Re: Numerical Computation and Ada95

[stift@fedelma.astro.univie.ac.at (Martin Stift)]

>From what you said I understand that nobody has ever compared Ada's timing
>performance with Fortran77/90/95, C and C++ in programs that require lots
>of arithmetic and logical operations (perhaps FFT).  I think I will try to
>do it when I learn Ada95 better

>The issue of performance is one of my biggest worries when choosing
>Ada95 over other languages, when speed is critical.  Comments elsewhere
>in this thread suggest that this is not normally a problem, but I
>have found that great care is needed to avoid inocuous constructs which
>result in slow code.  Sometimes this prevents writing the code in
>the most "natural" way.  Things may be better now with the latest
>GNAT (I developed most of my performance-critical code under GNAT 3.11, 3.12).
>Of course, you need to take great care in C or C++ too.  Fortran(77) seemed
>to be a lot more robust in performance terms, perhaps because it was
>quite limited in what you could do.  Try some simple example comparisons
>in your proposed environments.  They may not be representative, but
>at least you'll have some idea what to expect with "real" code.


Well, I have had the chance to compare the numerical performance of Ada83
and of Ada95 with Fortran77 with quite satisfactory results.

Back in 1994 I had a program written in Fortran77 which carried out spectrum
synthesis in magnetic stars. This program had been translated into Ada83 with
considerable restructuring of the code in order to take advantage of the superior
expressive power of Ada83. The code did mostly number-crunching.

Running the respective Fortran and Ada83 codes on a DEC Alpha under
OpenVMS resulted in the Ada code being about 5% slower than the Fortran
code. At that time I had 20 years experience in Fortran and 1 year experience
in Ada83.

In 1997 I definitely switched to Ada95, working on a Silicon Graphics Origin200
4-processor server. In the beginning the GNAT compiler running under IRIX
had to use the old and slow o32 ABIs whereas with Fortran one could use the
much faster new n32 ABIs. In some extreme cases, execution times could differ
by as much as a factor of 3. However, when using the o32 ABIs for both Fortran
and Ada95, performance was about the same to within 10-20%.

The situation has greatly improved since. Although Sgi is known for its advanced
Fortran compiler technology, the GNAT compiler by now performs impressively.
The famous "abstraction penalty" can be encountered both in Fortran and in Ada95.
The calculation of the Voigt-Faraday function, which is the most time-consuming
part of my code, in Ada95 takes about the same time (within 5-10%) as with Fortran.

I made another test with a celestial mechanics program which I translated directly
into Ada without any substantial modifications to accomodate Ada features. Again,
the program just does number-crunching, extensively using more-dimensional
arrays which should be treated very effectively in Fortran. Still, with GNAT 3.13,
the Ada95 version was  f a s t e r  than the Fortran77 version.

One of the major advantages of supercomputing with Ada95 is the fact that it
provides a thread-parallel model. You don't have to resort the MPI but can
parallelise your code with kernel language constructs. Protected objects have very
little overhead and are perfectly suitable for synchronisation. Tasks also are
easy to handle. At present, I am running my codes on an Origin2000, using up to
64 processors (Amdahl's law makes it impracticable to use more processors for
the problems at hand).

Summarising I can say that Ada95 is a viable alternative to Fortran. It is fast and
safe, the parallel constructs are easy to use. More information on supercomputing
with Ada95 can be found in

Stift, M.J., 1998.
Astrophysical Supercomputing: Ada95 as a safe, object oriented
alternative.
In Reliable Software Technologies --Ada Europe '98,
ed. L. Asplund, Lecture Notes in Computer Science, Vol. 1411, p. 128-139.

Stift, M.J., 1998.
Scientific programming with Ada95: object oriented, parallel and safe.
Computers in Physics, Vol. 12, No. 2, p. 150-158.


These publication give some ideas as to how the Ada95 features can be applied
to (astro)physical problems. Synchronisation as presented in 1998 is somewhat
outdated now, the solutions presented in the above-mentioned papers are only
useful on small machines like my Origin200. They have in the meantime been
replaced by much more efficient algorithms that work on large supercomputers.

Martin Stift


-- 
Posted from fedelma.astro.univie.ac.at [131.130.36.112] 
via Mailgate.ORG Server - http://www.Mailgate.ORG