Squid 3.0 Polygraph testing

Hi all,

I've run a couple of polygraph runs against Squid 3.0 last night and
today, with very encouraging results. I have seen zero fatal errors,
even under extreme overload conditions.

Performance with aufs appears to be very similar to that of a similarly
configured 2.5. It is a new test machine: a Via Mini-ITX based machine,
so similarly wimpy compared to my old AMD K6-2 test boxes--except this
one is working with the double disadvantage of having a 4500 RPM disk
instead of the 7200 RPM disks in earlier tests. So I won't know for
sure how performance stacks up compared to 2.5 until I perform tests on
2.5 on the same machine. However, since disk bandwidth is definitely
the limiting factor right now, I wouldn't expect to be able to make
interesting comparisons. I will do some RAM-only tests over the
weekend, to see what CPU usage looks like for the two Squids.

I've also been using it as the office proxy for two days, with no
complaints of errors.

Anyway, it looks like Robert, Henrik, and everyone have done a great job
with the re-factoring into C++. It seems to me a clear sign of very
good coders when such a major change can be made without disastrous
impact on the quality of the product, at least in the short term.
Congratulations on a job well done.

BTW-As an additional data point, for those who might not recall, the
first Squid to exhibit this kind of stability under extreme load was
2.4STABLE. All previous Squids were crashable on demand when placed
under impossible load. This one is absolutely solid and even continues
to reliably serve requests, though very very slowly (this may be because
the test machines run out of file descriptors eventually...placing a
hard upper limit on how overloaded things can get--but that's always
been the case in these tests).

Since epoll support seems to be already in production use, I will see
about giving it a run sometime soon.

I'll publish some results once I've got something interesting to show
(i.e. performance differences between 2.5/3.0, or differences between
3.0 aufs and epoll).

-- 
Joe Cooper <joe@swelltech.com>
Web caching appliances and support.
http://www.swelltech.com