Supercomputing Software and Hardware

Article # 27_2022

22 с.

PDF

Gennady G. Stetsyura
Synchronous execution of group operations in distributed supercomputer components and computer clusters

This paper proposes decentralized processes for synchronizing the actions of a distributed group of active components (objects) in supercomputers and computer clusters, allowing them to move to specified states or influence the external environment synchronously. The object action depends on the current state of the object and the external environment. The actions should start with the minimum delay after the possibility of their execution is detected. Synchronization is performed by exchanging optical signals over wireless communication channels through an optical signal repeater, combining one group of objects or sequences of groups of objects (layers). Accurate distance measurement performs the compensation of possible changes in distances between objects. Group operations accelerate synchronize and simultaneously receive data from a group of distributed objects. Data processing occurs during their transfer, without increasing the time. The operation time does not depend on the quantity of data processed by the operation. A group operation is performed in a repeater containing no computational means. (in Russian).

Key words: supercomputers, group operations, decentralized control, multilayer synchronization of object actions, distributed in-network computing.

article citation

DOI

https://doi.org/10.25209/2079-3316-2022-13-4-3-24

Article # 29_2022

30 с.

PDF

Viktor S. Podlazov
Multichannel non-blocking system area network with direct channels

An unblockable non-blocking self-routing network with direct channels has been developed, in which packet conflicts are resolved at the entrance to the network by the procedure of connecting sources to the first cascade of the network , providing a packet duality. Packets blocked during conflict resolution are retransmitted by sources with minimal delays. Within the network, the occurrence of conflicts is prevented by means of parallelization of the network itself. The network is designed in 2-, 4- and 8-stage versions with scaling of the number of channels from several hundred to many millions, keeping the same network performance. The network can provide 1- or 2-channel fault tolerance at the same link rate. The overhead cost of achieving these properties is higher complexity of the network, that becomes comparable to the complexity of the theoretical non-blocking Clos switch, which has no known practical implementation.The purpose of the proposed networks is photonic networks with routing information transmission in packet headers represented as one-time-use binary numbers. The proposed networks are made in an extended circuit basis, consisting of switches and separate multiplexers and demultiplexers. The paper presents the characteristics of the constructed networks with the specified method of presenting routing information. (in Russian).

Key words: supercomputers, group operations, decentralized control, multilayer synchronization of object actions, distributed in-network computing.

article citation

DOI

https://doi.org/10.25209/2079-3316-2022-13-4-47-76