Use this URL to cite or link to this record in EThOS:
Title: The effect of an optical network on-chip on the performance of chip multiprocessors
Author: Van Laer, Anouk
ISNI:       0000 0004 7231 3823
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Optical networks on-chip (ONoC) have been proposed to reduce power consumption and increase bandwidth density in high performance chip multiprocessors (CMP), compared to electrical NoCs. However, as buffering in an ONoC is not viable, the end-to-end message path needs to be acquired in advance during which the message is buffered at the network ingress. This waiting latency is therefore a combination of path setup latency and contention and forms a significant part of the total message latency. Many proposed ONoCs, such as Single Writer, Multiple Reader (SWMR), avoid path setup latency at the expense of increased optical components. In contrast, this thesis investigates a simple circuit-switched ONoC with lower component count where nodes need to request a channel before transmission. To hide the path setup latency, a coherence-based message predictor is proposed, to setup circuits before message arrival. Firstly, the effect of latency and bandwidth on application performance is thoroughly investigated using full-system simulations of shared memory CMPs. It is shown that the latency of an ideal NoC affects the CMP performance more than the NoC bandwidth. Increasing the number of wavelengths per channel decreases the serialisation latency and improves the performance of both ONoC types. With 2 or more wavelengths modulating at 25 Gbit=s , the ONoCs will outperform a conventional electrical mesh (maximal speedup of 20%). The SWMR ONoC outperforms the circuit-switched ONoC. Next coherence-based prediction techniques are proposed to reduce the waiting latency. The ideal coherence-based predictor reduces the waiting latency by 42%. A more streamlined predictor (smaller than a L1 cache) reduces the waiting latency by 31%. Without prediction, the message latency in the circuit-switched ONoC is 11% larger than in the SWMR ONoC. Applying the realistic predictor reverses this: the message latency in the SWMR ONoC is now 18% larger than the predictive circuitswitched ONoC.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available