The Paraguin Project is being conducted at the Department of Computer Science at the University of North Carolina at Wilmington . The purpose is to develop a message-passing parallelizing compiler for distributed-memory parallel computer systems. We are using the SUIF Compiler from Stanford University to build our compiler. In fact, the Paraguin Compiler is simply a SUIF compiler pass.
With the introduction of multi- and many-core systems, parallel
computing is now a main-stream computer science topic.
Traditionally, parallel computing was considered an advance computer
science topic. Not only do we currently see processors with
several cores and soon will see processors with hundreds of cores,
single core machines are no longer produced except for very specialized
and imbedded systems. Futhermore, GPUs are providing another
source of massively parallel architectures. Now it will and is becoming
a topic that is required by all computer science programs. There has
always been a lag between the development of hardware and the
development of software to take advantage of that hardware. Now
the gap between hardware and software development will grow
significantly larger.
Parallel computing is a difficult endeaver. Typically,
parallel software is develop by more gifted and talented computer
developers. This means that there will become a serious problem
with developing software that exploits the multi-core systems. At
present, there is already fewer students graduating with computer
science degrees than are needed by the software industry. Add to
that fact that many of the graduating students are neither trained to
program parallel systems nor have the talent, there is will be a woeful
shortaged of software developers for todays and tomorrows computers.
What is needed are two things: 1) as educators, we need to train
computer science students to develop correct parallel program that can
make use of todays multi-core systems; and 2) as scientists, we need to
develop tools that make the development of parallel programs easier and
more robost.
The Paraguin Compiler produces MPI
code suitable for execution on distributed-memory systems. The MPI
standard is considered still to be a low-level abstration for
expressing parallel algorithms. The Paraguin Compiler raises the
level of abstration to a similar level as OpenMP. Most experts would agree that
it is easier to develop a parallel algorithm using OpenMP versus
MPI. We have found in teaching parallel programming, that
undergraduates
students find programming with OpenMP to much easier than with MPI.
This is partially due to the use of shared-memory systems as opposed to
distributed-memory systems and partially due to the simplier interface
of OpenMP. Shared-memory is a simplier systems for which to
program because there is no inter-processor communication required; but
rather synchronization.
The Paraguin Compiler provides a similar abstraction as OpenMP but
for distributed-memory systems. Using a similar interface, the
user can have the compiler generate the MPI code. However, we feel that
such a level of abstraction is only a step in the right
direction. We will be introducing patterns (see [1] and [2]) in a
future release of
the software. Using patterns, the developer can choose a pattern
that matches their particular algorithms, provide the compiler with the
patterns and details of the individual steps and data that needs to
flow following that pattern, and the compiler will produce a parallel
program. Some patterns that will be implemented are:
Scatter/Gather, Workpool,
Pipeline, Stencil, Divide-and-Conquer, and All-to-All. We are in close
collaboration with the "Seeds
Framework", which is an environment for programming parallel
systems using the above patterns. Pattern programming using the Seeds
Framework as well as the Paraguin Compiler was presented at a Workshop
at SIGCSE 2013 [3].
In the current version of Paraguin, the Scatter/Gather and the 2-D
Stencil patterns have been implemented.
This page was last updated: May 12, 2014
Email:
