[Beowulf] Not all cores are created equal

Prentice Bisbal prentice at ias.edu
Tue Dec 23 06:41:14 PST 2008


John Hearns wrote:
> I'm surprised this has not been flagged up yet. Shamelessly passed on
> from Slashdot:
> http://www.gcn.com/online/vol1_no1/47765-1.html

John,

Thanks for posting. I just read it.

<unsolicited opinion>
The problems the researchers cite are nothing new. Those issues have
been around, and written about, for years the only difference is that
the processors are now on the same die.

Regarding the issue handling interrupts and how that can affect
performance is discussed in this paper, which determined that in
multiprocessors systems, it's best to let one processor do all the
interrupt handling (and nothing else)

http://hpc.pnl.gov/people/fabrizio/papers/sc03_noise.pdf

The other problem they mention, when data one processor needs is in
another processor's cache, and how that affects performance is nothing
new, either. That's why we have processor affinity. I learned about that
in my computer architecture class years ago - before multi-core
processors existed.

</unsolicited opinion>

The above is based on the GCN article, not the actual paper it refers
to. I haven't read all 12 pages of it yet, but I did skim through it.
It's interesting to note that the research was conducted only on the
Intel Architecture.

They mention that the AMD NUMA architecture is more complicated and that
they hope to do future research on it. Not including the AMDs in this
research lame. Showing the difference in performance between
architectures - *that* would be meaningful information that we could all
use. But I guess that leaves the authors an opportunity to publish an
additional paper on the same topic, so they can list two publications on
their CVs instead of one.

-- 
Prentice



More information about the Beowulf mailing list