[Beowulf] Re: Beowulf Digest, Vol 10, Issue 16

[Robert G. Brown]

On Tue, 14 Dec 2004, David Mathog wrote:

> Very unclear to me what, if anything is gained by using 
> registered ECC vs. unbuffered ECC on a smallish system.  On a big
> system, with lots of memory, I can see where registered ECC would
> be a plus.  But if the system only holds 1Gb of RAM in one
> or two memory slots the unbuffered memory should be slightly
> faster, and with the ECC enabled, just as reliable.  Correct?

I hope that is rhetorical instead of asking >>me<< in particular.  All I
know on the issue is from what people like Don Becker on list have
written about it over the years, and that's probably all muddled up.  To
my own personal experience, buffered/unbuffered ECC/nonECC memory all
pretty much works perfectly as long as one a) doesn't overclock; b)
doesn't overheat; c) aren't in the process of physically failing even
without overclocking or overheating.  At least, if I've dropped bits
over the years, I have little overt evidence of it, or else it seems to
occur just as often on systems with ECC memory (like those good old
2466's) as it does on my many cheaper systems without it.  More often,
even.

I vaguely recall Don saying something about ECC being essentially
redundant with checks that take place anyway and just slowing you down,
but it was a lot time ago and I also recall various also smart persons
arguing that ECC was essential (and offering an example of a cluster
where without it they would throw an error a week or something like
that.  Maybe Josip?  So I plead ignorance on the technical front and
inadequate personal/anecdotal experience.

   rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu