[daip] Y2K (the new DDT) - how quickly they forget

[Patrick P. Murphy]

On Tue, 09 May 2000 11:31:27 -0400, Ketan Desai <ketan at rentec.com> said:

(in response to Eric's message):

EG> A number of questions arise about the new DDT:

EG> 1. What should the master architecture be?  The official lies were
EG>     computed on Solaris for DDT93.

KD> I would say still a Sun machine.  There are too many variants of Linux
KD> machines to pick one as the master version.

But Intel/Linux will always have the same floating point behaviour, and
it's by far the most dominant platform in the Linux world.  There should
be no difference in F.P. behaviour between different flavors, e.g. Red
Hat, SuSE, Mandrake, Debian, etc. as long as they use the same libraries
(glibc) and hardware (Intel).

I think we should nominate two masters in the testing phase of the DDT2K:
one on Solaris/Ultra, a second on Linux/Intel; then decide on comparing
these which should be the "standard".  Chris can probably comment on the
wisdom of this vis-a-vis compliance with IEEE 754 (?) floating point
standards.

EG>  2. What machines should I test here?  I plan to run DDT large and
EG>     then medium and large Y2K.

In Socorro, you'd want to try and test "legacy" Linux systems like maruti,
(though without hdparm tuning, the results will almost certainly be
disappointing; I'm still convinced this is very much worth doing), a newer
Linux box if you can get it for the appropriate few hours exclusively, and
a smattering of Solaris systems.  You might also want to test ohsumi.

In CV, vulcan is of course a good candidate (PIII Xeon 550MHz).  Astral,
Tony's HP system, is now accessible but I've not shoehorned it into our
AIPS environment yet.  It has g77/gcc but not a HP supplied f77 compiler.

EG> 3. Ketin's machine has done DDT in 341 sec - should I scale the new
EG>    large to be ~10 X that?  More?

I think about an hour on modern hardware is in the right ballpark.
Whether that system, now getting on 1-2 years old, is "modern" is up for
debate of course!

KD> Is a target of an hour for the large DDT too short ?

Given the continued rate of improvement of processor speeds, disk access,
etc., maybe not.  But it's the right order of magnitude.  Anything longer
than a few hours is too much for now.

KD> To be really fancy, produce a DDT that automatically increases the
KD> size of the problem as a function of time in expectation of Moore's
KD> law.

I like this :-) but it would produce mass confusion among AIPSMark
consumers... 

				- Pat
-- 
  Patrick P. Murphy, Ph.D.            Division Head, Charlottesville Computing
  (804) 296-0372, 296-0236                National Radio Astronomy Observatory
  Home: http://www.goof.com/~pmurphy/   Work: http://www.cv.nrao.edu/~pmurphy/
   "Linux is Inevitable."  "Why?"  "Because it's alive!" - John MadDog Hall