Comparison of Different Architectures, Tools and Systems.
We believe that CORBA is the appropriate tool to use for
building a general infrastructure for communicating between different
statistical environments, applications and tools. We think it is
sufficiently open and extensible to adapt to the needs we will
discover as we create this integrated environment. We have some
experience with MPI/PVM and have read about a few others.
We are very interested in other people's opinions and experiences
with these systems and how they might be utilized for building an
environment for data analysis and statsitical research.
There are a myriad of different considerations, not least of which is
the cost of learning such a system. Below are some comments on the
advantages and disadvantages of the different tools. These should be
read in conjunction with a
list of ideal requirements for such a system.
In summary, each of the tools is useful for providing solutions
to different communication needs. For a general extensible system,
we still think CORBA is the right way to go.
From a non-technical viewpoint, one of the compelling reasons for
employing CORBA is that it has significant and growing support and
momentum.
MPI and PVM have been around a while and are used in quite specialized applications.
The RMI approach is limited to Java and DCOM doesn't have the same
wide appeal as CORBA in that it is Microsoft oriented.
Given the prevalent disposition of many developers to Linux, that
approach won't fill us with enthusiasm.
Since the end user will not see the infrastructure, we need
something that will not constrain us in the future and which
is itself well supported. If that involves an initial
learning cost, I believe the results will be worth it.
Additionally, this cost should bear great rewards in time
saved not reimplementing existing applications/tools.
We also hope to provide different tools to developers to
allow them simply to make their services CORBA-compliant.
These tools include a Java 2 IDL
generator,
C and Java libraries for manipulating common statistical data
types
and S level tools.
Alternative Sources of Comparisons
JavaWorld article
CORBA
versus DCOM
In short, while RMI is a viable option for
smaller-scale applications to be written entirely in Java, CORBA
provides the foundation for integration of existing objects with new
code and also the ability to scale for the future.
PVM
- Provides fault tolerance.
- Requires message content and message ID agreement.
- Process to process communication.
- Some tools supporting process visualization and debugging.
- Reasonably widely used in small domain of high performance
numerical computing.
- No support for self describing objects. This would require us
to build this infrastructure.
- Few tools for security, load balancing, etc.
- Successful research project
MPI
- MPI works only on homogeneous machines, not heterogeneous clusters.
- Similar issues as PVM regarding message contents,
communication, etc.
- Very fast implementations from different vendors of hardware.
- Work is being done to integrate MPI and PVM.
WebObjects
As a general tool for the goal of providing seamless connectivity
between different statistical components, WebObjects will not suffice.
Some potential problems include:
- Provides load balancing.
- Easier entry level learning curve/access than CORBA.
- Access to databases.
- Easy Web access.
- Doesn't do what we want for a general system
since it is apparently an intelligent form of CGI.
It should be used as such.
- Appears to have only one implementation which is commercial
This seems to be a significant prohibitive factor.
- Focus on Web rather than distributed objects with different configurations.
- Support or Unix? Doesn't appear to be.
- Limited support for C, just Objective C? Support for other
languages than C and Java - e.g. Perl?
- Non self-describing objects on which dynamic invocation
can be performed. In other words, the systems may not be
dynamically extensible by self discovery and may require
all the method invocations to be compiled.
- Load balancing may not be as flexible as other
implementations/configurations for different types of tasks
such as High Performance Computing, etc.
DCOM
- Many tools for developing applications for Windows.
- Very Microsoft specific and currently runs on only NT and
Windows. Recently ported to Sun's Solaris.
- Not an open standard.
-
DCOM is a very immature technology when compared
with CORBA.
(From
Java World article on CORBA and Java.)
- Not trivial to learn and use. Same as CORBA in this respect.
- Limitations of the interface language.
CORBA
- Provides all the facilities mentioned in the requirements.
- Available in commercial and non-commercial implementations. The
non-commercial implementations are not quite complete, but
getting there.
- Operating system independence.
- Communicates with other ORBs via IIOP allowing different ORBs
to be connected.
- Communicates with RMI compliant objects and also DCOM
connectivity.
- Significant development in other domains which we inherit
gratis.
- CORBA has emerged as a standard from experience with
distributed computing rather than as an abstract model. So it is
mature and flexible.
- Support for numerous languages C++, Ada, COBOL, Smalltalk, and
Java.
However, C++ and Java are the two primary languages supported.
- Steep learning curve - complicated by overwhelming number of
features and vendor specific extensions.
- Minimal additional programming complexity (when the basics are
understood & mastered).
- Full Exception support allowing for good programming structure
and error handling.
- IDL doesn't support polymorphism.
Duncan Temple Lang
<duncan@research.bell-labs.com>
Last modified: Fri May 22 08:02:23 EDT 1998