Documents

DOI

  • Prashant Joshi
Networks such as road networks, utility networks, computer and communication networks and even social networks are the backbone of human civilization. Network analysis enables quantitative measurement of the important criteria such
as delays, ease of routing and fault tolerance, and is required to build efficient and robust networks. Computer networks have evolved over the last five decades in parallel with technology which has grown exponentially tracking ‘Moore’s Law’, which projected exponential performance growth in computing. Notably though, the supercomputers of today pushing exascale performance are doing so, not primarily because of the improved performance of the microprocessors, but overwhelmingly due to the ability to network tens of millions of these microprocessors in systems. These systems depend very heavily on robust network topologies to achieve the exponentially growing performance seen over the last few decades. The network topologies in the world’s top performing supercomputers have evolved with the focus towards boosting performance by binding together an increasing number of processors in efficient networks over the years. Popular topologies have included torus, hypercubes, fat trees and some combinations thereof. The biggest drawbacks of the rapidly increasing number of devices networked together are the increased message delays, the declining ability to withstand various faults, and security issues. Building such supercomputers of today has very high down costs, and it is imperative that their utilization is maximized. This requires these high performance systems to be highly dependable also. This forms the motivation for the work in this thesis.
Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date7 Oct 2019
Print ISBNs978-94-028-1709-6
DOIs
Publication statusPublished - 2019

ID: 57078352