normal High parallelization

  • khaoj
  • khaoj's Avatar Topic Author
  • Offline
  • Fresh Boarder
More
13 years 7 months ago #393 by khaoj
High parallelization was created by khaoj
Hello,

I am currently simulating systems containing 33K beads on Martini and Elnedyn, and I can't get parallelization above 128 CPUs when changing -rdd (257 beads/cpu). Is this as far as we can push it ? If not, is there a way to directly find the optimal rdd ?

Thanks,
Jonathan.

Please Log in or Create an account to join the conversation.

More
13 years 7 months ago #394 by xavier
Replied by xavier on topic High parallelization
Considering the number of particles your system is made of, you can not expect so much scaling with large numbers of CPUs. How is the efficiency with 128 CPUs?

You may be able to go to a higher number of CPUs using the -pd (particle decomposition) and conserve some reasonable scaling. Note however that pd is slower than dlb.

Last the number of particles is one criterion to consider but the main problem will be the lower-limit of the box size due to bonded interactions that have to be kept together.

XAvier.

Please Log in or Create an account to join the conversation.

  • khaoj
  • khaoj's Avatar Topic Author
  • Offline
  • Fresh Boarder
More
13 years 7 months ago #397 by khaoj
Replied by khaoj on topic High parallelization
Hi Xavier,
Indeed, when comparing the efficiency of my 128 CPU (rdd 1.8nm) with a 8CPU MD, I hardly reach 70%. I just tried the particle decomposition, but it's way slower than domain decomposition.

If I'm correct, to determine a good rdd in an Elnedyn-Martini simulation, I have to set the rdd above the maximum bond length (including the elastic network bonds). But g_bond only gives average bond values, so before I start scripting, are you aware of any scripts/tools that can do that easily ?

Thanks,
Jonathan.

Please Log in or Create an account to join the conversation.

More
13 years 6 months ago #409 by xavier
Replied by xavier on topic High parallelization
70% on 128 CPUs is not so bad with 250 particles per CPU!
I would be happy with it :))

XAvier.

Please Log in or Create an account to join the conversation.

Time to create page: 0.099 seconds