Re: [squid-users] TCP send/receive buffer tuning

Am 01.02.2011 23:03, schrieb Amos Jeffries:
> On Tue, 01 Feb 2011 14:31:02 +0100, Jack Falworth<jackf.mail_at_gmx.de>
> wrote:
>> On 31.01.2011 23:53, Amos Jeffries wrote:
>>> On Mon, 31 Jan 2011 10:57:57 +0100, "Jack Falworth"<jackf.mail_at_gmx.de>
>>> wrote:
>>>> Hi squid-users,
>>>>
>>>> I have a question regarding the TCP send/receive buffer size Squid
> uses.
>>>> For my high-performance setup I increased both buffer sizes on my
> Ubuntu
>>>> 10.04 system. Unfortunately I found out that Squid 2.7 (as well as
> 3.x)
>>>> limits the receive buffer to 64K and the send buffer to 32K in the
>>>> configure.in script.
>>>>
>>>> In addition I found this bug report regarding this check:
>>>> http://bugs.squid-cache.org/show_bug.cgi?id=1075
>>>>
>>>> I couldn't really figure out the problem with Squid using higher
> buffer
>>>> sizes if it is the intention of the administrator to increase those
>>> values.
>>>> This check was included in CVS rev. 1.303 back in 2005, thus it's
> quite
>>>> old.
>>>>
>>>> Is this some legacy check or is it still important with today's
> systems?
>>>> Can I safely remove this check or will this have some side-effects,
> e.g.
>>>> say the some internal data structures won't be able to cope with
> higher
>>>> values?
>>> Note that setting ONLY affects the TCP buffers so 64K worth of packets
>>> can
>>> sit outside of Squid in the networking stack.
>>> This has side-effects on the ACK packets. While they are waiting in
> that
>>> buffer they are possibly ACKed but not actually received by Squid. If
>>> anything causes Squid to stop, crash or slow down on its read()'s and
>>> accept()'s the client can be left with incorrect information about the
>>> state of those bytes.
>>>
>> But this could also happen on a 64K buffer. If Squid crashes or goes
>> down for some reason then information is lost anyways.
>> Thus the only reason increasing the buffer size in a high-traffic
>> scenario would be a bad idea is if Squid is somehow overloaded and
>> slowed down on its read()'s and accept()'s? But if I make sure that
>> Squid can handle some peak traffic values without being overloaded it
>> would be safe to increase buffer sizes?
> It's the only reason I'm aware of. There may be another I'm not.
> Feel free to experiment and see for yourself of course.
>
> Amos
I digged a little bit deeper and found some problems that would occur
when increasing the receive buffer.

The receive buffer size is stored in SQUID_TCP_SO_RCVBUF. There are some
code parts where this size is assigned to size_t variables which are
normally 16bit sized. If greater value are assigned the size
would wrap around. Critically parts are in the functions httpReadReply
(http.c), sslSetSelect (ssl.c) and clientEatRequestBodyHandler
(client_side.c).
So in my opinion there is no way to get this working without (some) code
rewrite and verifying these critical sections. Correct me if I'm wrong.
Received on Wed Feb 02 2011 - 09:49:52 MST