September 27 from 4:00 – 5:30 pm (CDT)

SPEAKERS:
Matt Zekauskas, Internet2
Jeff Boote, Internet2

ATTENDEES
Laurie Kirchmeier, CharlesAntanew, Borje Josefsson, Stanislav Shalunov, Athanassios Liakopoulos, Eric Boyd, Joe St Sauver, John Moore, Stephen Bavington, Chris Heermann, Don Choi, Jim Ferguson, David Lapsley, George Brett, Yang Zia, Tanya Brethour, Jeff Boote, Leobino Sampaio, Jose A.S. Monteiro, Brian Tierney, Matt Zekauskas, Kevin Walsh, Terry Gray, Dave Pokorney, Rich Carlson, Clark Gaylord, Russ Hobby, Connie Logg, Chu-Sing Yang, Wayn Bullock, James Deaton, and Susan Evett.

Matt Zekauskas introduced himself and presented the agenda:

• BWCTL and OWAMP update
• Open Q&A on testing to Abilene
• First evidence of queuing on Abilene
• Research opportunities on the Abilene Observatory (AO)
• Open discussion

• 3-party communication – a user can , from their laptop, ask the Chicago and Los Angeles Abilene nodes to run BWCTL between them, and it will.
• Authentication options, that now allow a client to authenticate differently to each side of a test, which allows campus users to test between their GigaPoP and Abilene, for example.
• No need for a local server (BWCTL will detect a local daemon but, if one isn’t running, the test will spawn a tester point for the period of the test).
• Port ranges for bwctld peer connections that can now be specific (to allow BWCTL to co-exists with firewalls and NATs).

Boote gave an overview of OWAMP and described what is new in the latest version. All tests go from a client to a server and the client initiates. The client contacts the OWAMP daemon; there is a resource broker where the data is limited (by maximum bits per second, for example) and existing resources are reported (how many outstanding tests are taking place, how much resource tied up, etc.).

Connie Logg asked if the broker/ policy was something she’d need to write for herself or was there something she could just use. Boote reported that we have the broker/policy for the Abilene Measurement Infrastructure (AMI) and it could just be plugged in. Eric Boyd said that we want to make it possible for all these tests to be run on the same box (cost issues) because then you wouldn’t have to worry about BWCTL tests being run at the same time as OWAMP tests.

Open Q&A on Testing to Abilene
Matt Zekauskas gave an overview of the tools available via E2Epi. He reported on how the AMI tests were being run, showed how to get to NDT for on-demand testing. Zekauskas reported that access to OWAMP was more widely given, because the tool uses fewer resources. Zekauskas showed the form users need to send in to request access to tests into the middle of Abilene and explained how the process worked. For the most part, UDP tests are not allowed into the middle but TCP tests are permitted. When asked how users get the results back, Boote reported that the results go to both nodes being tested.

Boyd reported that someone from UMich reported today that they were having problems reaching Indiana so they wanted run tests into the middle to partially decompose the path. The value of this type of testing is linked to the number of locations across a path to which a user can run a test.

Q: If we have BWCTL on node b, and I want it to run to node c, can I go to node a and ask it to run a test between nodes b and c?
A: Yes. You can authenticate at one location or at both test locations (which is what would happen if you were crossing domains).

Q: If I’m writing code, and I run a test, is there some sort of security measure?
A: This is an in between measure – we didn’t want to create an API when Shibboleth is nearly completed and it will do the work for us.

Q: Why aren’t other people doing throughput tests?
A: BWCTL has more overhead – you need to have an AES key and a system administrator running things. Zekauskas noted that, with OWAMP, the user needs synchronized clocks and, with BWCTL, the user needs security issues clarified.

First Evidence of Queuing on Abilene
Matt Zekauskas noted that he’d seen his first evidence of queuing on the Abilene backbone (http://people.internet2.edu/%7Ematt/delay/). It was a 15 minute period on a Tuesday in August Caltech to CERN performing 10GE throughput experiment. The practical limit is closer to 7.5 Gbps – exactly what was being tested is unknown – Denver going to Houston experienced delay but it isn’t the same on the return! Flat everywhere except for the nodes from Denver to KC outbound and everything from KC to Sunnyvale on the return. It was only paths that traverse these nodes that show an additional delay. This intrigued them, so they looked at the SNMP data for the same period. It looked like a TCP ramp up – 9.5 Mbps – bumping up against 10 Gbps link. Abilene has both 5-minute and 1-minute SNMP graphs.

It appears that the testers exceeded 10 Gbps for a brief period. When there is a constant delay, it means the link is maxed out – the fourth and fifth events were 10-30 seconds long, so this is not a blip on the screen that could be an error. One of the reasons these tests were going on were to ensure they were ready for tests next year that require them to send data at 10 Gbps – to see what they needed to do. Zekauskas doesn’t have an absolute explanation for this but found it interesting.

Q: Could these variations be caused by routers? Juniper routers sometimes have these blips.
A: I don’t know enough to say – the way this particular router is, I don’t think so.

Q: Could these be NTP anomalies?
A: I don’t think so – you’d see mirror images in that case.

Q: Do you have any Netflow data that told you anything?
A: No. But, actually, we do have two scheduled tests – the anomalies might be being caused by them. Jeff Boote disagreed because he feels that the test is run 1x/hr/host. Stanislav Shalunov argued that it could be because the test is run for two tools on three different protocols and it could cause nearly that error rate.

Research Opportunities on the Abilene Observatory (AO)
There are two components to the AO – things Internet2 puts in and things users can get out. Having a piece of equipment co-located at an Abilene node might be valuable for some projects. AMP and PlanetLab are currently co-located and the NLANR network analysis project is collecting data at the nodes. Other projects include Netflow-like analyses. Zekausakas showed the links to data views for the data collected by the AMI – such as the 10 worst performing TCP throughput (IPv4) nodes.

New AO features include enhancements to data storage and access, changing the router buffer size based on test results, allowing a few high-rate SNMP tests, and adding a new measurement server. Outreach to understand Computer Science community needs is starting up and will include visiting folks to see if we can do anything for them.