On Fri, Mar 17, 2017 at 09:20:50AM +0100, Tru Huynh wrote:
> I am currently running the CC35 and will run CC52 just after :P.
>
...
> The elapsed time was dramtically reduced! 1+8 cores, 4 gpus:
> 2.0.3: 130 munites elapsed CC61
> git-201703015-d401f24: <89 minutes! with CC35
relion-git-201703015-d401f24-openmpi-1.10.6-libltdl-CC-35.bench
4+1 with 4 gpus: ~121-122 minutes elapsed
8+1 with 1 gpu : ~194-196 minutes elapsed
8+1 with 2 gpus: ~119 minutes elapsed
8+1 with 3 gpus: ~97-98 minutes elapsed
8+1 with 4 gpus: ~87-88 minutes elapsed
relion-git-201703015-d401f24-openmpi-1.10.6-libltdl-CC-52.bench
4+1 with 4 gpus: ~120 minutes elapsed
8+1 with 1 gpu : ~190-191 minutes elapsed
8+1 with 2 gpus: ~116-117 minutes elapsed
8+1 with 3 gpus: ~96 minutes elapsed
8+1 with 4 gpus: ~86-87 minutes elapsed
conclusions:
- simular performance can be obtained with 8+1/2gpus and 4+1/4 gpus,
- targetting CC 3.5 or 5.2 is borderline on the TitanX/Pascal.
I still need to run the scaling benchmark for 4+1 on 1/2/3 gpus.
Cheers
Tru
--
Dr Tru Huynh | http://www.pasteur.fr/research/bis
mailto:[log in to unmask] | tel/fax +33 1 45 68 87 37/19
Institut Pasteur, 25-28 rue du Docteur Roux, 75724 Paris CEDEX 15 France
|