[ovs-dev] [PATCH V3] ofproto-dpif-upcall: Remove the dispatcher thread.
alexw at nicira.com
Thu Feb 27 19:44:02 UTC 2014
I'm going to post them soon, gonna be in series.
On Wed, Feb 26, 2014 at 4:26 PM, Han Zhou <zhouhan at gmail.com> wrote:
> Hi Alex,
> I had a simple test with your V3 patch, and the improvements are
> In my 32 core machine I ran 10 handler threads, comparing hping3 test
> result (simulating short-lived flows) with/without the patch: with the
> patch the host can receive 80,000+ pkts/sec without datapath lost
> (ovs-dpctl show); without the patch it can only reach 60,000+ pkt/sec,
> or datapath starts dropping packets.
> This is attributed to the removal of dispatcher thread.
> Looking forward to your V4 patch.
> Best regards,
> Han Zhou
> On Wed, 2014-02-19 at 18:44 -0800, Alex Wang wrote:
> > Thanks a lot for pointing it out, Yamamoto,
> > I failed to cover this. will work on solving it.
> > Alex Wang,
> > On Wed, Feb 19, 2014 at 6:36 PM, YAMAMOTO Takashi
> > <yamamoto at valinux.co.jp> wrote:
> > > This commit removes the 'dispatcher' thread by allowing
> > 'handler'
> > > threads to read upcalls directly from dpif. vport in dpif
> > will
> > > open netlink sockets for each handler and will use the
> > 5-tuple
> > > hash from the missed packet to choose which socket (handler)
> > to
> > > send the upcall.
> > >
> > > This patch also significantly simplifies the flow miss
> > handling
> > > code and brings slight improvement to flow setup rate.
> > doesn't this make dpif-netdev reorder packets in a flow
> > (eg. a tcp stream) if running with multiple forwarders?
> > YAMAMOTO Takashi
> > _______________________________________________
> > dev mailing list
> > dev at openvswitch.org
> > http://openvswitch.org/mailman/listinfo/dev
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the dev