[Bloat] queuebloat

[Jim Gettys]

On 04/13/2011 12:29 PM, Bob Briscoe wrote:
> Jim,
>
> By the end I think I had already addressed a lot of the concerns you 
> stated at the start of the mail:
> - Yes, the name of this exercise is water under the bridge.
> - Buffers still have to be reasonably sized (my footnote covered that 
> already)
>
> However, three responses inline (prefixed "BB:")...
>
> At 15:30 13/04/2011, Jim Gettys wrote:
>> On 04/13/2011 07:19 AM, Bob Briscoe wrote:
>>> The problem is actually queuebloat, not bufferbloat. The buffer is 
>>> the memory set aside for the queue. The queue is how much of the 
>>> memory is used to store packets or frames.
>>
>> I think you are picking nits on the naming, though if you'd had the 
>> suggestion last fall, I might have gone for it
>
> BB: As I said, I'm picking nits on the naming, not suggesting it 
> should be changed at this stage.
>
> But having a misleading name does make the nuancing harder - there's a 
> lot of practitioners out there who don't need or want to understand 
> anything - they have no idea about why they should do things - they 
> just put together strings of feature buzz-words. That's how most of 
> the industry works.
>
> It only needs some researcher with only a partial grasp of the issue 
> to pick up the word bufferbloat as the new sexy research fashion, then 
> publish their research results showing that smaller buffers will make 
> things worse. Then we have to start explaining we didn't really mean 
> bufferbloat, yada yada, and it starts to make us look like we might 
> not have known what we were talking about. While our researcher friend 
> with half a brain starts running around crowing that his marvellous 
> new research has proved us wrong,... when all he's actually done is 
> proved that the word we chose as a name was not quite precise enough.
>

Heh.  We didn't have any term for this at all.  I went back and looked 
at the discussion in end-to-end interest when Dave Reed reported 3g 
bufferbloat, and the suggested alternatives were worse, and no consensus 
reached.

>> And there are buffers that hide in systems that are not packet 
>> queues, that people also should be aware
>> of (e.g. encryption buffers, error correction buffers, buffers in 
>> applications used for pipelining, etc).
>
> BB: Good point. I guess your point again is that many of these buffers 
> are not anything like as parsimoniously sized as they should be.

Yes, often they are infinite and dynamically allocated (e.g. the event 
queue inside of GUI applications and/or window systems themselves).

>
> But these buffers are harder to cut down below a certain minimum, 
> because they actually serve a function. There's no magic like AQM that 
> can keep these buffers unoccupied most of the time.
Sometimes yes, sometimes no. Often, as in packet queues, the buffers 
fill because flow control from lower layers of buffering/queuing have 
filled, and the software is not designed to elide unneeded operations 
when they can't keep up (again, causing buffers/queues to form just 
before the bottleneck).

I'm happy to also use a term queuebloat in places where it is 
applicable, where you have packet queues...    But bufferbloat a generic 
phenomena in communications programming, whether in network transports, 
or in applications using them.  I guess in this I'm an odd-ball, having 
mostly been a programmer who designed network based application.  Let me 
give a concrete example:

Oh, and I forgot about socket buffers, which on modern OS's may also 
automatically resize; these are not queues either.  Even worse, is that 
they will resize based on the underlying confusion induced by other 
bufferbloat/queuebloat underneath them.  These can be controlled by 
applications setting the socket buffer size, rather than taking default 
behaviour.  Again, at least for stream based protocols such as X, these 
aren't yet queues (though we then parse the stream, and generate a queue 
of X events).

A good (recent) example I've seen is in OpenOffice, which has had 
terrible behaviour on its slide arranging operations on Linux for years, 
not understanding it should discard unneeded mouse motion events (seems 
to be one of the things the LibreOffice guys may have fixed, thankfully; 
I talked to Michael Meeks about this a while back).  Bufferbloat affects 
applications just as much as network stacks.

>
>> So I'm not convinced that queuebloat is a better term, as it is less 
>> general than the phenomena I was trying
>> to describe.  In any case, I think it's water under the dam at this 
>> date.
>>
>>>
>>> We don't want vendors to (necessarily*) reduce the size of the 
>>> buffer, we want them to reduce the size of the standing queue. They 
>>> can do that with active queue management (AQM) (if we only knew how 
>>> to code it robustly). Ideally with ECN too, but AQM would be a good 
>>> start.
>>
>> Some of these buffers are truly bloated, and/or not sized even 
>> approximately related to the bandwidth available (e.g. the 1.2 
>> seconds of buffering I observed on my DOCSIS3 modem, or similar 
>> horror stories in DSL), or the 1000 packet transmit queue in Linux.  
>> These buffers are often sized by all the memory that is available, 
>> and the hardware vendors can't get small enough chips to "correctly" 
>> size them, (as though we knew what the bandwidth was, or the delay 
>> was, one of the mythologies that got us into this mess).
>>
>> One of the first steps (well short of the nirvana of AQM), is to at 
>> least get the buffers sized to something sane, and related to the 
>> bandwidth the hardware is being operated at. And as each generation 
>> of new kit is built (and often as a market requirement has to plug 
>> into downward compatible hardware), it's been getting worse.
>>
>>  This is what the cable folks are in the middle of doing; it's 
>> obviously safe to at least have the buffer sizes approximately 
>> proportional to the bandwidth at which the device is operating 
>> (similarly for the Linux transmit queue; if you are at 100Mbps, you 
>> can cut the size by a factor of 10 without any danger).  With the 
>> ability to go hundreds of megabits/second but most customers paying 
>> for 10-20Mbps, it is pretty obvious the buffer size had better be 
>> related to the bandwidth of operation, and never be a static buffer 
>> sized for the worst case.
>>
>> Let's not lose sight of immediate, safe mitigations that are at hand, 
>> while working on AQM with or without ECN, though that is the only 
>> real, long term solution.
>
> BB: The two stage fix might work for some types of product, where 
> continual fixes are the norm. But in other types of product, each fix 
> involves an engineer visit and a box swap out, which you don't want to 
> be doing more than once if you can help it.
>

Yup.  Would that we had AQM's that we knew worked in the face of highly 
variable bandwidth and workloads we could just recommend everyone go 
use: but we're not there yet.

At best, we have some not yet tested ideas and are still getting set up 
to try to run even simple tests (e.g. SFB, RED light when we can get our 
hands on it).

And we certainly *want* operators who could/should be running RED 
already to turn it on in places where it can be used.  My point is 
primarily that the enemy of the good is the perfect, and steps we can 
take to make the problem less severe while working on AQM that can 
handle the current edge are well worth taking.  Sometimes those steps 
may make the problem 1/10th the size it is today.  That doesn't get us 
where we ought to go, but it will reduce suffering.
             - jim