definition of etime(3F)

[uunet!cwi.nl!Dik.Winter]

 >Do you use etime for timing in Fortran?   It's been a part of f77-derived
 >Fortrans from at least 4.2 BSD, I think.   A currently contentious question
 >is how it should be defined on systems with multiple processors.

It is a problem indeed and there is not a good single solution.  For
accounting purposes obviously total user+system time is needed.  On
the other hand, for benchmarking parallel programs elapsed time is
needed (although not wall clock time, that is a different time again).

If you provide total time only you can expect some astonishing benchmark
results in future papers.  (I know of papers that mention on some 4
processor system a speedup by a factor of about 3.9.  Even on a dedicated
system from that manufacturer a factor of 3 is the best that can be
obtained.  The writers of those articles just take total time and
divide that by 4 "because we used 4 processors".  Not only one paper,
many papers have that.)

Cray provides some useful routines in addition to an etime equivalent.
One of those will fill an array with the first element the time the
program was running on one processor, etc.  This is very good for the
analysis of parallel programs.

dik
--
dik t. winter, cwi, kruislaan 413, 1098 sj  amsterdam, nederland
home: bovenover 215, 1025 jn  amsterdam, nederland; e-mail: dikacwi.nl