[Cake] [PATCH net-next v7 6/7] sch_cake: Add overhead compensation support to the rate shaper

[Sebastian Moeller]

Hi Toke,



> On May 2, 2018, at 17:30, Toke Høiland-Jørgensen <toke at toke.dk> wrote:
> 
> Sebastian Moeller <moeller0 at gmx.de> writes:
> 
>>> On May 2, 2018, at 17:11, Toke Høiland-Jørgensen <toke at toke.dk> wrote:
>>> 
>>> +		/* The last segment may be shorter; we ignore this, which means
>>> +		 * that we will over-estimate the size of the whole GSO segment
>>> +		 * by the difference in size. This is conservative, so we live
>>> +		 * with that to avoid the complexity of dealing with it.
>>> +		 */
>>> +		len = shinfo->gso_size + hdr_len;
>>> +	}
>> 
>> 
>> Hi Toke,
>> 
>> so I am on the fence with this one, as the extreme case is having a
>> super packet consisting out of 1 full-MTU packet plus a tiny leftover
>> in that case we pay a 50% bandwidth sacrifice which seems a bit high.
>> Nowm I have no real feling how likely this full MTU plus 64 byte
>> packet issue is in real life, but in the past I often saw maximum
>> packetsizes of around 3K bytes on my router indicating that having a
>> sup packet consisting just out of two segments might not be that rare.
>> So is there an easy way for me to measure the probability of seeing
>> that issue?
>> 
>> I am all for sacrificing some bandwidth for better latency under load,
>> but few users will be happy with a 50% loss of bandwidth...
> 
> Well, in most cases such GSO segments will be split anyway (we split if
> <= 1 Gbps). So this inaccuracy will only hit someone who enables the
> shaper *and sets it to a rate rate > 1Gbps*. Which is not a deployment
> mode we have seen a lot of, I think?

	Oh, I agree with that rationale; I was still under the impression that we want to go back to a (configurable) serialization delay based segmentation threshold and then this might become an issue (especially on puny routers will profit from the reduced routing cost* of GSO/GRO). Also I fear that 1Gbps service will become an issue rather sooner than later, even though I would assume that then dual segment super-packets should really be rare...



> 
> But sure, in principle you are right; I have no idea how to measure the
> probability, though. We could conceivably add another statistic, but,
> well, not sure it's worth it... I am certainly not going to do it ;)

	Again, I agree without proof that this is more than a theoretical issue, let's ignore this for now (especially since this is going to interfere with ATM encapsulation, but on ATM the bandwidth should always merit a segmentataion of supers (sorry, ISPs cake is not designed as the customer facing shaper on the DSLAM ;) ))

Best Regards
	Sebastian
> 
> -Toke



*) I guess the kernel's routing codsts pales in comparison with the actual shaping cost, so this might be a bad idea.