[Bloat] Cake "out of tree" on Debian

[Dave Taht]

On Sun, Jan 29, 2017 at 9:48 AM, Kristian Kielhofner <kris at kriskinc.com> wrote:
> On Sun, Jan 29, 2017 at 12:42 PM, Dave Taht <dave.taht at gmail.com> wrote:
>> tc should not be showing ?diffserv?
>>
>> did you pull your tc from here? https://github.com/dtaht/tc-adv -
>> fixed this yesterday, or so I thought.
>>
>> Also: I just refreshed iproute to head and rebased - (I needed working
>> ra expires support, and it's long past time we updated - regrettably
>> lede is on 4.4, can't change that, there). It's generally safe to be
>> running a later iproute2 than the kernel you have.
>>
>> https://github.com/dtaht/iproute2-head-cake
>>
>
> Now:
>
> UP:
> qdisc cake 8003: root refcnt 9 bandwidth 22Mbit diffserv3
> triple-isolate rtt 100.0ms raw
>  Sent 12868692 bytes 8521 pkt (dropped 0, overlimits 10941 requeues 0)
>  backlog 7570b 5p requeues 0
>  memory used: 44320b of 4Mb
>  capacity estimate: 22Mbit
>                  Bulk   Best Effort      Voice
>   thresh      1375Kbit      22Mbit    5500Kbit
>   target        13.2ms       5.0ms       5.0ms
>   interval     108.2ms     100.0ms      10.0ms
>   pk_delay         0us       7.0ms         0us
>   av_delay         0us       2.7ms         0us
>   sp_delay         0us       465us         0us
>   pkts               0        8530           1
>   bytes              0    12883790          42
>   way_inds           0           0           0
>   way_miss           0          10           1
>   way_cols           0           0           0
>   drops              0           0           0
>   marks              0           0           0
>   sp_flows           0           0           0
>   bk_flows           0           1           0
>   un_flows           0           0           0
>   max_len            0       10598          42
>
> qdisc ingress ffff: parent ffff:fff1 ----------------
>  Sent 264453 bytes 4795 pkt (dropped 0, overlimits 0 requeues 0)
>  backlog 0b 0p requeues 0


Groovy. I thought you were "Mr. VOIP"?

While the default fq scheme works really well in cake, if you test
marking packets as voip (EF, CS4,CS6,CS7,VA), it will end up in the
diffserv3 voice queue.

asterisk used to have an encapsulating protocol called iax2, which
generated a single flow as backhaul - is that still deployed?

We have a tool in flent based on dit-itg to test this. It's a bit
painful to setup the first time. I've longed to have a full
asterisk/freeswitch/jitsy test setup to look harder at voip/video
characteristics.

There are also now several as yet underdocumented options in cake:

"nat" will look at flows before they are natted so as to isolate them better.
When combined with the dual-dsthost or dual-srchost option (depending
on traffic direction) that gives you per host fq, along with per-flow
fq.

(nat triple-isolate should also do this but we're still sorting out a
bug on that:
https://github.com/dtaht/sch_cake/issues/46
)

wash: washes out dscp markings. Helpful when your provider (:cough:
comcast) remarks nearly all traffic to CS1.

I just found that appear.in is using the new "goog" marker, which
marks all videoconferencing traffic as AF41, which is more or less
appropriately handled in the "diffserv4" model.

In general I have always had good results with the simplest
(besteffort or diffserv3) settings.

There's also new support for a docsis mode using the new "mpu" idea.

https://github.com/dtaht/sch_cake/pull/45

>
> DOWN:
> qdisc cake 8004: root refcnt 2 bandwidth 130Mbit besteffort
> triple-isolate rtt 100.0ms raw
>  Sent 331715 bytes 4797 pkt (dropped 0, overlimits 610 requeues 0)
>  backlog 0b 0p requeues 0
>  memory used: 1792b of 6500000b
>  capacity estimate: 130Mbit
>                  Tin 0
>   thresh       130Mbit
>   target         5.0ms
>   interval     100.0ms
>   pk_delay        13us
>   av_delay         1us
>   sp_delay         0us
>   pkts            4797
>   bytes         331715
>   way_inds           0
>   way_miss           9
>   way_cols           0
>   drops              0
>   marks              0
>   sp_flows           0
>   bk_flows           1
>   un_flows           0
>   max_len          198

Blow it up. Tell us what happens. :)

>
> --
> Kristian Kielhofner



-- 
Dave Täht
Let's go make home routers and wifi faster! With better software!
http://blog.cerowrt.org