[Cerowrt-devel] Ideas on how to simplify and popularize bufferbloat control for consideration.

[Sebastian Moeller]

Hi David,

On Jul 27, 2014, at 00:53 , David Lang <david at lang.hm> wrote:

> On Sun, 27 Jul 2014, Sebastian Moeller wrote:
> 
>> Hi David,
>> 
>> On Jul 26, 2014, at 23:45 , David Lang <david at lang.hm> wrote:
>> 
>>> On Sat, 26 Jul 2014, Sebastian Moeller wrote:
>>> 
>>>> On Jul 26, 2014, at 22:39 , David Lang <david at lang.hm> wrote:
>>>> 
>>>>> by how much tuning is required, I wasn't meaning how frequently to tune, but how close default settings can come to the performance of a expertly tuned setup.
>>>> 
>>>> 	Good question.
>>>> 
>>>>> 
>>>>> Ideally the tuning takes into account the characteristics of the hardware of the link layer. If it's IP encapsulated in something else (ATM, PPPoE, VPN, VLAN tagging, ethernet with jumbo packet support for example), then you have overhead from the encapsulation that you would ideally take into account when tuning things.
>>>>> 
>>>>> the question I'm talking about below is how much do you loose compared to the idea if you ignore this sort of thing and just assume that the wire is dumb and puts the bits on them as you send them? By dumb I mean don't even allow for inter-packet gaps, don't measure the bandwidth, don't try to pace inbound connections by the timing of your acks, etc. Just run BQL and fq_codel and start the BQL sizes based on the wire speed of your link (Gig-E on the 3800) and shrink them based on long-term passive observation of the sender.
>>>> 
>>>> 	As data talks I just did a quick experiment with my ADSL2+ koine at home. The solid lines in the attached plot show the results for proper shaping with SQM (shaping to 95% of del link rates of downstream and upstream while taking the link layer properties, that is ATM encapsulation and per packet overhead into account) the broken lines show the same system with just the link layer adjustments and per packet overhead adjustments disabled, but still shaping to 95% of link rate (this is roughly equivalent to 15% underestimation of the packet size). The actual theist is netperf-wrappers RRUL (4 tcp streams up, 4 tcp steams down while measuring latency with ping and UDP probes). As you can see from the plot just getting the link layer encapsulation wrong destroys latency under load badly. The host is ~52ms RTT away, and with fq_codel the ping time per leg is just increased one codel target of 5ms each resulting in an modest latency increase of ~10ms with proper shaping for a total of ~65ms, with improper shaping RTTs increase to ~95ms (they almost double), so RTT increases by ~43ms. Also note how the extremes for the broken lines are much worse than for the solid lines. In short I would estimate that a slight misjudgment (15%) results in almost 80% increase of latency under load. In other words getting the rates right matters a lot. (I should also note that in my setup there is a secondary router that limits RTT to max 300ms, otherwise the broken lines might look even worse...)
>>> 
>>> what is the latency like without BQL and codel? the pre-bufferbloat version? (without any traffic shaping)
>> 
>> 	So I just disabled SQM and the plot looks almost exactly like the broken line plot I sent before (~95ms RTT up from 55ms unloaded, with single pings delayed for > 1000ms, just as with the broken line, with proper shaping even extreme pings stay < 100ms). But as I said before I need to run through my ISP supplied primary router (not just a dumb modem) that also tries to bound the latencies under load to some degree. Actually I just repeated the test connected directly to the primary router and get the same ~95ms average ping time with frequent extremes > 1000ms, so it looks like just getting the shaping wrong by 15% eradicates the buffer de-bloating efforts completely...
> 
> just so I understand this completely
> 
> you have
> 
> debloated box <-> ISP router <-> ADSL <-> Internet <-> debloated server?

	Well more like:

	Macbook with dubious bloat-state -> wifi to de-bloated cerowrt box that shapes the traffic -> ISP router -> ADSL -> internet -> server

I assume that Dave debated these servers well, but it should not really matter as the problem are the buffers on both ends of the bottleneck ADSL link.


> 
> and are you measuring the latency impact when uploading or downloading?

	No I measure the impact of latency of saturating both up- and downlink, pretty much the worst case scenario.

> 
> I think a lot of people would be happy with 95ms average pings on a loaded connection, even with occasional outliers.

	No that is too low an aim, this still is not useable for real time applications, we should aim for base RTT plus 10ms. (For very slow links we need to cut some slack but for > 3Mbps 10ms should be achievable )

> It's far better than sustained multi-second ping times which is what I've seen with stock setups.

	True, but compared to multi seconds even <1000ms would be a really great improvement, but also not enough.

> 
> but if no estimate is this bad, how bad is it if you use as your estimate the 'rated' speed of your DSL (i.e. what the ISP claims they are providing you) instead of the fully accurate speed that includes accounting for ATM encapsulation?

	Well ~95ms with outliers > 1000ms, just as bad as no estimate. I shaped 5% below rated speed as reported by the DSL modem, so disabling the ATM link layer adjustments (as shown in the broken lines in the plot), basically increased the effective shaped rate by ~13% or to effectively 107% of line rate, your proposal would be line rate and no link layer adjustments or effectively 110% of line rate; I do not feel like repeating this experiment right now as I think the data so far shows that even with less misjudgment the bloat effect is fully visible ) Not accounting for ATM framing carries a ~10% cost in link speed, as ATM packet size on the wire increases by >= ~10%.

> 
> It's also worth figuring out if this problem would remain in place if you didn't have to go through the ISP router and were runing fq_codel on that router.

	If the DSL modem would be debloated at least on upstream no shaping would be required any more; but that does not fix the need for downstream shaping (and bandwidth estimation) until the head end gear is debloated..

> As long as fixing bufferbloat involves esoteric measurements and tuning, it's not going to be solved, but if it could be solved by people flahing openwrt onto their DSL router and then using the defaults, it could gain traction fairly quickly.

	But as there are only very few DSL modems with open sources (especially of the DSL chips) this just as esoteric ;) Really if equipment manufactures could be convinced to take these issues seriously and actually fix their gear that would be best. But this does not look like it is happening on the fast track. (Even DOCSIS developer cable labs punted on requiring codel or fq_codel in DOCSIS modems since the think that the required timestamps are to “expensive” on the device class they want to use for modems. They opted for PIE, much better than what we have right now but far away from my latency under load increase of 10ms...)

> 
>>> I agree that going from 65ms to 95ms seems significant, but if the stock version goes into up above 1000ms, then I think we are talking about things that are ‘close'
>> 
>> 	Well if we include outliers (and we should as enough outliers will degrade the FPS and voip suitability of an otherwise responsive system quickly) stock and improper shaping are in the >1000ms worst case range, while proper SQM bounds this to 100ms.
>> 
>>> 
>>> assuming that latency under load without the improvents got >1000ms
>>> 
>>> fast-slow (in ms)
>> ideal=10
>> untuned=43
>> bloated > 1000
>> 
>> 	The sign seems off as fast < slow? I like this best ;)
> 
> yep, I reversed fast/slow in all of these
> 
>>> 
>>> fast/slow
>>> ideal = 1.25
>>> untuned = 1.83
>>> bloated > 19
>> 
>> 	But Fast < Slow and hence this ration should be <0?
> 
> 1 not 0, but yes, this is really slow/fast
> 
>>> slow/fast
>>> ideal = 0.8
>>> untuned = 0.55
>>> bloated = 0.05
>>> 
>> 
>> 	and this >0?
> 
> and this is really fast/slow


	What about taking the latency difference an re;aging it with a reference time, like say the time a photon would take to travel once around the equator, or the earth’s diamater?

Best Regards
	Sebastian



> 
> David Lang