Paper Number Cnnn: C532
Date: September 25, 2001
Paper Title: NetBuild: Transparent Cross-Platform Access to Computational 
Software Libraries 
Author(s): Keith Moore and Jack Dongarra

Referee Recommendation: Reject

Referee Comments (For Editor Only)

This paper is a description of a facility (NetBuild) that, while doing useful 
things, resides at almost the very bottom of the grid hierarchy. As such it 
doesn't qualify to be in a special issue on "Grid Computing Environments" 
whose emphasis should be on high-level, problem-oriented, application-specific
environments. NetBuild can definitely be part of a successful GCE but the
existing description reveals that it is little more than a glorified IMakefile
with a macro-processor. 

Referee Comments (For Author and Editor)

NetBuild builds on the successful repertoire of tools that have been developed
at the University of Tennessee. It automates many of the tedious details that
are involved in remote access of software libraries. 

However, the presentation makes it clear that NetBuild is little more than
an IMakefile system that is grid-aware. Some of the goals of this project
(e.g., "automate the process of selecting ...  software libraries over the 
Internet", "automate the construction of such libraries") are quite lofty.
But the superficial manner in which these topics are "solved" (by NetBuild)
indicate that the researchers have a very incomplete understanding of the
issues involved. For instance, selection of appropriate software is not
merely a matter of tagging libraries with architectural parameters and
selecting them. John Rice and colleagues at Purdue have been trying to
design such automatic algorithm selection systems for the past 20 years, and
the problem is one of the most open issues in numerical computing. Please
see the work by this group for more holistic approaches to this problem.

The section on "matching" assumes that ready-made lookup tables are available
to aid in this activity. There is no mention of how software performance
is to be assessed and how the set of attributes is used to distinguish
"good" performance from "bad." What are the decision algorithms to be used?
How sophisticated can they get?

It may be the case that the authors want to merely "facilitate access" once
the methodological details of the software selection are resolved. If this is 
the case, then NetBuild is really just a bunch of glue scripts that will form
part of a larger, high-level, problem-oriented, application-specific system.
I have heard Jack Dongarra speak about a vision of "adaptive software 
libraries" but will be quite surprised if this simplistic solution is what 
he had in mind.

There are also a lot of assumptions about how computations are conducted
on the grid. What if algorithms have to be switched dynamically during
a computation? What are the checkpointing facilities? Access to libraries
cannot simply be a matter of running them remotely.

In conclusion, I recommend that the authors address more serious issues and 
attempt to incorporate NetBuild in a larger context of dynamic software
selection, compilation, and linking technologies.