Re: tux3-2.6.8.1-A5 performance

[Sapan Bhatia <sapan.bhatia@inria.fr>]

Hi,

There have been other user-space servers that have outperformed TUX. X15
was one that got some attention a couple of years ago. This year's
USENIX had two papers that sped up user-space servers to get SPECWEB
performance comparable to TUX, so it's really not surprising (also
considering that you're using one of the featured servers). 

I'm curious - which OS did you benchmark Flash on? The USENIX paper had
numbers for Flash on FreeBSD (AFAIR). If it was on Linux, which family
of syscalls does it use for polling (select/poll/epoll)?

I think that being in kernel space should give TUX a big advantage in
truly high-concurrency workloads... since a lot of overheads, like
mappings from file descriptors to socket structures get bypassed for
every concurrent request, and operations like accepting a bunch of
connections happen in a tight in-kernel loop.

Which is to say, that although using httperf' is an excellent
indication, since it gives you the amount of uniform traffic your server
can handle before reaching failure; tolerance to high concurrencies is
an important indicator as well...

The problem with using httperf (assuming that your tool is adequately
similar to httperf) to test that is that load concurrency only gets
escalated when the server is close to failure. For the most, concurrency
does not exceed a few dozen requests at the same time. So I'd agree with
Kees' remark and suggest (if you want to evaluate this further) that you
try to benchmark with ab and vary the concurrency from a few connections
to a few thousand concurrent connections and observe the level of
performance degradation compared to TUX.

Sapan