You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 4 Next »

Performance Working Group at Joint Techs at Fairbanks, Alaska

12-July-2011

Welcome  -- _Carla Hunt

Internet2 Update -- Aaron Brown and Jeff Boote

  • Aaron Brown noted that Internet2 had recently released a minor release of the perfSONAR-PS Toolkit, version 3.2.0.1,
  • This version allows folks to download the RPMs from R+E mirrors.
  • Aaron asked folks to email him if they were interested in hosting one of these mirrors.
  • Aaron noted that they were working on an upcoming perfSONAR-PS Toolkit release, version 3.2.1,
  • This version should include bugfixes and some updated packages.
  • The release candidate should be out in the next week or so.
  • Aaron commented that Internet2 had been working with the ESnet developers on the OSCARS version 6 release, and that it should be available in the next few weeks.
  • On the DYNES front, he noted that Internet2 has been working with the DYNES group 'A' participants, and that they were looking to do the first set of DYNES installations within the next few weeks.
  • Aaron commented that Internet2 had been working on a circuit monitoring infrastructure as part of the DYNES work, and that he'd given a talk on it earlier.
  • Jeff commented that Internet2 was likely going to be working on an NDT project with Google,
  • Internet2 might be working with Verizon on a continuation of the work they'd done earlier in developing a measurement infrastructure for Verizon.

Carla Hunt talk:

Eric Boyd asked if the motivation for deploying the fiber database was to meet BTOP requirements. Carla said that it was partially motivated to support the BTOP grant, but the bigger interest was to be able to track the fiber, and for planning purposes.

Jim Warner asked if this software was being used for campuses, or just wide area. Carla said that the primary use is for the wide area. The idea was to make it easy for folks whose job was in the campus could easily access the data

Phil DeMar talk:

Eric Boyd asked if the 'hub' definition for FermiLab consisted of ESnet's FermiLab box, as well as FermiLab's boxes. Phil said that the concept was to have a user-centric view. They don't care who runs the perfSONAR box, just that they can obtain measurements applicable to FermiLab. Eric asked if this could be extended to a three-level system (e.g. a hub including regional networks and all universities). Phil thought that they would go from at the university level. The basic idea was just to keep users from having to select hostnames.

Carla Hunt asked if the 'hub' metadata was in the lookup service, or was local to the Web UI. Phil wasn't sure, but figured it was local to the Web UI.

Jim Warner asked if the cached data was kept around longer than the original perfSONAR services would keep it. A number of folks said that perfSONAR services tend to keep data forever.

Martin Swany:

No questions

Matt Zekauskas talk:

Matt said that Internet2 has been monitoring for a long time. They've been using firewall filters on the Junipers that are configured to soley measure counts, not block anything. They've seen 5G of v6 traffic on the backbone. The hardware to do proper flow monitoring is expensive though, so they've implemented a workaround. They duplicate all IPv6 packets to a local PC (<1G per site), and do flow generation from those packets. This came online just before IPv6 day, and has been monitoring ever since. They've been working on a backend to save and make avialable all these statistics. He noted that the data is available if researchers are interested in it, and that it is anonymized (80-bit zeroed, he thinks).

Brian Tierney asked if the backbone bwctl and owamp measurement traffic was using IPv6. Matt said that it was, but that the meaurements being shown was connector traffic, and not backbone traffic.

Jim Warner talk:

Jim Warner noted that Jim Getty has been promoting the "buffer bloat" issue. Jim Warner felt that the mailing lists and websites like fasterdata.es.net were telling people that bigger buffers aren't a problem. Carla Hunt asked what the recommendation that the mailing lists and websites are suggesting? Jim said that fasterdata.es.net recommended an outbound queue length of 1000 packets for 1G, and 10000 packets for 10G. He felt that was far too much buffer at home. Most of the buffers filling up are in the DSL boxes which are unchangeable. He noted that the outbound traffic can be shaped so that DSL queueing doesn't happen. Martin Swany felt that this was a TCP problem. He felt it was misguided to screw with sending so that TCP doesn't get confused. Jon said that TCP worked that way, and because of that, we can't ignore it. Jon thought the problem with buffering was that it was dependent on the link speed the router was attached to. One megabyte of buffering was nothing for 1G ethernet, but on DSL, it's forever. Jim noted that the effect of all this "hidden buffering" was a large percentage of a second (1/3 and higher). Prasad noted that it can be even worse on a noisy channels like wireless where retransmits happen after 60ms so you can get even higher blocking times. Jim said that wireless was a problem for lots of reasons, including that valid wireless speeds varied from 2Mbps to 100Mbps.

Summer 2011 ESCC/Internet2 Joint Techs
Hosted by the University of Alaska-Fairbankshttp://events.internet2.edu/2011/jt-uaf

  • Work is

Presentations:

pe

See the Presentation Slides linked from here

http://events.internet2.edu/2011/jt-clemson/agenda.cfm?go=session&id=10001551&event=1150

  • No labels