On Wed, Oct 1, 2014 at 5:53 PM, Justin Pettit <jpet...@nicira.com> wrote:
> On Wed, Sep 17, 2014 at 4:12 PM, Madhu Challa <cha...@noironetworks.com> > wrote: > >> >> I got a chance to try out your code with the example flows and measure >> raw packet throughput with and without connection tracking. The code works >> and the performance numbers look pretty good. >> > > Sorry that I missed this message earlier. Thank you very much for trying > this out and getting some performance numbers! > You are very welcome. > > >> I ran two sets of tests both with 4 iptable rules in chain input and 8 in >> chain forward, all default rules coming from a docker installation. The >> setup is a ovs connected to two packet generators via two ports. I am using >> dpdk-pktgen. >> > > I'm not sure I understand what you mean by using iptables rules and OVS. > I would have expected you just to be using OVS flows instead of iptables, > but maybe I misunderstand. > I had some ip tables rules that would need conntrack and I was not sure if it would affect the performance numbers. I had rechecked and looks like it does not. sorry about the confusion. > > >> In the first test I am sending random tcp 64 byte packets and the >> throughput difference is about 20% (5484/6819 Mbps) >> > > Are these two rates representing no conntrack (6819) and with conntrack > (5484)? > yes. > > >> In the second test I let a connection get established and then hijack the >> connection via pktgen and send a spoofed packet from one end to the other. >> Here I see a throughput difference of 15%. >> > > Why do you think the performance is better when hijacking the connection? > Is it that pktgen is detecting a slower rate when the caches are heating up > and never adjusts itself up? > I think its because the retransmitted data packets were slightly bigger in size. Thanks. > > Thanks again! > > --Justin > > > _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
