Re: MPI octave with hyperthreading

[Przemek Klosowski]

   > lambooting 2 nodes. ... enabling HT and lambooting 4 nodes often gives 
   > a better than 2X speedup. 

   Ouch! I got too dogmatical on that :-) Yup, I forgot this situation of 
   codes not requiring too much memory access, at least in comparison to 
   the amount of computation performed. I'm glad you can correct me on 

Well, I don't want to be a stickler for detail, but if the code is
completely compute-bound, it won't get any speedup on hyperthreading,
either, because there would be no stalls to use for the other thread;
remember that the HT CPU doesn't have two sets of arithmetic
units---just two sets of working registers.

I think your misunderstanding is caused by an assumption that all
threads on a SMP run ideally synchronized, banging on the same memory
locations.  If that were true, you would indeed have heavy contention,
and no speedup---but it will NOT be true. For starters, even if the
instruction streams were synchronized, they would touch on _different_
data, e.g.  on separate regions of the work array ('striping').
Secondly, in the absence of tight synchronization in the code, they
will get out of sync because of small variations in how they are
scheduled.



-------------------------------------------------------------
Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:  http://www.octave.org
How to fund new projects:  http://www.octave.org/funding.html
Subscription information:  http://www.octave.org/archive.html
-------------------------------------------------------------