This thesis argues that it is the communication costs of algorithms, rather than their computation costs, that will dominate future computing concerns. That, as we move to thousands of cores on a chip, the physical spatial locality of computation and data becomes critical to performance and cost. However, there is very little in the way of theory, models, or even characterisation of such locality for Chip-Multiprocessors (CMP). This thesis adapts and extends the existing theory and models of wire locality in VLSI circuits to the physical and temporal locality of software running on CMPs. It aims to provide a new foundation for characterising, modelling, predicting and exploiting the communication properties of software, which as we show, exhibits Rentian fractal scaling. In doing so, it lays a new communication-centric foundation for CMP software and hardware, and provides fundamental insights into their continued technological scaling.
This thesis argues that it is the communication costs of algorithms, rather than their computation costs, that will dominate future computing concerns. That, as we move to thousands of cores on a chip, the physical spatial locality of computation and data becomes critical to performance and cost. However, there is very little in the way of theory, models, or even characterisation of such locality for Chip-Multiprocessors (CMP). This thesis adapts and extends the existing theory and models of wire locality in VLSI circuits to the physical and temporal locality of software running on CMPs. It aims to provide a new foundation for characterising, modelling, predicting and exploiting the communication properties of software, which as we show, exhibits Rentian fractal scaling. In doing so, it lays a new communication-centric foundation for CMP software and hardware, and provides fundamental insights into their continued technological scaling.
Reihe
Sprache
Verlagsort
Verlagsgruppe
BCS Learning & Development Limited
Zielgruppe
Maße
Höhe: 254 mm
Breite: 203 mm
Dicke: 15 mm
ISBN-13
978-1-906124-92-2 (9781906124922)
Copyright in bibliographic data and cover images is held by Nielsen Book Services Limited or by the publishers or by their respective licensors: all rights reserved.
Schweitzer Klassifikation
Daniel Greenfield completed his PhD at the University of Cambridge as a Gates Cambridge Scholar, supervised by Dr Simon Moore. Prior to this he spent many years designing innovative GPUs, media processors and network processors in Silicon Valley. He obtained his previous degrees from the University of New South Wales, where his Masters developed new algorithms in systems biology. He has won multiple competitions in software, and has represented Australia internationally. He is currently Managing Director and co-founder of Fonleap Ltd, a high-tech startup in Cambridge.
Daniel Greenfield completed his PhD at the University of Cambridge as a Gates Cambridge Scholar, supervised by Dr Simon Moore. Prior to this he spent many years designing innovative GPUs, media processors and network processors in Silicon Valley. He obtained his previous degrees from the University of New South Wales, where his Masters developed new algorithms in systems biology. He has won multiple competitions in software, and has represented Australia internationally. He is currently Managing Director and co-founder of Fonleap Ltd, a high-tech startup in Cambridge.
1 Introduction
2 Motivations: communication vs. computation
3 Background
4 Rentian locality for Networks-on-Chip
5 Experimental evidence for Rent's rule
6 Generalising Rent's rule
7 Spatio-temporal Rentian model: CMP scaling implications
8 Effect of Locality on Asymptotic Complexity
9 Rentian scaling in neuronal networks
10 Conclusions and Future Work
