[ovs-dev] OVS-DPDK design summit notes

Kevin Traynor ktraynor at redhat.com
Thu Nov 17 20:49:29 UTC 2016


Hi,

There was lots of good discussion at the recent OVS-DPDK design summit
[1] with 40+ people. Robin and I put together a summary of notes/items
that people in the room thought should be worked on, along with some
names of people who volunteered. Feel free to correct/add.

thanks,
Kevin.

[1]
https://mail.openvswitch.org/pipermail/ovs-announce/2016-October/000230.html


Usability:
- A lot of progress in the last year
- Target removal of 'experimental' tag in next OvS release
- Port naming code to be reviewed (Daniele Di Proietto)
- In tree testing to be added (Daniele Di Proietto)
- Open mailing list discussion on distributions supporting older
instruction sets (Flavio Leitner)
- Remove aborts from DPDK init (Aaron Conole)
- Implement lazy DPDK init (Aaron Conole)
- Begin work on vHost server deprecation (Ciara Loftus)
- Implement single core Tx MQ (Jan Scheurich/Mathias)
- Documentation - add information on list of supported distributions
(Bhanu Bodireddy)
- Add/Remove PMD threads without taking down the datapath (Daniele Di
Proietto)


SW performance:
- Converge guest path optimisations around virtio development group
- Investigate optimisation/removal of vHost retries for performance
(Kevin Traynor)
- Continue implementation of efficient DPDK kernel interface for
container performance optimisation (John Fastabend and Mark Gray)
- Investigate optimisation of Ofproto performance (Jan Scheurich)
- Optimise SW Conntrack perf (Mark Gray)
- Automatic load balancing on RXQs
- Increase flow limit in upcall (Daniele Di Proietto)
- Rally around VSPERF+moongen for open source comparative benchmarking
and functional testing
- Prototype implementation of EMC per subtable
- Some investigative work required around the parsing of relevant
headers as opposed to the full tuple (ie like BPF)


HW acceleration:
- Agreed that HW acceleration is a must have feature in OvS-DPDK
- Benefits for both partial and full offload to Hw
- Share use cases where conntrack is not needed (John Fastabend)
- Implement biweekly OvS-DPDK community call to enable ongoing
discussion in the community (Kevin Traynor/Robin Giller)
- Follow stepwise implementation for HW acceleration -  evolution of
functionality. For example, preclassification -> forwarding / actions ->
dataplane
- Define use cases to create architecture specification for HW
acceleration solution through DPDK (Rashid Khan to set up shared doc)
- Converge on DPDK Generic Flow API for implementation. Review
in DPDK mailing list and cross-post to OvS ML
- Build prototype to flush out usability issues, estimate performance, etc


More information about the dev mailing list