Abstract
Optical Network on Chip (ONoC) is an emerging chip-scale optical interconnection technology to realize the high-performance and power-efficient inter-core communication for many-core processors. By utilizing the silicon photonic interconnects to transmit data packets with optical signals, it can achieve ultra low communication delay, high bandwidth capacity, and low power dissipation. With the benefits of Wavelength Division Multiplexing (WDM), multiple optical signals can simultaneously be transmitted in the same optical interconnect through different wavelengths. Thus, the WDM-based ONoC is becoming a hot research topic recently. However, the maximal number of available wavelengths is restricted for the reliable and power-efficient optical communication in ONoC. Hence, with a limited number of wavelengths, the design of high-performance and power-efficient ONoC architecture is an important and challenging problem.
In this thesis, the design methodology of wavelength-reused ONoC architecture is explored. With the wavelength reuse scheme in optical routing paths, high-performance and power-efficient communication is realized for many-core processors only using a small number of available wavelengths. Three wavelength-reused ONoC architectures and communication schemes are proposed to fulfil different communication requirements, i.e., network scalability, multicast communication, and dark silicon.
Firstly, WRH-ONoC, a wavelength-reused hierarchical Optical Network on Chip architecture, is proposed to achieve high network scalability, namely obtaining low communication delay and high throughput capacity for hundreds of thousands of cores by reusing the limited number of available wavelengths with the modest hardware cost and energy overhead. WRH-ONoC combines the advantages of non-blocking communication in each lambda-router and wavelength reuse in all lambda-routers through the hierarchical networking. Both theoretical analysis and simulation results indicate that WRH-ONoC can achieve prominent improvement on the communication performance and scalability (e.g., 46.0% of reduction on the zero-load packet delay and 72.7% of improvement on the network throughput for 400 cores with small hardware cost and energy overhead) in comparison with existing schemes.
Secondly, DWRMR, a dynamical wavelength-reused multicast scheme based on the optical multicast ring, is proposed for widely existing multicast communications in many-core processors. In DWRMR, an optical multicast ring is dynamically constructed for each multicast group and the multicast packets are transmitted in a single-send-multi-receive manner requiring only one wavelength. All the cores in the same multicast group can reuse the established multicast ring through an optical token arbitration scheme for the interactive multicast communications, thereby avoiding the frequent construction of multicast routing paths dedicatedly for each core. Simulation results indicate that DWRMR can reduce more than 50% of end-to-end packet delay with slight hardware cost, or require only half number of wavelengths to achieve the same performance compared with existing schemes.
Thirdly, Dark-ONoC, a dynamically configurable ONoC architecture, is proposed for the many-core processor with dark silicon. Dark silicon is an inevitable phenomenon that only a small number of cores can be activated simultaneously while the other cores must stay in dark state (power-gated) due to the restricted power budget. Dark-ONoC periodically allocates non-blocking optical routing paths only between the active cores with as less wavelengths as possible. Thus, it can obtain high-performance communication and low power consumption at the same time. Extensive simulations are conducted with the dark silicon patterns from both synthetic distribution and real data traces. The simulation results indicate that the number of wavelengths is reduced by around 15% and the overall power consumption is reduced by 23.4% compared to existing schemes.
Finally, this thesis concludes several important principles on the design of wavelength-reused ONoC architecture, and summarizes some perspective issues for the future research.