[ale] mosix clusters?

[Dow Hurst]

On our SGI Origin 200 servers with two processors we can get better efficiency
by running three computational jobs instead of two. Â The total time is greater
than two jobs but less than three sequential jobs. Â Each job shows 68% of
the CPU while running. Â These are FPU based MonteCarlo based algorithms for
molecular conformational analysis.
Dow


Jeff Hubbs wrote:
<blockquote type="cite"
 cite="mid1024890033.2876.672.camel at localhost.atl.ipsvc.net">
  On Sun, 2002-06-23 at 23:12, Joseph A Knapka wrote:
  
  
    Jeff Hubbs wrote:
    
    
      To revisit a bit, you'd need to know if, on a given 1-CPU machine, two
SAH WUs run at once is faster than, slower than, or as fast as two WUs
run end-to-end in the typical fashion.  If it's slower, then there is
absolutely nothing to be gained by running more SAH WUs at once than you
have CPUs.  My experience suggests that it would be slower, implying
that one instance per CPU is OPTIMAL; I say this because on a 2-CPU
machine, both CPUs get saturated, which tells me that memory I/O is not
an issue.  One instance per CPU is no different than what I do on every
machine I can get my hands on, Mosix-less.
      
    
    Surely it is impossible for two WUs to run simultaneously on
a single-CPU box in less time than they would run serially,
since in the simultaneous case both WUs would presumably
always be runnable, and thus you'd have context-switch
overhead that wouldn't be present if single WU were the
only runnable process on the box.

    
  
  
Joe -

That's what I would assume as well.  I don't have the 'fu to know just
how much context-switching overhead that would entail.  I do wonder,
though, if there might be some "sweet spot" that takes advantage of the
way the processes use the CPU/mobo architecture, what with instruction
pipelining and L1/L2 cache.  


---
This message has been sent through the ALE general discussion list.
See http://www.ale.org/mailing-lists.shtml for more info. Problems should be 
sent to listmaster at ale dot org.