On 19/05/15 14:04, Bless, Roland (TM) wrote:
Hi,
has anyone recently tested AQMs like Codel or PIE
at speeds of >=10Gbit/s? If so, where are the results available?
Pointers greatly appreciated...
Regards,
Roland
I know only background information on that.
1) fq_codel was written by Eric Dumazet of Google and tested at 10G.
Allegedly at 10G CoDel works fine, and fq_codel costs "2% of a single
modern CPU core". Sounds trustworthy but that doesn't give you any
details :). Cite: Jim Gettys.
Codel:
https://gettys.wordpress.com/2012/05/22/a-milestone-reached-codel-is-in-linux/
fq_codel: https://www.ietf.org/proceedings/87/slides/slides-87-aqm-6.pdf
[Whether router queue management can achieve anything for a 10g link
which also has high multiplexing and 50%+ exponential 'slow start' with
IW10... I think we didn't have an answer ready when that came up.]
2) Eric also wrote sch_fq and tested it at 10G. sch_fq is only designed
for end hosts and is not designed or implemented to reduce latency AIUI.
[I may finally understand this article
https://lwn.net/Articles/564978/
For our bufferbloat at the ISP network edge, the more important part of
the article is TSO sizing. Which does not depend on qdisc.
sch_fq just does TCP pacing (and flow queuing).
sch_fq pacing doesn't affect slow start _or_ saturating ("ack clocked")
flows, which are the causes of edge bufferbloat. It could do, but
Eric's next priority is to improve TCP congestion control.
https://lists.bufferbloat.net/pipermail/bloat/2015-April/002797.html
https://lists.bufferbloat.net/pipermail/bloat/2015-April/002801.html]
What does sch_fq actually pace? Well, "you no longer have to
slow start after idle", so it should help things like re-used HTTP
connections. But I think the implication is also for consistent flows
that don't saturate: media streaming.
https://lists.bufferbloat.net/pipermail/bloat/2015-April/002764.html]
Alan
_______________________________________________
aqm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/aqm
