eugene@eos.arc.nasa.gov (Eugene Miya) (12/07/88)
On the subject of superlinear speed up: the terminology in the field of
computing is bad enough (not unique to parallel processing, witness
security in recent days), that I would like to make a modest suggestion.
I unfortunately did not see the original posting on superlinear speed up
and I am aware of certain companies who make this claim without full knowledge
of what they are doing.
As yet another person with less than full knowledge, I am aware of the
concept of superlinear EXECUTION [as noted by Jones's paper in 1982?
Computer]. The question isn't linear as much it is unitary. If I add
one processor, I get the added benefits of one processor, ad nauseum
linearly. We can still have "linear" but sub unitary performance increase
[ASCII terminals are terrible for illustrating this type of graphic].
So the simple suggestion is to replace "linear" with "unitary."
In the case of "Super-linear" speed up I propose we have two types:
linear (which is what is typically reported) and non-linear (quite
desireable if it can be achieved). So we have this hierarchy:
Nonlinear Super unitary speed up
Linear Super unitary speed up
Unitary speed up (I call this "numerical breakeven").
Sublinear unitary Speed up
Log speed (Minsky 1970 Turing Limit)
It will be a measure of the influence of the net if we can change the
way people think about this terminology (I personally am skeptical,
but who am I). I also think by splitting a few hairs like this might
help us unstand what's going on.
Another gross generalization from
--eugene miya, NASA Ames Research Center, eugene@aurora.arc.nasa.gov
resident cynic at the Rock of Ages Home for Retired Hackers:
"Mailers?! HA!", "If my mail does not reach you, please accept my apology."
{uunet,hplabs,ncar,decwrl,allegra,tektronix}!ames!aurora!eugene
"Send mail, avoid follow-ups. If enough, I'll summarize."