Sleptsov Net Computing Resolves Problems of Modern Supercomputing Revealed in Turing Award Talk

For more than two decades, efficiency of supercomputers has been measured solving big dense random linear systems using LINPACK library developed by Jack Dongarra team. In his Turing Award talk, Jack Dongarra calls LINPACK a yardstick of high performance computing. Using this yardstick, a series of supercompures in USA and other countries has been developed, including the most powerful computer in the world at the present time – Frontier of Oak Ridge Laboratory, USA. Remind that, performance of computers is measured in FLOPS – floating point operations per second, and Frontier breaks exaflops barrier that amounts to ten to the eighteenth power flops. In the final part of his talk, Jack Dongarra reveals recently obtained information on efficiency of the best computers, mentioned in Top500 list, during solving real-life tasks in manifold application domains basically represented by simulators which lead to solving sparse linear systems. Frontier and many other top computers show efficiency about 0.8% that means Frontier has real performance about ten petaflops instead of an exaflops. Only Japanese computer Fugaku of Fujitsu, occupying recently the second place, reaches 30% efficiency.

Indeed, supercomputing is not so simple as cutting up plots of land to measure them with a single yardstick but requires a complex balanced approach of using a mixture of simulators for the top significant application domains. Dense systems provide regular load on processors and communication subsystem filling the cash well to smooth out the processor-memory bottleneck. Real life tasks, represented by simulators and sparse systems, require irregular communication pattern that is implemented rather well by the Fugaku’s six dimensional torus Tofu D Interconnect. Thus, design of communication subsystems in the form of a multidimensional torus represents a prospective direction of research.

Sleptsov Net Computing goes further; it uses computing memory, represented as two and higher dimension structures, when spatial peculiarities of modeled systems are taken into consideration. Doing computations in memory allows us to get rid of the traditional processor-memory bottleneck problem. Instead of sophisticated conventional architecture to utilize multicore CPUs and GPUs of distributed nodes, for instance with OpenMP-CUDA-MPI set tools, Sleptsov Net Computing offers homogenous concept of a unified graphical language for concurrent programming with very fine granularity of parallel computations. As grown from place-transition net theory, Sleptsov net inherits developed methods to analyses and prove correctness of parallel programs in the process of their model-driven development. Implementation of Sleptsov net processors in the form a matrix of computing memory leads to the nanosecond tack of time of massively parallel computations that means good ability to control hypersonic objects in real-time applications.

Software reliability is a separate issue. Let us imagine that reliability of a bulk of software operating in the world is based on the belief that it does not contain an error or rather contains an error with low probability. The belief is supported by a certain number of successful tests on which a program in question works correctly. Such software is embedded in cars’ and planes’ onboard computers, to make us believe in low probability of an accident, while we prove correctness of concurrent Sleptsov net programs in a formal way.

Recently, two master students Qing Zhang and Hongfei Zhao from XIDIAN University, Xi’an, China implemented software prototypes of a Sleptsov net processor and a compiler-linker of Sleptsov net programs and presented them at the International Conference “Problems of Infocommunications. Science and Technology” (PIC S&T′2022), 10 -12 October, Borys Grinchenko Kyiv University, Ukraine. We invite investors to start-up the project of enterprise level implementation of Sleptsov net computing paradigm, including hardware implementation of Sleptsov net processor and computer. To integrate it initially into the conventional infrastructure, we attach it, as an extension device, to a traditional computer, how it was mentioned in Jack Dongarra talk as one of prospective directions for further developing modern architecture of HPC.

Besides, at the present time, an SNC full-scale project background has been well prepared by manifold keynote talks and dedicated lectures as well as with about a dozen of publications, including Special issue Petri/Sleptsov net based technology of programming for parallel, emergent and distributed systems of International Journal of Parallel, Emergent and Distributed Systems.

Chapter 122 of Advanced Methodologies and Technologies in Network Architecture, Mobile Computing, and Data Analytics (2 Volumes), "Sleptsov Net Computing" presentes motivation for new models of hyper-computations. Sleptsov net was introduced compared to Petri and Salwicki nets. A concept of universal Sleptsov net, as a prototype of a processor in Sleptsov net computing, was discussed.