W3C home > Mailing lists > Public > ietf-http-wg@w3.org > January to March 2013

Re: http/2 prioritization/fairness bug with proxies

From: Martin Thomson <martin.thomson@gmail.com>
Date: Mon, 11 Feb 2013 15:35:00 -0800
Message-ID: <CABkgnnU4HtLcqC=6GiYrRgZLdH4WjHou68O84Sixi1tLFfaYqQ@mail.gmail.com>
To: Roberto Peon <grmocg@gmail.com>
Cc: William Chan (陈智昌) <willchan@chromium.org>, HTTP Working Group <ietf-http-wg@w3.org>
My bad, I jumped from that to the next point.  That is, once you start
de-prioritizing streams, you have to have a mechanism to prevent you
from being overwhelmed by incoming packets on that stream.

Note also that the proxy can apply prioritization based on more than
just the priority field.  It can actively avoid starvation based on
more than just the priority field.

(On a random hunch, could intermediaries need flow control windows
sized for the BDP of the complete path from origin to client?)

On 11 February 2013 15:27, Roberto Peon <grmocg@gmail.com> wrote:
> If that is what you're saying, then I definitely misunderstood. I thought
> you were advocating using flow control to reduce the effective priority of a
> long-running item?
> -=R
>
>
> On Mon, Feb 11, 2013 at 3:21 PM, Martin Thomson <martin.thomson@gmail.com>
> wrote:
>>
>> That's not at all what I had in mind, just a general re-iteration of
>> the basic principles:
>>
>> Primarily:  Send bits based on availability and relative priorities.
>> Secondarily:  Advertise flow control window availability based on when
>> those bits are consumed (i.e. sent onwards).
>>
>> On 11 February 2013 14:57, Roberto Peon <grmocg@gmail.com> wrote:
>> > I think that the overall complexity spend is lower when we put the
>> > complexity into prioritization mechanisms instead of stuffing heuristics
>> > into the flow control... (because that is just ewww and ouch in so many
>> > ways..)
>> >
>> > -=R
>> >
>> >
>> > On Mon, Feb 11, 2013 at 2:53 PM, Martin Thomson
>> > <martin.thomson@gmail.com>
>> > wrote:
>> >>
>> >> I'm thinking a simple mechanism.  The proxy can feed requests back to
>> >> clients based on its own prioritization and use flow control windows
>> >> on the back-end connections to ensure that it doesn't have to buffer
>> >> infinitely.  Basically, the same sorts of logic that would be needed
>> >> by a proxy that serves multiple clients.
>> >>
>> >> Yes, this is sub-optimal because it leads to either poor bandwidth
>> >> utilization, lots of buffering, or worse.
>> >>
>> >> Sure, it's very easy to get this wrong.  I don't believe it
>> >> unreasonable to consider this sort of thing to be marked "hard hat
>> >> area".  In both directions.
>> >>
>> >> On 11 February 2013 14:46, William Chan (陈智昌) <willchan@chromium.org>
>> >> wrote:
>> >> > Theoretically possible is one thing. But the moment we get into the
>> >> > game
>> >> > of
>> >> > trying to carve up portions of BDP via per-stream flow control
>> >> > windows
>> >> > for
>> >> > prioritization purposes in normal operation (as opposed to just
>> >> > trying
>> >> > to
>> >> > make reasonable progress under excessive load), I think we're in
>> >> > trouble,
>> >> > and likely to get into performance issues due to poor
>> >> > implementations.
>> >> > As
>> >> > I've stated before, I hope most implementations (and believe we
>> >> > should
>> >> > add
>> >> > recommendations for this behavior) only use flow control (if they use
>> >> > it
>> >> > at
>> >> > all, which hopefully they don't because it's hard) for maintaining
>> >> > reasonable buffer sizes.
>> >> >
>> >> >
>> >> > On Mon, Feb 11, 2013 at 2:39 PM, Martin Thomson
>> >> > <martin.thomson@gmail.com>
>> >> > wrote:
>> >> >>
>> >> >> Keep in mind that it is always possible for the intermediary to
>> >> >> apply
>> >> >> flow control to the infinite length stream so that, regardless of
>> >> >> priority, it doesn't consume more than its fair share.
>> >> >>
>> >> >> On 10 February 2013 14:40, William Chan (陈智昌)
>> >> >> <willchan@chromium.org>
>> >> >> wrote:
>> >> >> > First, I totally agree SPDY/4 prioritization changes are far more
>> >> >> > reaching.
>> >> >> > Let's not talk about them yet. I share your complexity concern and
>> >> >> > agree
>> >> >> > we
>> >> >> > need data before proceeding with many of those features, and I
>> >> >> > plan
>> >> >> > on
>> >> >> > getting data.
>> >> >> >
>> >> >> > As far as grouping, I think it's definitely a bug for the case
>> >> >> > where
>> >> >> > you
>> >> >> > have a forward proxy with users sharing the same HTTP/2.0
>> >> >> > connection
>> >> >> > to
>> >> >> > an
>> >> >> > origin server. You can have many high priority short-lived streams
>> >> >> > which
>> >> >> > can
>> >> >> > starve out others, unless we implement the vague notion of "don't
>> >> >> > starve
>> >> >> > out
>> >> >> > streams", which is difficult when it might be a sustained rate of
>> >> >> > high
>> >> >> > priority short-lived streams. It's easier in your latter case of
>> >> >> > infinite,
>> >> >> > high-priority streams.
>> >> >> >
>> >> >> > You're right that the high priority, infinite stream can starve
>> >> >> > streams
>> >> >> > within the same group. I don't think this means that grouping is
>> >> >> > not
>> >> >> > required, but that grouping is potentially insufficient. For
>> >> >> > intragroup
>> >> >> > starvation, I think it's debatable about whether the server should
>> >> >> > be
>> >> >> > smart
>> >> >> > and not allow streams to _completely_ starve other streams within
>> >> >> > a
>> >> >> > group,
>> >> >> > or that clients should have a reprioritization facility. I think
>> >> >> > this
>> >> >> > is
>> >> >> > a
>> >> >> > discussion worth having, but I'd personally classify it as
>> >> >> > separate
>> >> >> > from
>> >> >> > whether or not the grouping feature is necessary.
>> >> >> >
>> >> >> > On Sun, Feb 10, 2013 at 2:06 PM, Mike Belshe <mike@belshe.com>
>> >> >> > wrote:
>> >> >> >>
>> >> >> >>
>> >> >> >>
>> >> >> >>
>> >> >> >> On Mon, Feb 4, 2013 at 5:47 PM, William Chan (陈智昌)
>> >> >> >> <willchan@chromium.org>
>> >> >> >> wrote:
>> >> >> >>>
>> >> >> >>> I'm sorry if I am unclear in any way. Please continue to
>> >> >> >>> challenge/question my comments/assertions so I can clarify my
>> >> >> >>> position
>> >> >> >>> as appropriate.
>> >> >> >>>
>> >> >> >>> Just to be clear here, I stand by that it's a protocol bug
>> >> >> >>> currently.
>> >> >> >>
>> >> >> >>
>> >> >> >> +0.5.  :-)
>> >> >> >>
>> >> >> >> Mark - I think we could open a issue ticket with the current
>> >> >> >> HTTP/2.0
>> >> >> >> draft that this is a bug which will present itself for servers
>> >> >> >> that
>> >> >> >> implement naively.  This isn't strictly a "bug", since the spec
>> >> >> >> says
>> >> >> >> the
>> >> >> >> server can do whatever it wants with priorities, but this is
>> >> >> >> subtle
>> >> >> >> enough
>> >> >> >> (surfacing primarily in http/2 -> http/2 proxy situations), that
>> >> >> >> many
>> >> >> >> server
>> >> >> >> implementors won't think of it unless we mention it in the spec.
>> >> >> >>
>> >> >> >> The bare minimum would be to simply document it and tell servers
>> >> >> >> not
>> >> >> >> to
>> >> >> >> starve out streams.  However, this is probably a wimpy approach
>> >> >> >> and
>> >> >> >> I
>> >> >> >> think
>> >> >> >> we can do better.
>> >> >> >>
>> >> >> >> The SPDY/4 prioritization changes are far more reaching than just
>> >> >> >> fixing
>> >> >> >> this bug.  While I like grouping as a feature, I don't believe it
>> >> >> >> actually
>> >> >> >> fixes this bug:  a browser could open a high priority, infinite
>> >> >> >> stream,
>> >> >> >> which is competing across a shared proxy backend with other
>> >> >> >> streams;
>> >> >> >> unless
>> >> >> >> the user manually switches tabs (or does something to force
>> >> >> >> changed
>> >> >> >> groups),
>> >> >> >> the starvation can still occur.
>> >> >> >>
>> >> >> >> Mike
>> >> >> >>
>> >> >> >>
>> >> >> >>
>> >> >> >>
>> >> >> >>>
>> >> >> >>> I agree with adding more hooks to convey advisory priority
>> >> >> >>> semantics.
>> >> >> >>> That said, "advisory" is open to interpretation. I agree that
>> >> >> >>> the
>> >> >> >>> sender should ultimately be in control of how it orders
>> >> >> >>> responses,
>> >> >> >>> and
>> >> >> >>> indeed there are of course many situations where it's best for
>> >> >> >>> the
>> >> >> >>> sender to ignore the advisory priority. Yet, if the advisory
>> >> >> >>> priority
>> >> >> >>> semantics are generally not respected, then clients will not be
>> >> >> >>> able
>> >> >> >>> to rely on them, and will be forced to implement prioritization
>> >> >> >>> at
>> >> >> >>> a
>> >> >> >>> higher layer, which suffers from the link underutilization vs
>> >> >> >>> contention tradeoff I highlighted earlier.
>> >> >> >>>
>> >> >> >>> I appreciate the concern that we're adding complexity by
>> >> >> >>> introducing
>> >> >> >>> new semantics. I am arguing that because the existing mechanisms
>> >> >> >>> for
>> >> >> >>> addressing starvation are suboptimal, we should treat this as a
>> >> >> >>> protocol bug and thus change the protocol in such a way as to
>> >> >> >>> fix
>> >> >> >>> this
>> >> >> >>> problem. My suggestion for doing so was adding new priority
>> >> >> >>> "grouping"
>> >> >> >>> semantics. I am hopeful that these new semantics will not
>> >> >> >>> introduce
>> >> >> >>> an
>> >> >> >>> inordinate amount of specification, as the primary idea is that
>> >> >> >>> the
>> >> >> >>> current SPDY priority levels would apply within a "group". I
>> >> >> >>> think
>> >> >> >>> we
>> >> >> >>> can come up with a way to define a group that will be relatively
>> >> >> >>> easy
>> >> >> >>> to spec.
>> >> >> >>>
>> >> >> >>> SPDY/4 introduces other prioritization semantics beyond just
>> >> >> >>> grouping,
>> >> >> >>> but I wanted to focus on this one first, as I believe this is a
>> >> >> >>> bug
>> >> >> >>> that we *need* to fix. The other SPDY/4 priority changes are of
>> >> >> >>> a
>> >> >> >>> performance optimization nature, and I believe they will need to
>> >> >> >>> be
>> >> >> >>> justified by data. I have no plans to raise them up in this
>> >> >> >>> group
>> >> >> >>> until we have said data.
>> >> >> >>>
>> >> >> >>> On Tue, Feb 5, 2013 at 5:34 AM, Poul-Henning Kamp
>> >> >> >>> <phk@phk.freebsd.dk>
>> >> >> >>> wrote:
>> >> >> >>> > Content-Type: text/plain; charset=ISO-8859-1
>> >> >> >>> > --------
>> >> >> >>> > In message
>> >> >> >>> > <42A54D15-0AA3-4172-94F7-E94C86E84D7F@niven-jenkins.co.uk>,
>> >> >> >>> > Ben Nive
>> >> >> >>> > n-Jenkins writes:
>> >> >> >>> >
>> >> >> >>> >>So the idea is the protocol contains enough 'hooks' to
>> >> >> >>> >> sufficiently
>> >> >> >>> >>express the different priorities between & within groups that
>> >> >> >>> >> folks
>> >> >> >>> >>would like to express but isn't prescriptive about how anyone
>> >> >> >>> >> uses
>> >> >> >>> >> or
>> >> >> >>> >>implements different prioritisation, scheduling, etc schemes.
>> >> >> >>> >
>> >> >> >>> > That was clearly not how the original poster presented it:
>> >> >> >>> >
>> >> >> >>> >         "I consider all those options as suboptimal, and thus
>> >> >> >>> >         consider this issue to be a protocol bug. Our SPDY/4
>> >> >> >>> >         prioritization proposal addresses this by [...]"
>> >> >> >>> >
>> >> >> >>> > --
>> >> >> >>> > Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
>> >> >> >>> > phk@FreeBSD.ORG         | TCP/IP since RFC 956
>> >> >> >>> > FreeBSD committer       | BSD since 4.3-tahoe
>> >> >> >>> > Never attribute to malice what can adequately be explained by
>> >> >> >>> > incompetence.
>> >> >> >>>
>> >> >> >>
>> >> >> >
>> >> >
>> >> >
>> >>
>> >
>
>
Received on Monday, 11 February 2013 23:35:30 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Monday, 11 February 2013 23:35:34 GMT