[nSLUG] no networking

Peter Cordes peter at llama.nslug.ns.ca
Fri Jan 9 13:40:52 AST 2004


On Tue, Jan 06, 2004 at 04:42:49PM -0400, Dop Ganger wrote:
> On Tue, 6 Jan 2004, Peter Cordes wrote:
> 
> >  For anything other than a number-crunching machine, HT is probably a good
> > thing.
> 
> HT's also a good thing for (some) number crunching apps. We did a test on
> the cluster I installed at Dal and found that one processor with HT
> enabled gave 80% of the performance of two processors without HT enabled.

 Interesting...  Maybe I'll run some benchmarks on a dual Xeon I have an
account on.

> If you're running different apps instead of one I suspect you won't get
> the performance due to cache issues.

 I would guess that total amount of memory traffic/cache use would be more
of a factor.  I guess running two copies of the same program would lead to
memory accesses at the same offsets within pages, but don't the caches
have enough associativity (L2 is 8-way associative, L1 Dcache is 4-way, see
www.sandpile.org) that this isn't a big problem.

> > Maybe there's an HT interface that the
> > kernel can use to interrogate the hardware, but laptop CPUs support it but
> > only have one logical CPU?
> 
> I think that's the situation. It's probably cheaper for Intel to have a
> mask that has everything enabled and picks out parts that are "good" for
> different apps, but any chip that doesn't cut the mustard as (say) a P4
> with HT can be reused as a P4 without HT. I remember that's what they did
> with 486 SX and DX, and it wouldn't surprise me if that were still the
> case...

 That seems a little unlikely: don't they have to make laptop chips on
purpose, since they have different speed-efficiency tradeoffs?

-- 
#define X(x,y) x##y
Peter Cordes ;  e-mail: X(peter at cor , des.ca)

"The gods confound the man who first found out how to distinguish the hours!
 Confound him, too, who in this place set up a sundial, to cut and hack
 my day so wretchedly into small pieces!" -- Plautus, 200 BC




More information about the nSLUG mailing list