Project

General

Profile

Task #3082

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

Feature #2915: GPU direct communications

move launch/synchronization points to clarify task dependencies

Added by Szilárd Páll 3 months ago. Updated 2 months ago.

Status:
New
Priority:
High
Assignee:
-
Category:
mdrun
Target version:
Difficulty:
uncategorized
Close

Description

The direct communication changes introduce new dependencies which are currently satisfied using function calls that lead to position dependent brittle code and as a side-effect suboptimal code essentially eliminating asynchrony.

Data dependencies should be reassessed after integration:
  • data movement should be initiated as soon as a producer generates data (rather than right before the communication call invocation);
  • synchronization should be async whenever possible and when synchronous, done as late as possible, close to the consumer of the data rather than bundled with the data transfer function.
Concrete cases to consider:
  • TODO

Related issues

Related to GROMACS - Feature #2890: GPU Halo ExchangeNew

History

#1 Updated by Szilárd Páll 2 months ago

#2 Updated by Szilárd Páll 2 months ago

  • Priority changed from Normal to High

Also available in: Atom PDF