From chemistry-request@ccl.net Tue Dec 10 21:41:53 1991
Date: Tue, 10 Dec 91 21:20:49 -0500
From: fredvc@esvax.dnet.dupont.com
To: %chem@ccl.net
Subject: About those workstations....
Status: RO

  RE:	Attack of the killer workstations 
From:   Joe Leonard
        jle@world.std.com


>>Reviewing the recent discussion regarding C vs. Fortran and structured vs.
>>spagetti coding, isn't it reasonable to consider the time of the programmer
>>as more important that the code's runtime (within reason)?  Given that
>>approx. $20k workstations are comparable to Cray machines, isn't 10-ish%
>>overhead more than worth the human time savings through better-thought-
>>out programming (structured, modular, OO, etc.)?  I hope the day's past
>>for really needing to count cycles in loops - as one can wait for the next
>>version of workstation xxx to gain appreciable speedups.

We have found it to be a mistake to depend upon workstation performance
(MIPS/MFLOPS), per se.  There is a lot that goes into getting good perform-
ance out of a box besides processor speed.  This becomes painfully clear as
the problem gets large.  Caching is commonly used in such systems, and this
can quickly become a performance bottleneck.  What you pay for in getting a
CRAY is a balanced system in terms of cpu, memory access, and disk access 
speeds.

>>Obviously, this isn't the case for mpp machines, for which the consensus is
>>that one must rethink and/or reimplement algorithm's in order to use
>>the hardware to advantage.  Still, this might prove to be an example where
>>clearer coding styles might turn out to be a win after all...


Good code runs better than bad code on ANY machine.  The "edge" tends to be
bigger on a CRAY, but the edge is still there on a VAX, SUN, IBM 6000, etc.
We recently optimized a program for the CRAY and found that the optimized
version ran better on the ALL machines we tested.

		CRAY speed up: ~63x
		 VAX speed up: ~17
	    IBM 6000 speed up: ~20

Mearly putting the original code on a box with a faster cpu would not give
this kind of payback.  There was nothing "pathological" about the program.  
It just needed some TLC!!  

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
                           FREDERIC A. VAN-CATLEDGE

Scientific Computing Division         ||   Office: (302) 695-1187
Central Research & Development Dept.  ||      FAX: (302) 695-9658
The Du Pont company                   ||
P. O. Box 80320                       || Internet: fredvc@esvax.dnet.dupont.com
Wiilmington DE 19880-0320             || 
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++