Performance

[litman90]

Hello,

I am trying to compute farthest point sampling for my data set. It contains ~10000 frames and I want to get 1000 points. Currently, I am running in a cluster with 10 nodes (400 CPU) and in 11 hours it computes only the first ~ 130 points. I am using the following flags: --kernel rematch -n 12 -l 8 --gamma 0.1 -c 4.5 --peratom --distance and my system contains 38 atoms. Is this the expected performance? Is there any trick I can do to accelerate the calculation beyond reducing the number of considered atoms and/or the data set size?

Regards, Yair

[ceriottm]

use --kernel fastavg. rematch has to do a ton of calculations.if you really need rematch (which tends to be better to disambiguate similar structures) you should make sure that you have compiled the cython library for sinkhorn, otherwise glosim will fall back on a python version which brings slow to another level.

On Wed, 5 Dec 2018 at 16:24, litman90 [email protected] wrote:

Hello,

I am trying to compute farthest point sampling for my data set. It contains ~10000 frames and I want to get 1000 points. Currently, I am running in a cluster with 10 nodes (400 CPU) and in 11 hours it computes only the first ~ 130 points. I am using the following flags: --kernel rematch -n 12 -l 8 --gamma 0.1 -c 4.5 --peratom --distance and my system contains 38 atoms. Is this the expected performance? Is there any trick I can do to accelerate the calculation beyond reducing the number of considered atoms and/or the data set size?

Regards, Yair

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/cosmo-epfl/glosim/issues/13, or mute the thread https://github.com/notifications/unsubscribe-auth/ABESZ6JSHS23RH97JX1e08awh2utewxAks5u1-UngaJpZM4ZC9EL .