the bottleneck?

[Bohn Christopher A Capt AFRL/IFSD]

Good day,

> From: James Sasitorn [mailto:camus@rice.edu]
> Sent: Thursday, June 24, 1999 6:31 PM

> what is considered the most crucial aspect of a beowulf 
> cluster? I'm aware
> that in the past that there have been serious bottlenecks in 
> networking. How
> exactly does several 100Mbit ethernet cards compare to a 
> gigabit ethernet?
> What kind of bandwidth utilization occurs?

First, the easy answer: "it depends on your application."  In general,
though, the network is (and I dare say always will be) the bottleneck for
all but the most coarse-grained problems.  As for multiple fast ethernet
networkss vs gigabit ethernet, the "smart-guy" answer is that a gigabit
ethernet network is ten times faster than one fast ethernet network.  In
practice, I understand (but have no personal experience with gb ethernet or
channel-bonded fast ethernet) there are hardware limitations that would make
gb ethernet and two or three channel-bonded fast ethernet networks perform
similarly on a standard x86 system.

> Is it better to 
> get some decent
> celeron 300s or k6-2s or is it better to get PIIIs?

Depends.  On your application.  (sigh - It's starting to sound trite, but
it's true with so many things.)  I understand the K6's do not have as good
of a FP core as the P6's.  But it is cheaper, and if you don't have much FP
going on, that may be for you.  As for Celeron vs Pentium III, the big
differences are the speed of the system bus, the size of the L2 cache, the
speed of the L2 cache, and the price.  Do not Do Not DO NOT use a basic
Celeron 300 -- if you use a Celeron, get a Celeron 300a or 333 & up.  The
basic Celeron 300 & slower do not have an L2 cache >at all<.  The 300a and
333 & up have a 128KB full-speed cache; the Pentium III has a 512KB
half-speed cache.  All Celerons (to date) have a 66MHz system bus, while the
Pentium IIIs has a 100MHz system bus (133MHz allegedy by the end of the
year).  Depending on your memory access patterns, the difference in speed to
main memory may be significant, or it may not.  As for whether the L2 cache
for one or the other is more suitable, a few more factors come into play (if
you spend a lot of time dealing with chunks no larger than 128KB, then the
512KB won't do you much good, and the faster access to L2 will be a real
boon; for larger than that, you'll have to give it some analysis).

It also depends on what you want your system for.  If you want to do real
computational science/engineering investigations, then obviously you want to
build the fastest machine you can.  If you're doing education or research in
parallel/distributed applications, then being able to scale applications may
be more important to you than the actual speed, and you'd probably want to
buy the less expensive hardware so you can buy more of it.

> Also how much of a
> performance gain is there after a certain threshold of memory 
> is reached and
> for scsi harddrives.

As for memory, there's no way around it ... it truly depends on your
application.  How big is it?  Keeping it all in main memory is much better
than swapping to disk, even with SCSI drives.

Hope that helps a little.  Take care,
cb
*-*-*-*-*                                                     *-*-*-*-*
Christopher A. Bohn, Capt, USAF     || christopher.bohn@sn.wpafb.af.mil
Digital Simulation Systems Engineer || cbohn@computer.org
Collaborative Simulation Technology ||
          and Applications Branch   || v (937)255-4429x3576  (DSN785)
Information Directorate             || f (937)255-4511       (DSN785)
Wright Research Site                ||
Air Force Research Laboratory       || http://members.aol.com/EngrBohn/
         http://www.if.afrl.af.mil/div/IFS/IFSD/IFSD_home.html
*-*-*-*-*                                                     *-*-*-*-*