Gene, Sounds like a great meeting. Too bad I couldn't make it. Some extremely minor corrections and additional information. Guy presented this but I did the actual testing.
UPDATES: ANS router software activity Software enhancements: RS960 buffering and queueing microcode updated
- increased number of buffers, also went from max MTU sized buffers to 2+kB chainable buffers (max FDDI will fit in two buffers with room to spare.
We are using 2kB buffers. A FDDI packet fits into 3 buffers. The advantage is that most real world packets are still ethernet MTU or less and take up less space using the new scheme. We still got the buffering up for FDDI packets and used full FDDI MTU in testing.
[ ... ]
The conditions and results were summarized on two slides:
+ Single flow Van Jacobson random early drop:
41Mbps at 384k MTU cross-country (PSC to SDSC?)
This was on our testnet. We took NY to Ann Arbor down and went by way of Texas (MCI) giving us 68 msec RTT. NY to SF is 70 msec so it is roughly cross country equivalent. Of course we couldn't step on poor unsuspecting users in the middle of the night by congesting the net, and they couldn't provide a "realistic background load" for our testing. We'd like to see a (brief) validation of results after various steps in deployment and have support from PSC and (I think) SDSC to do this.
This code (V4.20L++) is likely to be deployed in a month or so.
It doesn't have an official name and has no firm deployment plans. A month or so would be very optimistic. Some of the changes are already deployed since the Maui testing and others will deploy soon but others (RED) have no plans (yet). We'll validate progress as this stuff gets deployed and hopefully it will all get deployed (soon).
By way of comparison Maui Supercomputer center to SDSC was 31Mbps using an earlier version of code with 35 buffers. Windowed ping with the same code did 41Mbps.
MHPCC (Maui) to SDSC is a 50 msec RTT. So we went faster on a longer RTT path.
+ Four flow Van Jacobson random early drop:
42Mbps at 96kB MTU.
All the numbers are with full forwarding tables in the RS960s
We inject full routing into the testnet but don't allow packet forwarding between testnet and production net. Curtis BTW - I gave numbers to Guy that I've since revised down slightly. I changed the way I estimate link utilization in mutiple flow tests so that I will generally be more accurate and will underestimate performance if inaccurate. It's closer to 40 Mb/s for 1 flow and 41 Mb/s for 4 flows (we did hit 42 Mb/s on 8 flows) on the above tests and those results were before we tested RED. The RED code was written just before IETF and wasn't performance tested until just after. In the enthusiasm over "we have RED" I may not have conveyed that the graphs were prior to having had a chance to test RED. Could you just quietly revise the numbers down by one in the minutes. I plan to give a detailed update on the performance testing at NANOG.