Microcontrollers/ASICs v. General Purpose Microprocessors -- WAS: Opteron Vs. Athlon X2
Mark Hahn
hahn at physics.mcmaster.ca
Sat Dec 10 05:08:34 UTC 2005
> > I'm personally not running a cluster, but if you look into the
> > Beowulf list archives, people are citing impressive numbers.
>
> Beowulf performance is _inversely_proportional_ to communication load.
that's either tautological or wrong.
> Applications scale far more linearly when the slower 10GbE are not used
> than the local, HyperTransport interconnects (assuming Opteron). From
oranges are easier to peel than apples. what was the question?
10GE is a different category of thing than HT.
> what I've seen, no one is getting anywhere close to 1GBps over layer-2
> communication with 10GbE for clustering. I'm talking about when you're
myri 10G does.
> because 1GBps won't cut it. Beowulf is not even applicable!
you have some weird ideas of what Beowulf is.
> Furthermore, the new, preferred solution is to use HTX (HyperTransport
> eXtension) Infinibind cards, which will give you about 1.8GBps real-
are you trying to spell "InfiniPath"?
> (in raw transfer terms) it's faster than GbE. But even then, you don't
> use Beowulf for applications where you're taxiing the interconnect
> (database, file, etc...).
afaikt you're claiming that beowulf is low-bandwidth. that's just weird!
More information about the amd64-list
mailing list