More than Just slightly off topic

[Bohn, Christopher A.]

On 9/17/98 10:52 AM, Bob Glamm [SMTP:glamm@ece.umn.edu] wrote:

[...]
> sometimes cache effects will allow more than linear speedup (i.e. 
> 10x CPUs = 10x the cache of one machine).

It isn't uncommon for large problems that have to be swapped to & from
secondary storage on one processor to see phenomenal speedup when the
problem can finally also fit in the aggregate main memory.  In these cases,
though, many authors base speedup calculations on the fewest number of
processors for which the problem fits in-core.

[...]
> IMHO: anyone doing any parallel work should know Amdahl's law backwards
> and forwards; it is _fundamental_ to performance evaluations and comes
> up repeatedly.

In fact, I'd say that anyone interested in ANY computer work should be
familiar with the more general application of Amdahl's law, namely that the
performance gain by any optimization is limited by how much the optimization
is used.  This is why, for example, a 200MHz Pentium system is not twice as
fast as a 100MHz Pentium system:  the processor may be twice as fast, but
the memory isn't.  Same concept applies when improving software.

Another speedup metric I've seen, besides Amdahl's fixed-load speedup, are
Gustavson's fixed-time speedup, which measures how much larger your problem
size can grow and still run in the same amount of time, as you add more
machines -- Gustavson's law provides considerably more optimistic (too
optimistic?) results than Amdahl's law.  And then there's memory-bounded
speedup, in which you grow the problem size to fit in the aggregate main
memory, without concern for execution time.

MTC

take care,
cb
*-*-*-*-*-*-*-* 
Capt Christopher A. Bohn 
Graduate Student, Electrical (digital) Engineering 
Air Force Institute of Technology     Phone (937)255-3636 (DSN 785) 
AFIT/EN638                              Lab x4606   Voicemail x6638 
2950 P St, Box 4638                         email cbohn@afit.af.mil 
Wright-Patterson AFB OH 45433-7765                 EngrBohn@aol.com 
               http://members.aol.com/EngrBohn/ 
*-*-*-*-*-*-*-*