httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jason S. Clary" <>
Subject Re: ZD Net tests, IRIX, and MAX_SOCKETS
Date Mon, 20 Jan 1997 17:30:28 GMT
> In the immortal words of Randy Terbush who said, "I drank what?":
> Is this benchmark package available anywhere?
err.. yes, on  I saw yesterday.. the controller runs on NT, the
run on nt, 95, and mac...  (it controlls multiple machines to hit one server)

Looks like its free..  The ZDNet test done for the article used a 100mb local
with 60 Win95 clients and an NT server as the controller...

here's the addy...

There's also a full system benchmark package...  I think its only for windows
mac, though...  

btw, even though under these benchmarks the Cyrix 686 processors that I use
are on average 23% faster than intel versions of the same P rating (the Cyrix
is actualy a 133mhz processor, but its 23% faster than the average Intel P166
... but, this didn't explain the speed difference I was seeing between my
server and
a SunOS based web server (NCSA) running on a Sparc20 with 256meg of ram..

Further checking revealed the reason behind the difference noted (mine ran
twice as fast) was because of the MAX_SOCKET setting in its kernel.  It was
to 512 whereas mine was set to 1024.

I had (wrongly?) assumed it was general knowledge that you need up the
MAX_SOCKET on most unix systems in order to make your web server perform for
high volume.

The newest linux kernels even have a compile time option to optimize as a
server in which case it bumps the MAX_SOCKET setting up considerably from the
standard 256 setting it gets when you don't optimize for network.

It makes the kernel use more memory during the run life, of course, as it
allocates the
memory for the fd's at boot time and doesn't ever free it, but it doesn't make
difference on a machine thats got 128meg+.. ;)


View raw message