Date: Mon, 25 Apr 2011 17:14:52 -0700 (PDT)
From: david@xxxxxxx
To: Alex Rousskov <rousskov@xxxxxxxxxxxxxxxxxxxxxxx>
Cc: Marcos <mczueira@xxxxxxxxxxxx>, squid-users@xxxxxxxxxxxxxxx,
squid-dev@xxxxxxxxxxxxxxx
Subject: Re: Res: squid 3.2.0.5 smp scaling issues
On Mon, 25 Apr 2011, Alex Rousskov wrote:
On 04/25/2011 05:31 PM, david@xxxxxxx wrote:
On Mon, 25 Apr 2011, david@xxxxxxx wrote:
On Mon, 25 Apr 2011, Alex Rousskov wrote:
On 04/14/2011 09:06 PM, david@xxxxxxx wrote:
In addition, there seems to be some sort of locking betwen the
multiple
worker processes in 3.2 when checking the ACLs
There are pretty much no locks in the current official SMP code.
This
will change as we start adding shared caches in a week or so, but
even
then the ACLs will remain lock-free. There could be some internal
locking in the 3rd-party libraries used by ACLs (regex and such),
but I
do not know much about them.
what are the 3rd party libraries that I would be using?
See "ldd squid". Here is a sample based on a randomly picked Squid:
libnsl, libresolv, libstdc++, libgcc_s, libm, libc, libz, libepol
Please note that I am not saying that any of these have problems in
SMP
environment. I am only saying that Squid itself does not lock anything
runtime so if our suspect is SMP-related locks, they would have to
reside elsewhere. The other possibility is that we should suspect
something else, of course. IMHO, it is more likely to be something
else:
after all, Squid does not use threads, where such problems are
expected.
BTW, do you see more-or-less even load across CPU cores? If not,
you may
need a patch that we find useful on older Linux kernels. It is
discussed
in the "Will similar workers receive similar amount of work?"
section of
http://wiki.squid-cache.org/Features/SmpScale
the load is pretty even across all workers.
with the problems descripted on that page, I would expect uneven
utilization at low loads, but at high loads (with the workers busy
serviceing requests rather than waiting for new connections), I would
expect the work to even out (and the types of hacks described in that
section to end up costing performance, but not in a way that would
scale with the ACL processing load)
one thought I had is that this could be locking on name lookups. how
hard would it be to create a quick patch that would bypass the name
lookups entirely and only do the lookups by IP.
I did not realize your ACLs use DNS lookups. Squid internal DNS code
does not have any runtime SMP locks. However, the presence of DNS
lookups increases the number of suspects.
they don't, everything in my test environment is by IP. But I've seen
other software that still runs everything through name lookups, even
if what's presented to the software (both in what's requested and in
the ACLs) is all done by IPs. It's a easy way to bullet-proof the
input (if it's a name it gets resolved, if it's an IP, the IP comes
back as-is, and it works for IPv4 and IPv6, no need to have logic that
looks at the value and tries to figure out if the user intended to
type a name or an IP). I don't know how squid is working internally
(it's a pretty large codebase, and I haven't tried to really dive into
it) so I don't know if squid does this or not.
A patch you propose does not sound difficult to me, but since I cannot
contribute such a patch soon, it is probably better to test with ACLs
that do not require any DNS lookups instead.
if that regains the speed and/or scalability it would point fingers
fairly conclusively at the DNS components.
this is the only think that I can think of that should be shared
between
multiple workers processing ACLs
but it is _not_ currently shared from Squid point of view.
Ok, I was assuming from the description of things that there would be
one DNS process that all the workers would be accessing. from the way
it's described in the documentation it sounds as if it's already a
separate process, so I was thinking that it was possible that if each
ACL IP address is being put through a single DNS process, I could be
running into contention on that process (and having to do name lookups
for both IPv6 and then falling back to IPv4 would explain the severe
performance hit far more than the difference between IPs being 128 bit
values instead of 32 bit values)
David Lang
