fnord is meant to be used under Linux with the diet libc
(http://www.fefe.de/dietlibc/).  These are actual apache bench results,
all on localhost on a 100k JPEG test file, 1000 requests with a
concurrency of 10.

To be fair, I linked thttpd, mini_httpd and fnord against the diet libc.
I did not try this with apache, though.  Since apache does not exec
anything, it should not matter much, though.

mini_httpd forks for each request, apparently does not support
keep-alive and compared to fnord does not incur the overhead of execve
for each request.

thttpd is the fastest web server known to me.

Values are time in seconds for the whole transaction (1000 downloads, 10
parallel connections).

server software		keep-alive		no keep-alive
----------------------------------------------------------------
mini_httpd 1.15c	1.690			0.943
apache 1.3.22		1.236			1.178
mathopd 1.4gamma	3.988			1.497
boa 0.94.12		1.107			1.291
thttpd 2.21b		0.896			0.839
fnord			1.008			1.331
fnord w/ sendfile	0.316			0.912

That number looks like a measuring bug, but I copied it straight from
the apache bench output ;)

Please note that fnord actually plays in the same league as others even
without keep-alive and sendfile support.  That is surprising since fnord
has one fork() _and_ one execve() as overhead for each request!  As the
difference between keep-alive and non-keep-alive shows, that difference
is not very large.  That is the achievement of the diet libc, which
reduces the normally significant libc overhead to zero.