[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    Re: First impressions of the famed C450A
From:       Kristian Soerensen <elof () image ! dk>
Date:       1999-06-21 23:20:27
[Download RAW message or body]


On Sun, 20 Jun 1999 getrpml@terrorist.math.ntu.edu.tw wrote:

> 	Note-- at bootup, I noticed how sluggish response was; Using a
> pair of PP200/512s and the memory-interleaved P6DOF (about equivalent
> to using a LX dual PII/233 with old SDRAM timings, or so I am told), I
> got very accustomed to quick response times way back.  The small size
> of the cache apparently really hurts.  I had originally planned to put
> the older PII/333's that I inherited into a toy box.  Now I will stay
> with them, and put the Celerons into the toy box instead.
> 
> 	This is my first attempt to do a `real' overclock.  I'm almost
> convinced that overclocked Celerons really isn't all that hot now and
> there *IS* a reason to making larger caches, even slower ones. 
> 
> 	Any comments?  Anyone with different experiences?  

Cache misses are the enemy #1 these days for most Linux-worthy tasks.
It doesn't really matter how quick the CPU is if it spends a lot of it's
time sitting idle while waiting for data or code to be fetched from main
memory or the slowest of the caches due to frequest cache misses.

Actually this has been the most interesting difference between PC's and
workstations for the last many years. The workstations has had way better
memory bandwith then the PC's, only with the best of the most recent
hardware have the PC's caught up with low end workstations.


The STREAM benchmark tries to measure this memory bandwith problem.
http://www.cs.virginia.edu/stream/analyses.html

I have found the runtimes of non-trivial Java applications and 3D
rendering tasks to map rather lineary to the STREAM benchmark numbers for
the systems I have access to. Only for jobs that's not operating on vary
large pieces of memory at a time does the more CPU oriented benchmarks
make any sort of sense.



Gone are the days when we counted CPU instructions while building data
structures and thinking up algorithms, these days it's memory bandwith
that's the limited resource and therefore flyweights are kings.



--------------------------------------------------------------
The Linux Resource Exchange             http://www.linuxrx.com
Kristian Elof Soerensen    elof@image.dk     (+45) 45 93 92 02 

-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to majordomo@vger.rutgers.edu

[prev in list] [next in list] [prev in thread] [next in thread]