[Tmrg] Measuring "burst completion times"
lachlan.andrew at gmail.com (Lachlan Andrew) Fri, 25 July 2008 23:52 UTC
From: "lachlan.andrew at gmail.com"
Date: Fri, 25 Jul 2008 16:52:43 -0700
Subject: [Tmrg] Measuring "burst completion times"
Message-ID: <aa7d2c6d0807251652u3877453dn608caea709580679@mail.gmail.com>
Greetings all, An internet draft of the test suite has been submitted <http://www.ietf.org/internet-drafts/draft-irtf-tmrg-tests-00.txt>. It has lots of questions and points for discussion (in bold in the html version <http://netlab.caltech.edu/lachlan/abstract/draft-irtf-tmrg-tests-00.html>), so feel free to start a thread. Tom Q is busy implementing it on WAN-in-Lab. We're currently having an issue with the "transfer time per flow versus file size" statistic for the "basic" scenario (Section 3.1.3). (As an aside: That is listed as an optional extra, but seems an important statistic. Should it be added to the core statistics?) It was agreed to use Tmix, which generates non-greedy connections. a) Does everyone agree that we should measure the transfer times of *bursts* (Tmix "ADUs") vs size, instead of connection times vs size? b) If so, should we use the time (and length) of a "request-response" pair, rather than an individual burst? Reasons for this are: i) That is what the user actually observes ii) Tmix currently records that (at least the Linux version does) iii) Measuring the sending time of a single burst would be difficult in a distributed network: The sending application can't know when the last byte was received -- it just sees when it was swallowed by the socket layer. This is slightly problematic as it could contain two slow-starts, but again, this is what the client actually observes. c) Tmix allows pauses between the request and response. We propose that we massage the traces by moving that pause to be after the response. That will allow us to measure the impact of TCP on the request-response pair but will keep the load roughly unchanged. d) The Tmix traces we're using have some "concurrent" connections, which are not sequences of request-response pairs. We propose that we ignore these for calculating the burst duration vs burst size statistic. Cheers, Lachlan -- Lachlan Andrew Dept of Computer Science, Caltech 1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA Ph: +1 (626) 395-8820 Fax: +1 (626) 568-3603 http://netlab.caltech.edu/lachlan
- [Tmrg] Measuring "burst completion times" Lachlan Andrew