[Starlink] Fwd: Pie not fully implementable on tofino p4

Dave Taht dave.taht at gmail.com
Wed Jun 23 16:02:38 EDT 2021


One of the things that has long bugged me about the p4 spec was the
lack of an invsqrt which is a single assembly language instruction or
about 3k gates, for a newton approximation.

Anyway, I wish I had direct comparison data for pie vs codel on p4,
two good papers below.

---------- Forwarded message ---------
From: Ralf Kundel <ralf.kundel at kom.tu-darmstadt.de>
Date: Wed, Apr 21, 2021 at 9:07 AM
Subject: Re: Pie not fully implementable on tofino p4
To: Dave Taht <dave.taht at gmail.com>, bloat
<bloat at lists.bufferbloat.net>, <codel at lists.bufferbloat.net>


Hi,

regarding the question: how timestamps come from ingress to egress pipe?

Well this depends (on the P4 abstraction) on the P4 version you use. As
far as I remember (it is a long time ago) in P4_14 metadata (which can
store the timestamp) are handed over implicitly from ingress to egress.
In P4_16 you have do do this explicitly in the Deparser of Ingress and
Parser of Egress: Simply said you just write the timestamp before the
Ethernet header. You can see this here in line 64 to 80:
https://github.com/ralfkundel/p4-codel/blob/master/targets/tofino/srcP4/l1switchCodel.p4


best,
Ralf

Am 21.04.21 um 04:42 schrieb Dave Taht:
> https://www.comsys.rwth-aachen.de/fileadmin/papers/2021/2021-kunze-aqm-tofino-p4.pdf
>
> For a codel comparison, see:
>
> https://arxiv.org/pdf/2010.04528v1.pdf
>
> Where the authors show that fq_codel is feasible on the PISA platform...
>
> but a direct comparison between the two papers is not possible. What I
> don't understand about the above paper is how codel timestamps are
> passed from ingress to egress.
>
--
________________________________________________

Ralf Kundel, M.Sc.

Technical University of Darmstadt
Dept. of Electrical Engineering and Information Technology Multimedia
Communications Lab (KOM)
Adaptive Communication Systems Group (ACS)

Postal address: TU Darmstadt, FB 18, KOM, Rundeturmstr. 10, 64283
Darmstadt, Germany
Location: S3|20 - 216
Tel: +49 6151 16-20855
Email: Ralf.Kundel at kom.tu-darmstadt.de
Web: http://www.kom.tu-darmstadt.de
___________________________________________________________



-- 
Latest Podcast:
https://www.linkedin.com/feed/update/urn:li:activity:6791014284936785920/

Dave Täht CTO, TekLibre, LLC



More information about the Starlink mailing list