> He has some "smoking" figures for some servers, but they're run in loopback
> (no network latency).  You can get high hits/sec figures with small pages and
> high Kb/sec with large pages, so these can be tailored as well.  Also omitted
> was the concurrency levels, the size of the server pool (where applicable),
> and other *very* influential factors.  Not that he was trying to bias the
> figures (don't get that impression) just saying "your mileage may vary".
 
As I stated before, I did not intend on trying to benchmark "typical" usage
(as there really is no such thing). This is partly why I did not state the size
of the server pool, concurrency level or whatever. Basically the deal was, "tune
the client and server however you can for the highest throughput", to test the
absolute upper bound of the server.
 
This means, if you get a higher number with 10 concurrent connections, then
use that. If you get a higher number with 100 concurrent connections, then
use that instead. Tweak for best results.
 
If you have a hard time grasping the concept, then repeat after me, "upper
bound, not typical usage" ten times. If you feel the concept slipping from
your mind, repeat it a few more times. :-)
 
I did not do a concurrency vs. connections/sec breakdown although I imagine
it might be interesting to see how various servers perform as the concurrency
level gets really high.
 
> I can't speak for a high-end 3000 (the 960 is the best I have at the moment)
> but I suspect they could be competitive with OpenMarket compared to other
> platforms.  The 960 running OpenMarket does smoke the 386 WinHTTPD server,
> but our Suns running NCSA 1.5 (not ported to MPE due to BSD socket quirk that
> has not yet been fixed).  As I've said before, MPE's Posix doesn't fork()
> well at all, and that is what kills the freeware NCSA httpd port.  OpenMarket
> uses a server pool and posix threads and doesn't have to fork() for simple
> connections although it can be burdened by cgi applications.
 
Note that servers like Spinner don't server pool at all -- they multiplex
all connections through a single process, thus avoiding fork() and socket
passing overhead altogether (Spinner "spins" on a socket, hence the name :-)
This is partly why it "screams" compared to even commercial servers like
OpenMarket.
 
Note that you _can_ start multiple Spinner servers to take advantage of SMP,
but by default only one runs because that is generally all that is needed.
 
Spinner makes use of a lot of POSIX stuff that MPE does not support, therefore
spinner would be a tough port, although it would be worth it -- Spinner is
a _very_ nice web server.
 
For those interested in Spinner, visit http://spinner.infovav.se/
 
> Dan - can you supply the zbench package and actual benchmark pages on your
> web site so that those of us who can contribute numbers can obtain "equal
> ground" environment for comparison?
 
The zbench package is available through a link on the web page. For those
of you who missed it, it's http://www.anime.net/httpd_benchmarks.html
 
The hits/sec page was an ascii file of 1024 bytes (any will do) and the
throughput page was a binary file of 1048576 bytes (again, any will do).
 
-Dan