Did you see my reply to this question on the Mininet list? POX's task scheduler can lead to a batching effect. It's not optimized for low latency reactive applications. POX's current development branch (dart) has an option (which may become the default) which generally does much better for such uses. It's activated with the --unthreaded-sh commandline option, as indicated on my reply on the Mininet list. I suggested that you try it and see if the performance was more like you were expecting. Have you tried this?
-- Murphy On Jan 16, 2014, at 2:33 PM, Xiaoye Sun <[email protected]> wrote: > Hi all, > > I created a chain network topology with 3 ovs kernel switches. Each switch at > the end of the chain connects to a host. The topology is like this: > h1--s1--s2--s3--h2 > > These switches are connect to a pox_carp controller running on the local > machine. > The pox controller runs the openflow.discovery and a modified > forwarding.l2_multi applications (not using barrier message and not > installing entries for the reversed path). the l2_multi application will > install exact matching rules, specifying mac address, ip address and TCP port. > > The experiment I did is described here. > h1 keeps sending TCP syn packet to h2 with different TCP source port number > at a specific rate. > After h2 receives the syn packet, it will immediately replies a syn-ack > packet to h1. > Each syn and syn-ack pair will be consider as different flows by the switches > since controller only install exact matching rules (they have different tcp > ports). > Thus, every syn and syn-ack packet will be sent to the controller by the > switches. > > In the experiment, I change the sending rate of the syn packet and compute > the RTT of each syn and syn-ack pair. However, I found that when the packet > rate is high, for example, 200 syn pkt per sec, the RTTs usually are much > less than the RTTs of a lower sending rate, for example 50 pkt per sec. > > This result is strange since when the sending rate is high, the controller > and ovs should have more workload. The RTT should not be less than the RTT in > a network having less workload per second... > > My question is that does POX implements any mechanism (like batching...) that > can process the packet-ins faster when they are coming at a higher rate? > > Xiaoye > > -- > Xiaoye (Steven) Sun, Ph.D. Student > Department of Electrical and Computer Engineering (ECE) & Department of > Computer Science (CS) > George R. Brown School of Engineering > Rice University, Houston, Texas, USA
