Laboratory of Parallel Algorithms
and Architectures
Head of the Laboratory: Sergey
V. PISKUNOV
SUMMARY.
Main topics include investigations of different
formal models of parallel processes, as well as simulation of parallel
algorithms and structures, and elaboration of methods for high performance
parallel architecture synthesis.
IN MORE DETAILS.
Modern technology of parallel computing
organization relies on two types of parallelism: coarse-grained
and fine-grained ones.
Such a differentiation is rather conventional, nevertheless the
following is worth to be noted.
Coarse-grained parallelism is inherent in computer systems,
composed of a number of (tens, hundreds) powerful interconnected computers,
forming a network.
Fine-grained parallelism is inherent in computing systems,
including a huge number (tens, hundreds thousands) of relatively simple
processing elements. Connection between them have a
regular structure and frequently (but not always!) are organized
according to local interactivity principle. As a rule, such systems are
highly specialized. The term itself "fine-grained parallelism" exhibits
simplicity and promptness of any computing action. A characteristic feature
of fine-grained parallelism is that the approximate equality of computing
intensity and that of data exchange.
Fine-grained parallelism has a long history: it is the
most "ancient" variety of parallelism. Its theory has been developed
concurrently with that of sequential computations, being
associated with the name of John fon Neumann. His theoretical model of
fine-grained parallel computer, referred to as "cellular automaton" is
widely known.
In the laboratory the investigations of fine-grained parallelism
are held in many directions.
We consider fine-grained parallelism to be attractive,
because in its framework it is possible to find the best (for
example, according to temporal characteristics) parallel algorithms for
solving many practically important problems both numerical and
nonnumerical ones. Moreover, some problems may be solved only in the
framework of a certain fine-grained parallel model of computations
(for example, solution of hard to be formalized problems by neural
networks using learning process). Practical importance of fine-grained
parallelism is in two facts: first, that it serves as a source of methods
for solving complex problems on the modern multiprocessor computer
systems, and second, that many of both real and hypothetical
special-purpose processors have fine-grained parallel architecture.
INVESTIGATIONS OF FINE-GRAINED PARALLELISM
CARRIED OUT
IN THE LABORATORY
Elaboration and investigation of fine-grained parallel algorithms
(as well as of any other) is performed using a certain
computation model. Nowadays, the following classes of models are
used for fine-grained computations: "cellular automaton", "systolic
array", "associative processing system", "neural network", "cellular neural
network". The model class chosen for constructing and exploring a
parallel algorithm predetermines the architecture of the processing
unit (may be, an abstract one), which realizes the given algorithm. The
problem is in searching such an algorithm, which is
the best, according to a certain criterion ( for example, a number of
steps) to be realized in a given architecture. The role of
such investigations is hard to be overestimated. They are of great
practical and theoretical importance. Researchers, dealing
with models of one class, may easily compare the results and join
their efforts to search for the best parallel algorithms. Many
real and hypothetical processors architecture is in fact a realized in
hardware model of computation from one of the above classes.
According to mentioned above, the following research themes are
under investigation in the laboratory.
-
Development of cellular technology for parallel algorithm and structure
synthesis, which is based on a model of distributed computations called
Parallel Substitution Algorithm .
-
Analysis and synthesis of multiplanar cellular architecture (universal [2]
and specialized [3]), oriented to 3D pipe-lined computations.
-
Elaboration and investigation of associative parallel algorithms for
nonnumerical (especially, graph-theoretical) problems solution [4].
-
Design and investigation of distributed functional
structures [5] and high performance special purpose processors
construction on this base.
-
Development and investigation of cellular neural algorithms for image
processing [6], distorted patterns recognition [7],
and discrete optimization [8].
-
Recently, the investigations associated with simulation of physical
phenomena in discrete space (cellular automaton, cellular neural
network) are intensively evolved. We don't stay aside of this scientific
direction: some problems are stated which are related to
problems from mathematical physics using cellular neural approach.
Particularly, cellular neural networks, simulating autowaves and
dissipative structures are under investigation [9].
-
Research in fine grained parallelism is performed with
the help of program tools, elaborated in the laboratory. Systems,
based on the languages STAR [10]
and VEPRAN [11] are applied to
associative algorithms investigation. These systems are permanently
modified and improved. Cellular algorithms and structures
(including neural and cellular neural networks) are studied with
the help of a simulating system ALT [12]. A new more powerful
simulating system WinALT [13]
is under development.
-
Design of combined architecture [14] of computer systems,
comprising special purpose fine grained parallel processors.
-
Elaboration and investigation of associative algorithms for numerical
processing. This research has been transformed into a work on
creation a program package for highly precise computations on
computers of different types [15]. Now this work proceeds in the
laboratory of parallel program synthesis [16].
-
Something new is often a well forgotten past. In order to move forward,
the past should be well learned. So, computer science hystory in
Russia is studied in the laboratory [17].
PERSPECTIVES.
Future investigations are seen in the following:
-
in transferring the ideas of fine-grained algorithms
on the construÓtion of parallel algorithms for multiprocessor
computer systems;
-
in expanding spheres of fine grained parallelism application first
of all on the investigation of space-time dynamics
of active media, and on the evolution processes;
-
in system WinALT realization on supercomputers, which
should provide real-time computation while solving pattern
recognition problems, image processing, simulation in physics and
biology in cellular space etc.;
-
in constructing libraries of cellular algorithms and
structures models for WinALT implementation, which are oriented
to design of special purpose processors based on time-space
massive parallelism:
-
in designing and investigating associative parallel
algorithms for optimization on graphs and visualization the
execution process.
Last update October
2, 2000
Maintained by Elvire Kuksheva.