Subject: (C413) JGSI Review
Resent-Date: Thu, 30 Sep 1999 23:16:54 -0400
Resent-From: Geoffrey Fox <gcf@npac.syr.edu>
Resent-To: p_gcf@npac.syr.edu
Date: Sun, 12 Sep 1999 14:20:44 -0700
From: Matt Welsh <mdw@cs.berkeley.edu>
To: Geoffrey Fox <gcf@npac.syr.edu>

Paper number: C413
Title: Locality optimization in JavaParty by means of static type analysis
Authors: Michael Philippsen and Bernhard Haumacher

Overall comments:

This is an excellent paper, very well-written, and quite detailed. I don't
have any comments which suggest ommissions or modifications to the existing
work.

However, I am sure that the type-analysis optimizer has come some way since
the EuroPar workshop in 1998, and would like to see a new section added to
the end of the paper discussing current status, implementation details, and
performance. This paper sets the stage by describing in detail the
algorithms
used and the basic technique for cloning methods, however, it does not
adequately cover the myriad implementation details which are so vital for
this approach to be incorporated into the JavaParty system as described.

Performance is a serious concern. Considering that cloning methods requires
the addition of new parameters and a lookup table per method, I am curious
as to what the overhead of this approach is (both in terms of CPU time as
well as memory usage). I am also interested in whether the authors have
considered compiler optimizations for this approach -- i.e. whether the
standard JIT compilers could be extended to understand this lookup table and
generate efficient code for it.

Benchmark results would be another important addition to this paper. I would
like to see some simple concurrent applications and how well they perform
with and without this type-inference optimization. While it is clear that
this
kind of optimization has the potential to improve concurrent performance in
the kind of applications which JavaParty targets, it is important to
evaluate
the bottom line to see if this was a success!

Apart from some additional discussion of results and performance numbers,
I have no other suggestions for improvement!