← Back to team overview

yade-users team mailing list archive

Re: Triax profiling on cluster


We are (finally...) starting to benchmark Yade parallel on 3SR cluster with Remi.
Just a few questions :
- from triax-perf.py, I understand we are not using the "fast" option by default, right? - are we starting the collider at each time step or do we use the version that is only working each N timesteps based on velocity extrema?
By default, fast is not used (unless set differently in the params table, if you use yade-multi). With fast, collider runs only every few steps (see source of TriaxialTest). Be aware however that due to fast motion, tthere is no effect of this in first few hundreds of steps.
As a preliminary result, Remi said "weird, it using very few cpu". He still investigating this.

It will run on 16 cores by default. Don't run it on more than 3 or 4 cores, or your performance will actually go down.

Cheers, V.

PS. I am at a http://www.launchpad.net/esys-particle/ course right now; I am sorry to say that, but if you need something high-performance, get esys; much better for computing than yade. They can run easily on 512 cores with almost linear scaling (meaning 10^6 particles in routine simulation). Why? They implemented relatively few algorithms (constitutive laws, grid-based collision detection, MPI parallelization), but they've done them well.