[Cake] The two SCE tests I have in mind

Sebastian Moeller moeller0 at gmx.de
Sun Mar 24 18:31:09 EDT 2019


Hi All,

while I am late to the party, I could run a client side test tonight. Nothing special, just a residential O2/Telefonica (via L2-bitstream access) VDSL2 50/10 link in Germany. Would a capture of that be of any help? If yes I am happy to test, otherwise I will wait for the dust to settle.

BTW, later this week it would be great to create sharable client and server VMs, just as the LLLLS folks should start distributing, but that certainly is for after the meeting.

P.S.: how likely is that meeting to be on time (I need to figure out how to schedule work around this)?

Best Regards, or better Ahoy to the crew of the buffer boat

	Sebastian



> On Mar 24, 2019, at 11:30, Dave Taht <dave.taht at gmail.com> wrote:
> 
> On Sun, Mar 24, 2019 at 10:30 AM Luca Muscariello
> <luca.muscariello at gmail.com> wrote:
>> 
>> We have done something like that a year ago in our team in Cisco
>> for a project that you Dave are aware of but that is out of scope for these lists.
>> 
>> It is important to have vantage points in residential and enterprise networks.
> 
> It is hard to get dedicated testing points in those networks. However
> providing tools so more volunteers can participate in a larger test
> has always proven worthwhile.
> 
>> Cloud to Cloud never goes to transit and the big Clouds have global footprints.
>> So you can traverse the planet inside the Cloud WAN (100% bit consistency) and then traverse a safe peering
>> point where the adjacency  has been well set (or not).
> 
> cloud to cloud should be interesting.
> 
> Anyway, the first sce enabled version of cake for public testing is
> now up at cake-sce-de.teklibre.net. (dns still propigating) It is rate
> limited to 40Mbits and thus with 10 flows going, self congests and
> thus marks with both SCE and CE. Inititial tests across our cloudy
> backbone have been good, no reordering, no drops, while SCE marking
> like crazy. The needed code to setup your own server is in my github
> sch_cake and tc-adv repos.
> 
> We hope to roll that out over linode, google cloud, and aws over the
> coming days.
> 
> A typical test right now, looks like this:
> 
> # server side install if you want to set up your own test server
> # after an:
> 
> apt-get install flent netperf irtt build-essential flex bison
> git clone https://github.com/dtaht/sch_cake
> git clone https://github.com/dtaht/tc-adv
> git clone https://github.com/dtaht/fq_codel_fast
> git clone https://github.com/dtaht/fq_codel_fast
> git clone https://github.com/tohojo/sqm-scripts
> 
> for i in sch_cake tc-adv fq_codel_fast sqm-scripts
> do
> cd $i; make; make install; cd ..
> done
> 
> tc qdisc replace dev eth0 root cake sce bandwidth 40mbit
> 
> # Client side test
> 
> tcpdump -i eth0 -s 128 -w mylocationinfo.cap
> flent -H  cake-sce-de.teklibre.net -l 20 -t my_location_info
> --te=download_streams tcp_ndown
> killall tcpdump
> 
> Go looking at the cap for sack, drops, reorders, and the sce bit being excerted.
> 
> (flent's --socket-stats option got broken in a recent tc release which
> we need to fix for that and qdisc stats)
> 
> if you are running your own server, you can now also see the
> sce marking rate via:
> 
> tc -s qdisc show dev eth0
> 
> qdisc cake 8002: root refcnt 2 bandwidth 40Mbit diffserv3
> triple-isolate nonat nowash no-ack-filter split-gso rtt 100.0ms raw
> overhead 0 sce
> 
> Sent 37542361 bytes 25522 pkt (dropped 0, overlimits 30923 requeues 0)
> backlog 0b 0p requeues 0
> memory used: 124568b of 4Mb
> capacity estimate: 40Mbit
> min/max network layer size:           42 /    1514
> min/max overhead-adjusted size:       42 /    1514
> average network hdr offset:           14
> 
> 
>                   Bulk  Best Effort        Voice
> 
>  thresh       2500Kbit       40Mbit       10Mbit
>  target          7.3ms        5.0ms        5.0ms
>  interval      102.3ms      100.0ms      100.0ms
>  pk_delay          0us        528us        347us
>  av_delay          0us        201us         10us
>  sp_delay          0us          4us          5us
>  backlog            0b           0b           0b
>  pkts                0        25209          313
>  bytes               0     37500795        41566
>  way_inds            0            0            0
>  way_miss            0          106            5
>  way_cols            0            0            0
>  drops               0            0            0
>  marks               0           24            0
>  ack_drop            0            0            0
>  sp_flows            0            1            0
>  bk_flows            0            0            1
>  un_flows            0            0            0
>  max_len             0         3028         1752
>  quantum           300         1220          305
>  sce                 0        23105            0
> 
> 
> setting up the sqm-scripts to use fq_codel_fast is a matter of setting
> the right parameters in a /etc/sqm/eth0.iface.conf file for bandwidth
> (40mbit) and qos model (simplest.qos) and
> 
> IQDISC_OPTS="ce_threshold 2us" # these are artificially low - 1-2ms
> should be closer to "right" on cloudy devices for a real, rather than
> test, deployment
> EQDISC_OPTS="ce_threshold 2us"
> 
> 
>> IPv4/IPv6 of course we'll see a big difference in favour of IPv6.
>> What Michael has shown is his paper makes a lot of sense to me.
>> Depending on the use case hit ratio can be very high or very low.
> 
> I am very behind on reading papers this week.
> 
>> 
>> 
>> 
>> 
>> On Sun, Mar 24, 2019 at 9:55 AM Dave Taht <dave.taht at gmail.com> wrote:
>>> 
>>> 1) The research into whether bit flipping to the extent that SCE will
>>> do has not been done yet. The study of ECT(0) vs ECT(1) behavior
>>> transiting to CE was a little lightweight.
>>> 
>>> To test this we going to fire up a ton of nanodes in various data
>>> centers, with low SCE thresholds, and low bandwidths, to flip lots of
>>> bits, and test between the data centers and from as many vantage
>>> points around the net as we can get - do packet captures as well as
>>> flent tests
>>> 
>>> as a control, set up identical boxes, with SCE disabled, in the same
>>> data centers.
>>> 
>>> Setup flent, irtt, iperf3.
>>> 
>>> 2) Diffserv bit preservation test
>>> 
>>> The research going by on the tsvwg mailing seems a bit dated. It is
>>> very straightforward to use irtt to test to see what udp codepoints
>>> survive e2e, and to also leverage this testbed setup. Similarly,
>>> netperf can easily be used to mark tcp. We do not have a good packet
>>> cap tool to verify that the bits are being set right, however I think
>>> irtt can be modified to check for correctness here and produce a
>>> report.
>>> 
>>> On Sun, Mar 24, 2019 at 1:45 AM Jonathan Morton <chromatix99 at gmail.com> wrote:
>>>> 
>>>>> On 24 Mar, 2019, at 8:37 am, Pete Heist <pete at heistp.net> wrote:
>>>>> 
>>>>> I should theoretically arrive at the boat today some time after 3pm, having picked up a mini HDMI to HDMI adapter, which we can use with the cable that’s there…
>>>> 
>>>> Awesome.  I'm also setting up a Linux VM on my Mac, which should help things along.
>>>> 
>>>> We're bringing up some actual hardware with the SCE-enabled Cake on it now.  Dave wants to investigate various theoretical phenomena the Internet might exhibit with a mixture of ECN codepoints; I just want to be sure it actually works as intended, before I move on to fiddling with TCP.
>>>> 
>>>> - Jonathan Morton
>>>> 
>>>> _______________________________________________
>>>> Cake mailing list
>>>> Cake at lists.bufferbloat.net
>>>> https://lists.bufferbloat.net/listinfo/cake
>>> 
>>> 
>>> 
>>> --
>>> 
>>> Dave Täht
>>> CTO, TekLibre, LLC
>>> http://www.teklibre.com
>>> Tel: 1-831-205-9740
>>> _______________________________________________
>>> Cake mailing list
>>> Cake at lists.bufferbloat.net
>>> https://lists.bufferbloat.net/listinfo/cake
> 
> 
> 
> -- 
> 
> Dave Täht
> CTO, TekLibre, LLC
> http://www.teklibre.com
> Tel: 1-831-205-9740
> _______________________________________________
> Cake mailing list
> Cake at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cake



More information about the Cake mailing list