An
Interprocessor Communication Interface for Message Passing via Shared Memory
Modules - Design and Performance
G.Lj.
Djordjevič,
M.K. Stojčev
Abstract.
In this paper the interprocessor communication interface intended for
realization of multiprocessor (MMC) system is described.
The MMC system is implemented as a Fully_Connected_n-side_Pyramid (FCnP).
The base of the pyramid consists of n
processors and it acts as an accelerator to the host computer that is placed at
the top of the pyramid. Communication between any two processors takes place
through Shared_Memory_Module (SMM) independently accessed by both processors
involved in current data transfer. The SMMs are realized with two-side
accessible memory chips of FIFO RAM type. For the processors we use standard
Single_Board_Computers (SBC) extended with a communication hardware referred to
as the Communication_Module (CM). The main task of the CM is to provide
efficient DMA transfer between the SBC's local memory and SMMs. Attaching the CM
to the SBC requires only some minor modification of the SBC's hardware. In order
to connect one SBC with several SMMs a special bus named Shares_Memory_Bus (SMB)
is provided. Higher FCnP's performances, in comparison with the common bus
biased MMC systems, are obtained mainly due to: increased communication
bandwidth, possibility to use heterogeneous processors, and configuration
flexibility of system topology. This paper deals with hardware structure of
constituent parts of the communication interface (CM, SMM, and SMB), and system
operation concerning the message transfer. Further on, performance evaluation
for the proposed communication interface related to system efficiency,
communication throughput, and message latency are carried out. Simulation
analysis is also included.
p-adic Arithmetic and
Parallel Symbolic Computation: An Implementation for Solving Linear Systems Over
Rationals
C.
Limongelli, R. Pirastu
Abstract.
In this work we describe the use of truncated p-adic
expansion of handling rational numbers by parallel algorithms for symbolic
computation. As a case study we propose a parallel implementation for solving
linear systems over the rationals.
The
parallelization is based on a multiple homomorphic image technique and the
result is recovered by a parallel version of the Chinese remainder algorithm.
Using a MIMD machine, we compare the proposed implementation with the classical
modular arithmetic, showing that truncated p-adic arithmetic is a feasible tool for solving systems of linear
equations working directly over rational numbers. A safe algorithm for computing
the p-adic division operation is
proposed.
The
implementation leads to a speedup of up to seven by ten processors with respect
to the sequential implementation.
Implementation
of the Self-Organizing Feature Map on Parallel Computers
V.
Demian, J.-C. Mignot
Abstract.
In this paper, we propose two implementations of the SOFM on parallel computers.
One is for a MIMD computer, the other one is for a SIMD computer. We propose a
new learning method for the SOFM using a block strategy. This allows to exploit
the high performance level of the new generation of parallel computers. We show
that the block strategy performs well on several examples outperforming
classical implementations. A model to describe the performance of this algorithm
is proposed and compared with experimental data. Finally, we compare
experimental results
on the two classes of parallel computers.
Evaluation
of Differential Methods for Image Velocity Measurement
P.
Handschack, R. Klette
Abstract.
For eight point-based differential methods of image velocity measurement,
including two new methods, quantitative evaluations are reported based on
synthetic images, and qualitative evaluations based on real images.
A new method
(cluster method) is
characterized by the determination of the nearest point to all constraint lines,
and the other (VF expansion method) by
local approximation of the motion vector field. For six methods, solutions for
algorithmic use are compiled, e.g. also iterative solutions for the Schunck and
the Nagel constraint. The given quantitative evaluations may be used for
characterizing situations where methods may be applied, and where not, and what
parameters may be suggested.