[Bloat] [Cake] 450 flow tests

Dave Taht dave.taht at gmail.com
Wed Apr 22 22:08:42 EDT 2015


I did a little post processing on that data set (with sch_fq hitting
it's packet limit),
attached is the result and script that generated it, as well as the test script.

Based on this extremely limited analysis, I might argue that sch_fq does not,
actually, scale well to millions of flows at gigE, and that there are
more things
we could do to reduce host buffering than we have done already, in anything,
with tons of local flows extant.

But my kernel and test rig are suspect, and I would suggest that someone
with a testbed (and not on vacation) attempt a gigE test with and without
offloads, with the qdisc changing on the host, only, on a modern kernel.

Each one of these is 50 flows up, 1 flow down, just tracking the up.

Qdisc sch_fq pie fq_codel cake cake_flowblind
1 154.011054 139.730713 184.544654 179.255445 172.34438
2 138.095866 109.192431 164.956621 163.068599 159.865739
3 131.337959 121.666566 85.271696 63.340337 127.387201
4 128.151294 103.581081 155.710963 168.49141 156.477068
5 127.481834 104.298855 155.354039 168.913933 149.34434
6 128.053238 116.933986 153.856399 128.460848 148.357292
7 131.207182 141.397305 157.879694 141.654782 138.197248
8 137.152774 105.374206 165.909706 126.396167 137.494603
9 151.162839 122.407064 179.41467 173.291788 177.100583


Total 1226.65404 1064.582207 1402.898442 1312.873309 1366.568454

On Wed, Apr 22, 2015 at 11:52 AM, Jonathan Morton <chromatix99 at gmail.com> wrote:
>
>> On 22 Apr, 2015, at 18:59, Dave Taht <dave.taht at gmail.com> wrote:
>>
>> * Cake totally controls things
>
> Yay!  And there really aren’t very many collisions there, so the FQ logic is still mostly working as designed.
>
>> * As does cake flowblind, but less so
>
> That is surprising, actually.  With shaping off, all packets having the same DSCP and with FQ disabled, it should behave similarly to ns4_codel (as I used codel4.h as a basis for codel5.h).  That the three “way” stats counters remain at zero confirms that the FQ logic is being completely bypassed (since if the hash was fixed at zero and the FQ logic still ran, you’d see at least some misses and indirect hits as well as a lot of collisions, since the set-associative logic would still search the first 8 queues).
>
> But I didn’t notice an ns4_codel test run in your post.  A direct comparison might be apt, and then we can look at the differences between ns4_codel and the other codels in isolation.
>
>  - Jonathan Morton
>



-- 
Dave Täht
Open Networking needs **Open Source Hardware**

https://plus.google.com/u/0/+EricRaymond/posts/JqxCe2pFr67
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 450flowgoodput.ods
Type: application/vnd.oasis.opendocument.spreadsheet
Size: 45832 bytes
Desc: not available
URL: <https://lists.bufferbloat.net/pipermail/bloat/attachments/20150422/5d116215/attachment-0003.ods>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: sumstats.sh
Type: application/x-sh
Size: 379 bytes
Desc: not available
URL: <https://lists.bufferbloat.net/pipermail/bloat/attachments/20150422/5d116215/attachment-0006.sh>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: t.sh
Type: application/x-sh
Size: 138 bytes
Desc: not available
URL: <https://lists.bufferbloat.net/pipermail/bloat/attachments/20150422/5d116215/attachment-0007.sh>


More information about the Bloat mailing list