Re: [tsvwg] Requesting TSVWG adoption of SCE draft-morton-tsvwg-sce

[Jonathan Morton <[e-mail]>]

> On 19 Nov, 2019, at 6:30 pm, Ingemar Johansson S <ingemar.s.johansson=[e-mail]> wrote:
> 
> Yes. I know the paper discussed CoDel, what I wanted to point out however that CoDel has its own share of issues related to link underutilization at higher RTTs, meaning that the 5ms threshold is not good enough in all cases. My interpretation is that for classic traffic you need to set the thresholds differently depending on what RTT you expect your e2e traffic to experience.

My reading of Codel literature is that 'interval' is intended to be set at an estimate of native path RTT, accurate to within about an order of magnitude, and that 'target' is to be set at nominally 1/20th of that.  The default settings of 100ms 'interval' and 5ms 'target' are appropriate for typical Internet paths, which I believe some literature quotes at 80ms average.  I say this as someone who has worked extensively with Codel implementations.

If you expect to handle a lot of traffic over GEO satellite paths, it would be sensible to increase these values about tenfold: 1000ms interval, 50ms target.  Some locations, especially in the southeastern corner of the world, may prefer settings optimised for intercontinental paths: 300ms interval, 15ms target.  On slow links it may be necessary to increase target to avoid capturing serialisation delay, without changing interval.

When congestion control exhibits a deep sawtooth (as it traditionally does), there's always a tradeoff between peak latency and overall utilisation; Codel broadly sacrifices the latter to optimise the former, but not recklessly IMHO.  A major goal of several recent proposals, including BBR, SCE, and L4S, is to remove that deep sawtooth, and thus allow operating closer to the ideal of maximum throughput with minimum latency.

 - Jonathan Morton