds
> | NonSetup wall time: 1230 seconds
> | Total wall time: 1230 seconds 0.34 hours
>
> I've tried it in different calculations (minimizations, MDs with or without restraints, aMD and so on) and I always experienced the same 1/10 of performance. Actually, I don't remember if the card started behaving like this after some software update, but my other machine which has same OS, same amber version, same NVIDIA driver version and mounts a Titan black card works as usual.
> Does anyone has an idea on what can be causing this problem?
>
> Thank you
>
> Alessandro
>
>
>
> --
> Alessandro Contini, PhD
> DISFARM - Sezione di Chimica Generale e Organica "Alessandro Marchesini"
> Università degli Studi di Milano
> Via Venezian, 21 20133 Milano
> Tel. +390250314480 Fax. +390250314476
> e-mail alessandro.contini.unimi.it
> skype alessandrocontini
>
> ****************
> Per destinare il 5x1000 all'Universita' degli Studi di Milano: indicare nella dichiarazione dei redditi il codice fiscale 80012650158.
>
> http://www.unimi.it/13084.htm?utm_source=firmaMail&utm_medium=email&utm_content=linkFirmaEmail&utm_campaign=5xmille
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Sep 22 2015 - 08:00:03 PDT
Custom Search