Bug 1628669 - [NFV] Add OVS 2.9 with round-robin based rxq-pmd assignment into OSP 10
Summary: [NFV] Add OVS 2.9 with round-robin based rxq-pmd assignment into OSP 10
Keywords:
Status: CLOSED CURRENTRELEASE
Alias: None
Product: Red Hat OpenStack
Classification: Red Hat
Component: openvswitch
Version: 10.0 (Newton)
Hardware: Unspecified
OS: Unspecified
high
high
Target Milestone: z11
: 10.0 (Newton)
Assignee: Kevin Traynor
QA Contact: Yariv
URL:
Whiteboard:
Depends On: 1616001
Blocks:
TreeView+ depends on / blocked
 
Reported: 2018-09-13 16:58 UTC by atelang
Modified: 2019-06-07 18:02 UTC (History)
21 users (show)

Fixed In Version: openvswitch-2.9.0-56.el7fdp.1
Doc Type: Enhancement
Doc Text:
With this update, to specify round-robin assignment of Rxqs to PMDs according to the port/queue number in OVS 2.9, use the `pmd-rxq-assign=roundrobin` configuration. Round-robin assignment of Rxqs to PMDs was the default prior to OVS 2.9 and can be preferable for systems with volatile traffic and configuration. OVS 2.9 also includes the recorded processing cycles from assigning Rxqs to PMDs.
Clone Of:
Environment:
Last Closed: 2019-06-07 18:02:04 UTC
Target Upstream Version:
Embargoed:


Attachments (Terms of Use)


Links
System ID Private Priority Status Summary Last Updated
Red Hat Bugzilla 1662419 0 high CLOSED [NFV] Add OVS 2.9 with round-robin based rxq-pmd assignment into OSP 13 2022-03-13 16:38:15 UTC

Description atelang 2018-09-13 16:58:50 UTC
Description of problem:
(a) what OSP-10 change is the root cause of the regression:     
Prior to OVS 2.9 automatic assignment of Rxqs to PMDs
    (i.e. CPUs) was done by assigning Rxqs in an ascending
    port/queue order, round robined across the available
    PMDs.
    
    That was changed in OVS 2.9 to order the Rxqs by the
    measured processing cycles for each, in order to try
    and keep the busiest Rxqs on different PMDs.
    
    For the most part the new scheme should be better, but
    there could be situations where a user prefers a
    port/queue scheme because Rxqs from a single port are
    more likely to be spread across multiple cores, and/or
    traffic is very bursty/unpredictable.
    
    Add the option to have a port/queue based assignment.

(b) when was that change introduced: In OVS2.9 
(c) the impact of that regression on customers: Possible performance degradation from OVS2.6 to OVS2.9 
 
Related OVS BZ https://bugzilla.redhat.com/show_bug.cgi?id=1616001

Comment 1 Daniel Alvarez Sanchez 2018-09-18 12:17:58 UTC
Link to upstream patch: https://patchwork.ozlabs.org/patch/964332/
Will it be possible to get the backport to 2.9? Any ETA?
Thanks!

Comment 2 Kevin Traynor 2018-09-19 10:11:33 UTC
(In reply to Daniel Alvarez Sanchez from comment #1)
> Link to upstream patch: https://patchwork.ozlabs.org/patch/964332/
> Will it be possible to get the backport to 2.9? Any ETA?
> Thanks!

Yes, I will make backports for 2.9 (and 2.10) and run through our CI. If no issues, then expect by EOD tomorrow.

Comment 3 Daniel Alvarez Sanchez 2018-09-19 12:57:40 UTC
++ Thanks a lot Kevin for your swift response. Great job!

Comment 4 Kevin Traynor 2018-09-20 15:11:00 UTC
The option to have a simple round robin assignment is backported to OVS 2.9 in 'openvswitch-2.9.0-56.el7fdp.1'. Please note the '.1'.

It is available in brew here:
https://brewweb.engineering.redhat.com/brew/buildinfo?buildID=770069

The option is default *off* and can be enabled with: 
ovs-vsctl set Open_vSwitch . other_config:pmd-rxq-assign=roundrobin

Comment 5 Daniel Alvarez Sanchez 2018-09-25 12:22:15 UTC
I guess we need a separate bugzilla then to enable this option? Either from neutron or from OSPD?


Note You need to log in before you can comment on or make changes to this bug.