Project

General

Profile

Documentation #1611

Document network dependent rdma thresholds, above which benefits of the zero copy API can be gained

Added by Nitin Bhat 4 months ago. Updated 9 days ago.

Status:
Implemented
Priority:
High
Assignee:
Category:
Documentation
Target version:
Start date:
06/15/2017
Due date:
% Done:

0%


Description

It's unclear to the user as to when the Zero Copy API should be used. So, it'll be good to document the buffer sizes for different networks to help the user choose between regular sending and zero copy sending appropriately.

History

#1 Updated by Nitin Bhat 4 months ago

  • Target version set to 6.8.1
  • Status changed from New to In Progress
  • Category set to Documentation

Performance Evaluation before 6.8.0 release: https://docs.google.com/spreadsheets/d/1dL2AYO-kKZYGeM1PLDRX3k9z_uL95eDI7aU9qQjBZmo/edit?usp=sharing\

Current numbers for GNI and Verbs aren't great (around 2 MB mark). This will improve with upcoming rdma features like recv-side, registration api, dynamic decision to use regular send for small arrays etc.

For now, documentation for 6.8.0 mentions that benefits from the zero copy api are mostly because of reduced memory footprint.
Gerrit: https://charm.cs.illinois.edu/gerrit/#/c/2734/

#2 Updated by Nitin Bhat 23 days ago

Since the memory registration API and registration cache optimization hasn't been done for the Rdma Nocopy Entry method API and the performance numbers are still bad for gni and verbs, I think this ticket should be moved to a later version.

#3 Updated by Nitin Bhat 14 days ago

  • Status changed from In Progress to Implemented

#4 Updated by Phil Miller 9 days ago

  • Target version changed from 6.8.1 to 6.9.0

Not going to hold the 6.8.1 release for this.

Also available in: Atom PDF