Project

General

Profile

Bug #3334

Check for DD when increasing rlist is broken

Added by Pascal Merz 8 months ago. Updated 7 months ago.

Status:
New
Priority:
Normal
Assignee:
Category:
mdrun
Target version:
Affected version - extra info:
Affected version:
Difficulty:
uncategorized
Close

Description

prepare_verlet_scheme(...), called from the runner, tries to increase the verlet buffer size (in increaseNstlist(...), pairlist_tuning.cpp). If domain decomposition is used, the trial buffer size is tested for DD (pairlist_tuning.cpp:294). The call to prepare_verlet_scheme does, however, occur before DD is set up, so this test for DD is never happening

History

#1 Updated by Berk Hess 8 months ago

  • Category set to mdrun
  • Assignee set to Berk Hess
  • Affected version - extra info deleted (probably also earlier versions are affected, but needs to be checked)

I wonder if this is really an issue.
Now rlist might limit the domain decomposition, whereas before the domain decomposition would limit rlist. Both have advantages and disadvantages.

#2 Updated by Szilárd Páll 8 months ago

Berk Hess wrote:

I wonder if this is really an issue.
Now rlist might limit the domain decomposition, whereas before the domain decomposition would limit rlist. Both have advantages and disadvantages.

Tricky tradeoff: this will limit scalability and cause more hassle for finding suitable DD setup. At the same time, for GPU simple non/less-scaling runs, we want even longer nslist which may require longer rlist.

I have no good idea what action to take now, but going forward, I think we might want to have the ability to strike a balance and tailor the automation around nstlist/rlist to take into account the DD requirements.

#3 Updated by Paul Bauer 7 months ago

  • Target version changed from 2020.1 to 2020.2

bumped

Also available in: Atom PDF