Newsgroups: comp.parallel.pvm
From: john@elroy.jpl.nasa.gov (John Wright)
Subject: Re: Definition of Granularity
Organization: Image Analysis Systems Group, JPL
Date: 17 Nov 1994 15:38:10 GMT
Message-ID: <3aftd2$i2o@elroy.jpl.nasa.gov>

In article <3ad6q8$qp4@muller.loria.fr>,
Frederic Montoya <montoya@loria.fr> wrote:
>In article <3a7572$hmq@fstgds15.tu-graz.ac.at>, goller@fcggsg32 (Goller Alois) writes:
>|> D. Subramani (Raj.Subramani@ncl.ac.uk) wrote:
>|> : Could some kind soul please post/e-mail me a clear
>|> : and concise definition of Granularity (or its 
>|> : reference). I have yet to see one (clear and concise,
>|> : that is) from the six textbooks I have so far
>|> : reffered to.
>|> 
>|> : Mucho gracias.
>|> 
>|> : Cheers
>|> 
>|> : -raj.subramani@newcastle.ac.uk
>|> 
>|> A coarse grained algorithm is generally easier to parallelize
>|> than a fine grained one :-)
>|> 
>|> Seriously: As I looked for the solution of this question, I
>|> found many answers from many authors but most of all were
>|> different from each other. Who is authorized to give such 
>|> a definition?
>
>Hi;
>
>I was thinking that the ratio computation time over communication time
>was a commonly accepted definition (and obviously precise).
>I would be interested to know about different ones.
>
>Regards;
>Frederic.
>
I've always used the approach of taking the serial execution of the
algorithm, unfolding it so the entire problem is a linear series of
instructions, and then identifying the processors on which each
block of code executes.  The ratio of the size of each processor's
chunk to the total size is the granularity.  It works for me.  The
simple example is generating a number of frames for an animation.
The serial process generates a frame at a time, a line at a time, a
pixel at a time, a calculation at a time.  Puttine each frame on a
processor is coarse grain, each line might be medium grain, each pixel
would be fine grain, and each calculation (a pipeline perhaps) would
be extremely fine grain.

I don't like using the computation/communication time ratio because
it's dependent on the system as much as the algorithm but I don't
claim my way is best or objective or anything else.  I just like it.

TTFN (that's Ta Ta For Now)
John



