RE: Advice on a HUGE configuration: !!!+800 Hits/sec!!!.

From: Jordan Mendelson <>
Date: Sat, 25 Apr 1998 14:52:04 -0400

> > different vendors. Not sure who puts FreeBSD on hardware though.
> I'm not sure but alpha pc architecture doen't much differ from intel pc
> architecture.
> Would you give a bit more explanation about why alpha is
> better than intel
> pc ?

Oi, I don't want to get into this. Using PC's as servers has been the big
thing these past few years, but before this.. no one in their right mind
would even consider putting mission critical services on a PC.

First off, the Alpha PC line is not the same as the Alpha server line. They
use two different CPUs. The Alpha PC has a stripped down L2 and L3 cache
which lowers performance.

However, Alpha's (21164 and soon to be 21264) have in general huge CPU
cache, which help dramatically in servers. They also have 64 bit wide PCI
buses. Multiple PCI buses on the same motherboard is also not uncommon in
Alphas (so you have > 5 PCI slots). Multiple Alpha processors also tend to
give higher performance than multiple Intel processors (I've heard horror
stories about 3-4 Intel processors in a single machine).

Also note that the Alpha being put out by Samsung is a 21264. It has up to 5
GB bandwidth for L2 cache and the ability to swap the CPU with an AMD. The
initial chip KP21264-2.0X delivers about twice the Integer performance of
the 500MHz KP21164 processor or Intel's Pentium-II 333MHz processor. Of
course, the high end chip the KP21264-3.5X puts out about 5x the performance
of a P2 333Mhz processor. These are just the ones being put out by Samsung,
Alpha technology has been licensed to quite a few other companies as well.

Oh, and one more thing... if you ever plan on using > 4096 file descriptors
under linux under one process, don't use an Intel machine. The same is
probably true for FreeBSD.

I always have a hard time figuring out the number of fd's required for Squid
however. One for each incoming socket from the client, one for each outgoing
socket to the server which the client requested, and one for the file being
saved to disk? So you'd roughly need to multiply incoming connections * 3?
If you figure each connection is going to take on average 4 seconds, I would
guess that 800 conns/sec would use up 9600 filedescriptors. I could be wrong
here though... I'll leave that to the Squid experts :)

In the end, I guess it's a matter of perspective, you can buy a machine made
to be a server, or you can buy a PC and try to make it a server. It also
depends on how much money you want to spend. Is an extra few thousand worth
the performance?

Geez, I sound like a salesman now.


Jordan Mendelson     :
Web Services, Inc.   :
Received on Sat Apr 25 1998 - 12:00:16 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:39:57 MST