Project

General

Profile

Bug #2177

tip4p_continue fails with -ntomp 8 -nt 8

Added by Roland Schulz over 2 years ago. Updated over 2 years ago.

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

Description

In the past we always supported running test up to 8 (MPI/OpenMP) of each and all other tests still pass. tip4p_continue was broken by 819dc6c7f8be.
It works by adding "-dd 4 2 1". Why does the auto domain decomposition pick 8 1 1 for such a small input?

Associated revisions

Revision 4af4dbb7 (diff)
Added by Berk Hess over 2 years ago

Make DD setup obey PME restrictions

When running with many MPI ranks relative to the system size
combined with OpenMP, the domain decomposition could choose a setup
with to few PME grid lines per rank along x, which resulted in
a fatal error. Now the PME grid restrictions are checked during
the domain decomposition setup.

Fixes #2177.

Change-Id: I2f3ded51d9f447a0571f78e7d6ead4d262f599d5

History

#1 Updated by Gerrit Code Review Bot over 2 years ago

Gerrit received a related patchset '1' for Issue #2177.
Uploader: Berk Hess ()
Change-Id: gromacs~master~I2f3ded51d9f447a0571f78e7d6ead4d262f599d5
Gerrit URL: https://gerrit.gromacs.org/6633

#2 Updated by Berk Hess over 2 years ago

  • Category set to mdrun
  • Status changed from New to Fix uploaded
  • Assignee set to Berk Hess
  • Target version set to 2018
  • Affected version changed from git master to 2016

Nx1x1 minimizes the PME communication, which apparently is more costly than the reduced PP communication for other setups here.
This setup now fails, because it uses OpenMP and it could not with the group scheme. I uploaded a fix for this.

#3 Updated by Berk Hess over 2 years ago

  • Status changed from Fix uploaded to Resolved

#4 Updated by Berk Hess over 2 years ago

  • Status changed from Resolved to Closed

Also available in: Atom PDF