Page 22 - Kotov

PS-x000
Some specialized applications, such as geophysics, fluid dynamics,
real-time process control, and similar needed relatively cheap high-
speed computers. A joint effort of the Institute of Control Prob-
lems and the Institute of Electronic Control Machines produced in
1970s the PS-2000 and in 1980s PS-3000 multiprocessors for that
applications.The computers had some features in common, yet had
significantly different architectures.
PS-2000 consisted of a parallel SIMD (Single-Instruction Mul-
tiple-Data) processor with between 8 and 64 processing units to-
gether with a monitoring host subsystem and an external storage. The SIMD multiprocessor PS-2000
The processing units are linked together, each element with two
its neighbors plus a serial bus linking all of them, and can be also segmented into clusters. The processing units
worked under a single control units that broadcasts instructions and operands for simultaneous execution on all
(or a part of) the units. The maximal speed was around 200 mln instruction per second. . The main disadvantage
was the 24-bit word length. Effective use of the PS-2000 parallelism was difficult. The majority of software
developed for PS-2000 has been written in an assembly language. There were around 150 PS-2000 systems in-
stalled by the mid 1980s.

PS-3000 was a 32-bit multiprocessor system used primary for real-time
process control. It was a MIMD (Multiple-Instruction Multiple-Da-
ta) system consisting of four SIMD multiprocessors. The developers
claimed the maximal speed of 3 bln operations per second. However,
the system did not went into serial production and not widely used.

The PS-3000 multiprocessor system

Experimental Developments of the 1980s

As in the USA and other western countries, new ideas and solutions in computing were born in academia. The
first Soviet computers has been designed in the USSR Academy of Sciences (or in a close cooperation with it).
New architectural ideas for supercomputers were also emerging in the Academy . However, the Academy had
no enough funds to do ambitious projects and had no access to the development and production facilities. Some
most entrepreneurish researchers were able to sell their ideas to the government and industrial institutions in the
late 1970s and early 1980s and initiated joint experimental projects in the supercomputer architectures. However,
the perestroika in the 1980s and collapse of the Soviet Union canceled any possible follows up for those projects.

ES-1766. The theoretical foundations for the ES-1766 were developed by Viktor M. Glushkov in 1970s and a
prototype was built in 1984. The ES-1766 was a “macro-pipeline” MIMD multiprocessor, in which different
processors repeatedly execute different large fragments of a program and send the output to their neighbors (it
was a sort of non-von Neumann data-flow architecture). With a complete set of 256 processors, the computer
should achieve 0.6 bln operations per second. To be efficiently executed on the macro-pipeline multiprocessor,
a program should be parallelized into fragments with balanced execution and communication time slots. So, the
ES-1766 had a software development environment that hides all hardware details from the user. The operating
system managed the assignment of program fragments to processors.

MARS-M. MARS stands for the Modular Asynchronous Reconfigurable System and grew from the late 1970s
idea of an open hierarchical architecture with functional partitioning of subsystems at each level. The partitioning
was implemented on the set of asynchronously communicating functionally specialized modules (processing, or
memory, or control, or communication processors). Idealistically, such an architecture can be viewed as a fractal
– 20 –