[Beowulf] Re: computing on Altix? (Andrew Piskorski)

[Isaac Dooley]

>>I submit a task to Altix 350 share memory parallel machine.  The
>>task is to compute a graph, and rquests 6, 9, 12, 16 processors
>>respectively. The results show that the running time is decreasing
>>with the rise of the size of processors for same graph, while each
>>processor runs same program, and communicates with fixed number of
>>neighbours periodically. why?
>>    
>>
>
>Google for "superlinear speedup".  Most likely, as you split up your
>fixed problem size among more processors, more and more of it fits
>into the processor cache, where it runs much faster due to fewer main
>memory accesses.
>

This cache effect is quite profound on Altix since some of these have 
something like 9 MB cache per processor. You can see this result on 
NAMD. http://www.ks.uiuc.edu/Research/namd/performance.html.

An additional cause for superlinear speedup, which I have been seeing in 
an application recently is that the serial case and also therefore the 
parallel case is inefficient. Thus you may have some room for improving 
performance, even though your speedup numbers look good. Basically by 
decomposing the problem more and more, you may shrink some portion of 
the computation which might have been O(N^2).

Isaac Dooley