Re: an idea about dual cpus

From: Darren Steven <dsteven@dont-contact.us>
Date: Wed, 17 Nov 1999 09:39:35 +1100

Dustin Byford wrote:

> Alex Rousskov wrote:
> >
> > On Mon, 15 Nov 1999, Dustin Byford wrote:
> >
> > > Our setup is currently consuming 80% of the cpu (average) with regular
> > > peaks to 100%. The machine is not io or network bound. So in an effort
> > > to fix the situation I was thinking about using smp. What does everyone
> > > think about this.
> > >
> > > Run 2 squid processes under different config files/log files/etc. Let
> > > them use icp to talk to each other through loopback. Give each process
> > > 2 of the 4 disks. This way each squid process will get scheduled on 1
> > > cpu. Of course the overhead of icp may be to inefficient. What do you
> > > think?
> >
> > a) Are you already using dual CPUs? I would not be surprised if a
> > single Squid installation benefits from using smp.
>
> No this is a single PII-400 over clocked to 500 MHz, 125 MHz bus, 41 MHz
> PCI, 1G RAM, 4 18G u2w drives.
>
> RAM, network, and disk are all doing very well. It's just the CPU
> that's holding us back. Currently it looks like we will get around 3M
> hits a day. Soon this should go up as we double our outgoing
> connection. First of all does this sound normal? Should I track down
> what is using all the CPU and simply quit doing it?
>

We have noticed that disk IO transactions per second is a problem. We had a box
with PIII500, 768M ram and 4x9G UW scsi, with about 1500-2000 http req per
minute. The total disk IO was on the order of 1MByte per second, way below the
expect thruput of the SCSI bus, or the 100M ethernet. However, individual disks
were sustaining up to 120 requests per second, which can be a problem if they
are random IO, not sequential - they become seek bound. I turns out that the
large number of IO's was due to raid (which we were experimenting with - and no
longer use). We still see large transaction rates and associated seek latencies
from time to time, but never the thrashing that we used to get - the machine
sounded like a washing machine when all disks were seeking themselves to death.

Darren Steven
Applications Specialist
Networking Tasmania
Telstra Australia
Received on Tue Nov 16 1999 - 15:52:04 MST

This archive was generated by hypermail pre-2.1.9 : Wed Apr 09 2008 - 11:57:32 MDT