ISO/IEC JTC1/SC22/WG5-N1689

                 Letter to the Convener from John Wallin

Hi John,

I just got a read of the British Standards Institution response to the
Fortran 08 draft. I am probably preaching to the choir, but here are a few
comments.

I am very disturbed to see the suggestions about slipping the schedule two
years and/or dropping major features in the language. In particular, I am
concerned about dropping co-arrays. (I am not doing to talk about the macros
or BITS features.) Although this is a new feature, it is coming at a
critical time for the language. With Intel creating 80-core chips scheduled
for release in the next five years, we need a way to take advantage of this
new computing power. Adding on to this, the speed of network connections
between boxes is making the walls between CPUs less problematic. In five
years, it would be surprising if many machines don't have two to four CPU
chips, each with a few dozen (or more cores). The target for our audience
will probably be higher end versions of the machines for scientific use -
with perhaps several boxes tied together as well. The machines will probably
be non-homogeneous and hierarchical, particularly in terms of memory access
times. As a data point, it is useful to note that my mid-end current home
machine has 96 floating point pipelines in its graphics card and a dual core
CPU. I believe it is critical to realize that parallel computing is here now
for commodity PCs, and this isn't some concept that is going to happen in
the distant future.

These hardware changes are going to affect users at the desktop level
tremendously, as programmers struggle to optimize their codes for this new
hardware. I think the biggest mistake in the document is that "co-arrays
would be relevant to only a minority of users." I don't believe this will be
true in five or ten years for anyone who wants to get the performance out of
their machine they need to solve scientific problems, especially given the
current trends in hardware evolution. Right now, as you know, the options
for doing parallel computing are fairly limited. You can use HPF extensions,
but it limits the types of codes that you can program. The same is true with
OpenMP. For many complex problems, MPI is the only workable option
especially if you are crossing over a network to access remote data.  MPI is
a very flexible, general, and yet horrific way to do work on complex
parallel code. It is very difficult to write and debug code using MPI, and
the book keeping aspects tend to dominate the code.

I teach a course in high performance/parallel computing. Seeing how students
respond to MPI and the other paradigms out there (threads, sockets, openMP,
etc), I believe there is a huge need to have a language that can express
parallel concepts easily. Co-arrays fit very nicely into what people need to
do. It isn't perfectly intuitive, but it is infinitely better than most of
the other options around AND it fits the array model of Fortran perfectly.
It is also clear that the demand for this type of paradigm is going to grow
greatly over the next few years as hardware changes. We haven't barely begun
to address these issues in the Computational/Computer Science curriculum,
but I suspect industry is going to start demanding our students know how to
take advantage of the hardware that is coming. Having something we can teach
them that is straightforward would greatly help make this possible.

I also should note that the argument that we need more experience with
Fortran 03 doesn't quite make sense to me. Co-arrays really only address
issues that were in the 95 draft and are largely independent of the
revisions to 03. Yes - the feature does add complexity to the language, but
it doesn't interfere or interact with 03 in a deep way.

In summary, I think the hardware on commodity PCs is going to force us to
use parallel programming constructs on any code we want to run efficiently. 
The current paradigms are not well suited for the task in many types of
complex codes. If we don't have constructs available to make this work, the
language will be in serious trouble. Fortran 08's co-arrays are critical to
taking advantage of the new hardware that is on the way.

Well, that's my two cents. I hope all is going well for you and hope to see
you at the November meeting.

-John Wallin