METHODS OF EFFECTIVIZATION OF SCALABLE SYSTEMS: REWIEW
Keywords:: effectivization, scalable systems, high performant computing, architecture, topology
The article discusses the problem of inefficiency of modern systems and horizontal scaling as a method of increasing productivity. The main issues that make up the mentioned problem are highlighted: parallelism constraints, mismatch between the task and the system, the complexities of programming and the question of the balance between cost and performance. A classification for possible solutions was proposed, according to which they were divided into architectural and network, and an overview was carried out. As part of the architectural class, such approaches as quantum computing and the dataflow paradigm were reviewed, the most promising solutions were analyzed.
The comparative analysis shows that by their nature dataflow and quantum computing do not contradict each other, moreover, they complement each other in the context of the problem. Thus, specialized D-Wave quantum computers, in contrast to universal quantum processors, provide large computing power at a relatively modest price, while the dataflow solution, represented mainly by Maxeler processors, is universal and efficient, but inferior to quantum systems in a number of tasks.
At the same time, both types of processors require a certain network for communication, which makes the issue of topology relevant. At the network level, 2 topologies - Fat Tree and Dragonfly - were considered, and their main properties were highlighted. The analysis showed that in the context of the problem Dragonfly is slightly better due to decentralization and smaller diameter, however, both solutions provide good topological characteristics and support for the main modern routing technologies.
In the conclusions, the main aspects of problem formulation and review are indicated, further prospects and possible methods are considered. First of all, a promising idea is the combination of quantum and non-quantum solutions in one system. This approach allows you to significantly speed up certain calculations, while ensuring the universality of the system. However, a more general issue is the mutual integration of solutions as such. The problem of efficiency has many partial solutions, but not all of them are compatible, therefore, the development of complex methods on the basis of already known ones is a key perspective of the subject area.