Project

General

Profile

Task #3089

Feature #2816: GPU offload / optimization for update&constraits, buffer ops and multi-gpu communication

Feature #2890: GPU Halo Exchange

relax dlb scaling limit when that would suit GPU halo exchange

Added by Mark Abraham about 1 month ago. Updated 4 days ago.

Status:
Closed
Priority:
Normal
Assignee:
Category:
mdrun
Target version:
Difficulty:
uncategorized
Close

Description

The new GPU halo exchange is limited to 1D decompositions that need only one pulse.

We should consider automatically relaxing the DLB scale limit to allow GPU DD halo communication.

Associated revisions

Revision 3f1e9ce0 (diff)
Added by Berk Hess about 1 month ago

Add option request 1D DD with 1 pulse

The option is currently controlled by an (undocumented) environment
variable GMX_DD_1D_1PULSE, but is intended to be used when requesting
GPU DD halo communication.
Note that requesting 1D DD and 1 pulse increases the lower limit
on the system size. We should consider automatically relaxing the DLB
scale limit to allow GPU DD halo communication.

Refs #3089

Change-Id: I6152c859072765c803fb99ca45ace15ae10dcfba

History

#1 Updated by Szilárd Páll 7 days ago

  • Tracker changed from Feature to Task

Do we have plans to complete this by beta2? Do we need testing before updating the heuristics?

#2 Updated by Mark Abraham 6 days ago

  • Status changed from New to Resolved

I think this is already fixed, as Berk's change made it possible to adapt

#3 Updated by Paul Bauer 4 days ago

  • Status changed from Resolved to Closed

Also available in: Atom PDF