Attached are the minutes from two weeks ago, which included lively discussions 
on some pitfalls with the current kernel verbs approach and the potential for 
kfabric as a mechanism for stabilizing the API w.r.t. the consumer.  The 
discussion included:

  *   Continuing to develop kfabric as a match for LNET
  *   Exploring the level of abstraction of kfabric - do we need an easy button?
  *   CPU/Affinity - want to be able to hand hints down to the provider in 
order to maintain CPU affinity all the way down to the provider.

All agreed that a workshop session highlighting some of the challenges that 
Lustre faces using kernel verbs (e.g. memory registration complexity) and how 
kfabric could be a potential solution.  Unfortunately, Doug is unable to attend 
the WS this year.

I'm not precisely sure where this discussion needs to go next, but it feels 
like there is more to be done here.

Regards,
-Paul

Advanced Technology Group
Cray, Inc.
Office  - (503) 620 - 8757
Mobile - (503) 703 - 5382

Attachment: OFWG_DSDA_minutes_2017-02-07.docx
Description: OFWG_DSDA_minutes_2017-02-07.docx

_______________________________________________
ofiwg mailing list
[email protected]
http://lists.openfabrics.org/mailman/listinfo/ofiwg

Reply via email to