> On Mar 3, 2019, at 13:13, Jonathan Morton <[email protected]> wrote:
>
>> On 3 Mar, 2019, at 1:26 pm, Sebastian Moeller <[email protected]> wrote:
>>
>>>> Doesn't this look like ingress magic being applied selectively to the
>>>> users based on number of flows? I thought that the idea behind the ingress
>>>> keyword is to effectively shape harder the more bulk flows are coming in.
>>>
>>> No, it simply counts dropped packets against the shaper, as well as those
>>> actually transmitted.
>>
>> Sure, but the question is how is the resulting "pressure" to drop/mark
>> distributed over the existing (bulk) flows.
>>
>>> There shouldn't be that many packets being dropped to make this much of a
>>> difference.
>>
>> My intuition (probably wrong) is that is not the few packets dropped, but
>> the fact that the the dropping does seem to be restricted to the flows of
>> the IP with more flows, no?
>
> As long as there is a backoff response to a dropped packet, the extra
> back-pressure should be contained to the flows experiencing drops, and other
> flows should see no difference - so you should see a slight reduction in
> goodput on the 16 flows, but *no increase* on the single flow in parallel.
Sure, it looks as if that single flow would see much less drops than
each of the 16 flows of the other IP?
>
> Even if that were not the case, the single flow should take longer on average
> to recover from a cwnd reduction (even in CUBIC) to the fair BDP. That
> should result in a greater reduction in goodput on the single flow than the
> many flows - but we see the reverse here.
If the single flow gets roughly the same per-flow drops/marks as the 16
other flows, that would be the expected behavior. Would it be fair to assume
that this looks like dual-host isolation together with the ingress keyword
results in unequal drop/mark probability based on the IP-address isolation?
>
> So I'm not entirely sure what's happening here, but at least the asymmetry
> isn't too bad; it's achieving significantly better host-fairness than a pure
> flow-fair system would.
>
> - Jonathan Morton
>
_______________________________________________
Cake mailing list
[email protected]
https://lists.bufferbloat.net/listinfo/cake
