[Cake] cake flenter results round 0

[Dave Taht]

Pete Heist <peteheist at gmail.com> writes:

> I have a script (called flenter) which can run flent with different parameter
> variations and produce an html report. I’m very sorry again that this doesn’t
> yet use flent’s batch feature- it was started before I knew about that.
>
> I’ll use it to produce some “rounds” of cake tests, with notes/analysis of each
> round and plans for future rounds. If there’s any feedback on anything, results
> or what to change, let me know, otherwise I’ll just take it where I’m able to.
>
> I’m calling this round 0, because these tests weren’t designed originally for
> cake at all, but it’s a vastly stripped down version of tests I was in the
> middle of doing for point-to-point WiFi. The tests need a lot of changes for the
> next round to focus more on cake and those are noted at the end.
>
> Round 0 index of all tests:
>
> http://www.drhleny.cz/bufferbloat/cake/round0/
>
> *** Notes/Analysis ***
>
> * When “over-limited” to 200mbit, pfifo bloats everything, sfq improves the UDP
> flows but bloats TCP rtt, and cake clearly holds the lowest rtt for the ping/udp
> flows as well as TCP rtt, even when compared to fq_codel:
>
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_pfifo_200mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_sfq_200mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_fq_codel_200mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_cake_200mbit/index.html
>
> * At 950 mbit rate limiting, fq_codel holds latency of the ping and udp flows a
> bit lower than cake, whereas cake holds tcp rtt a bit lower. sfq does pretty
> well actually and pfifo is, well, pfifo:
>
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_pfifo_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_sfq_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/eg_csrt_rrulbe_eg_cake_950mbit/index.html
>
> * This is a little surprising, at 950mbit rate limiting with nflows 8/8, 16/16
> and 32/32, fq_codel seems to outperform cake both in TCP rtt and latency of the
> UDP flows. Is cake’s “ethernet” keyword is really crucial here, or is there a
> difference between Cake and fq_codel at these high bandwidths? I’ll add it in my
> later tests. Also I’m losing the queue with 64/64 flows so will lower the
> bandwidth limit.
>
> http://www.drhleny.cz/bufferbloat/cake/round0/nflows_32_32_eg_pfifo_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/nflows_32_32_eg_sfq_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/nflows_32_32_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/nflows_32_32_eg_cake_950mbit/index.html
>
> * Obviously cake (with diffserv3 and not besteffort), roundly defeats everything
> else in the torrent test because of the de-prioritization of the background
> flows:

I think you can safely drop pfifo from future tests.

I'd rather like combined charts so it is possible to eyeball these
differences directly. Just between fq_codel and cake. Or, is there a
tarball available to browse this stuff?

Another nice thing to try capturing is queue depth/loss/marks/etc, which
is --test-parameter qdisc_stats_hosts=X,y,z and qdisc_stats_interfaces=

There's also capturing the tcp statistics on the server that is
possible.

>
> http://www.drhleny.cz/bufferbloat/cake/round0/tor_rrultor_eg_pfifo_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/tor_rrultor_eg_sfq_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/tor_rrultor_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/tor_rrultor_eg_cakebe_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/tor_rrultor_eg_cake_950mbit/index.html
>
> * The benefit to lowering cake’s rtt parameter in an Ethernet environment can be
> seen on the TCP rtt:
>
> http://www.drhleny.cz/bufferbloat/cake/round0/cake_rtt_100ms_rrulbe_eg_cake_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/cake_rtt_60ms_rrulbe_eg_cake_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/cake_rtt_20ms_rrulbe_eg_cake_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/cake_rtt_10ms_rrulbe_eg_cake_950mbit/index.html
>
> * I’m a little surprised that fq_codel holds UDP flow latency a little lower at
> "target 1ms interval 10ms" than cake’s "rtt 10ms”. It almost seems like a trend
> that Cake outperforms at lower bandwidths and fq_codel at higher bandwidths.

Since fq_codel supports superpackets and cake peels them, we have a cpu
and latency hit that originates from that. Also the codel derived
algorithm in cake differs quite significantly from mainline codel, and
my principal gripe about it has been that it has not been extensively
tested against higher delays.

>
> http://www.drhleny.cz/bufferbloat/cake/round0/fq_codel_ti_t5ms_i100ms_rrulbe_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/fq_codel_ti_t3ms_i60ms_rrulbe_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/fq_codel_ti_t2ms_i20ms_rrulbe_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/fq_codel_ti_t1ms_i10ms_rrulbe_eg_fq_codel_950mbit/index.html
>
> * ECN helps marginally with UDP flow rtt, but I’ve never seen ECN do very much.
> When does it help the most?
>
> http://www.drhleny.cz/bufferbloat/cake/round0/ecn_off_rrulbe_eg_cake_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/ecn_on_rrulbe_eg_cake_950mbit/index.html
>
> * Cake’s “ethernet” parameter helps a bit, I’ll add it to all other rate limited
> tests in my next round:
>
> http://www.drhleny.cz/bufferbloat/cake/round0/cake_overhead_rrulbe_eg_cake_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/apu2-eth/cake_overhead_rrulbe_eg_cakeeth_950mbit/index.html
>
> * Cake’s host isolation clearly works, but I’m a little surprised that
> “srchost/dsthost” is more fair on a host level than “dual-srchost/dual-dsthost”
> (I usually find it easiest to just scroll to the bottom of the page and look at
> the numbers):
>
> http://www.drhleny.cz/bufferbloat/cake/round0/hostiso_eg_fq_codel_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/hostiso_eg_cake_src_cake_dst_950mbit/index.html
> http://www.drhleny.cz/bufferbloat/cake/round0/hostiso_eg_cake_dsrc_cake_ddst_950mbit/index.html
>
> * Anyone see anything in my “Flow Isolation Mix” tests? Those are a little hard
> to read. :) They used to be combined with a VoIP test but I don’t have a d-itg
> setup now.

I look forward to you adding OWD irtt based tests.

> *** Round 1 Plans ***
>
> - Update cake to latest (will do this with every round)
> - Remove all “full-duplex limiting” (egress and ingress) tests as I don’t see
> the use here
> - Add cake’s “ethernet” keyword to all rate limited tests
> - Lower standard rate limit to 900mbit ensure no queue loss (particularly for
> nflows tests)
> - Take standard rrulbe tests to even lower bandwidths: 1mbit, 10mbit, 50mbit,
> 100mbit
> - Add bql tests (no rate limiting)
> - Add 100us, 1ms, 2ms, 3ms, 5ms, 8ms to Cake RTT tests
> - Add nflows tests at lower bandwidths
> - Fix UDP flood tests (no suitable iperf binary found)
> - Remove or improve flow isolation tests
> - Add ethtool output to host info
>
> *** Plans for Future Rounds ***
>
> - Add flow isolation tests with rtt variation (to look again at problem I
> reported in an earlier thread)
> - Use netem to make a spread of rtts and bandwidths (maybe the most useful of
> all?)

Yes.

> - Add ack filtering tests
> - Add VoIP tests (I hope to do this with irtt)
> - Test BBR
> - Use qemu to test other archs (I may never get to this, honestly)
>
>
> _______________________________________________
> Cake mailing list
> Cake at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cake