Re: [ixp1200] Large packets on the IXP

Hi Abhijeet, Thanks for your interest on Genesis/NetBind and feedback. Here are some quick answers to your questions:
Right Thanks. So figure 12 reports 900,000pps , each packet = 64 bytes, across all 4 ports, which means a bandwidth of 115Mbps across EACH port.
Is this interpretation correct?
Not exactly. We receive packets from 4 ports and forward them to the same output port. So these numbers are from a four-to-one configuration. Also the receiving rate is not ~100Mbps but mucher larger. To measure the maximum possible throughput we assume that the rfifo always has a packet received. The size of this packet is 64 bytes. So our code repeatedly gets a minimum size packet from the RFIFO and forwards the packet to the same output port. Our measured maximum throughput is in accordance with the maximum throughput measured by our Princeton colleagues. However in their measurements they use sixteen contexts for fowarding as opposed to four in our case.
Also, how do you saturate the IXP's 100 Mbps links through external sources? Do you use packet generating hardware or just plain PC's with
One way to do it is to use IXPs as packet generators, if you have many boxes to play with. So the packet generating code runs in the microengines of an IXP1200 system.
some kernel/user level packet generators? I am finding it preety hard to get high rate sources on PC's without missing packet send deadlines on the senders.
Hmm.., maybe I could use multiple sources per port !!
Also, how did you measure the 225 cycles per packet processing time? In the paper you mention you measure it through some program on the StrongArm. So basically do you pump a large number of packets, so that your IP code runs through a large number of iterations and then take coarse grained measurements of packet count from the StrongArm and then divide it by the total time taken to forward packets?
correct. Hope this helps, best, Michael
participants (1)
-
Michael E. Kounavis