The rest of the dissertation is organized in seven chapters.
Chapter 2 reviews the current emerging interconnect fabrics proposed by many research groups for efficient communication of future large-scale NoC architecture and existing methodologies used for mapping topology on NoC platforms.
Chapter 3 describes the basic heterogeneous architecture of our framework, defines network performance metrics, and introduces the network traffic scenarios that is used in this dissertation.
Chapter 4 presents the benefits of using heterogeneous link types for NoC de-signs. We obtain optimized heterogeneous NoC architectures by using three dif-ferent technology-driven link types and evaluate and compare the network perfor-mance in terms of throughput and energy with homogeneous and 2D regular mesh networks.
Chapter 5 analyzes the structure of the obtained heterogeneous NoC architec-ture by using tools from complex network science, such as community detection and small-worldness, to understand how heterogeneous link types are placed to improve the network performance.
Chapter 6 presents NoC architectures by using ten different abstract link types.
We show that optimized networks with ten different link types outperform not only homogeneous and regular mesh networks but also networks with three different
technology-driven link types.
Chapter 7 describes details of the basic architecture of our core assignment framework and shows that application mapping onto application-specific optimized NoC architecture with heterogeneous link types supports better network perfor-mance compared to mesh NoC architectures.
Chapter 8 concludes the dissertation and presents future work.
2
Related Work
2.1 Emerging On-Chip Communication Media
Network-on-Chip (NoC) have been introduced as a promising solution for scal-able communication of future large-scale multi-cores systems and to support the performance and power requirements of future applications. Traditional NoC ar-chitectures are based on packet-switching networks. Each packet goes through switches/routers and interconnect links, thus generating large amounts of multi-hop long-range communication. This can significantly affect the network perfor-mance and power consumption [105].
In the last few years, several solutions were proposed to improve the network performance. Ogras and Marculescu [121] have proposed inserting a few range links to standard mesh NoC topologies. The results show that adding long-range links reduces the average distance between source and destination nodes, which increases the network throughput and reduces the average packet latency.
However, the authors did not consider cost and scalability in their paper.
Using existing metal interconnects will be highly inefficient and become harder to satisfy the design requirements, such as power, delay, and reliability for future large-scale electronic systems [34, 151, 152]. This is because as interconnects are scaled due to the increase in the number of components per single chip, their resis-tance increases. Therefore, using metal as a long-range link will be challenging for interconnect delay and power dissipation [80]. Researchers have recently proposed different emerging interconnect technologies to be used as NoC communication
links.
Photonic interconnects were introduced as a promising new technology for NoC communication [29, 72, 86, 145, 171]. A few recent works have shown that photonic NoCs provide lower latency, lower power, and higher bandwidth compared to wired interconnects [86, 99, 119, 143, 144]. Photonic networks can support transmitting large amount of data across long distances at low latencies [10, 90]. For example, Bahirat and Pasricha [10] introduced photonic ring-based hybrid photonic NoC architecture. The photonic ring waveguide, used as a global on-chip communication channel, is built on top of a regular 2D mesh network to improve the latency and reduce power dissipation (see Figure 2.1 (a)). However, the components are expensive and there are still open problems that need further investigation, e.g., photonic NoCs require to integrate true speed on-chip light sources [21, 72], high-radix photonic switches [145], and thermal management [160].
Another promising interconnect technology is multi-band RF interconnects (RF-I) introduced by Chang et al. [31–33]. They considered various implemen-tation challenges of hybrid NoC architecture using RF-I technology (RF transmis-sion line) laid in a zig-zag pattern on top of a wired mesh (see Figure 2.1 (b)) and demonstrated the performance achievements in [31]. The benefit of using RF interconnect is that it transmits electromagnetic (EM) waves, which travel at the effective speed of light along the wires. The RF interconnects provide fast data transport over long distances. The work shows that higher bandwidth and lower latency can be obtained by using this technology [31].
Floyd et al. [60] first introduced on-chip wireless interconnects as a new inter-connect technology for clock distribution. Zhao and Wang [173] proposed CMOS
(a) (b)
(c)
Figure 2.1: Hybrid NoC architectures. (a) Photonic ring-based hybrid photonic NoC. Source: [10]. Photonic ring interconnect placed on top of a regular mesh network. (b) RF-interconnect NoC. Source: [31]. Z-shaped RF-interconnect is built on top of a regular mesh network. (c) Hybrid wired/wireless NoC. Source:
[65]. Wired links are used to communicate intra-components in a sub-network and wireless links are used to communicate between hubs.
ultra-wideband (UWB) wireless technology for high-data rate low-power short-range communication. Lee et al. [97] have proposed a wireless on-chip architecture, which uses a hybrid wireless and wired architecture to interconnect cores. They have shown that simple transmitters/receivers operate at the 100-500 GHz sub-terahertz frequency through miniature antennas, which reduces latency and power consumption compared to a 2-D mesh network. In Deb et al. [50], the authors intro-duced a design of a hierarchical small-world network with long-range and low power mm-wave wireless links for NoC (mWNoC) and showed a performance improve-ment for both uniform and non-uniform traffics. Ganguly et al. [65] introduced a new on-chip antenna based on Carbon NanoTubes (CNTs) for on-chip wireless communication (see Figure 2.1 (c)) and evaluated latency, throughput, and energy dissipation. The network is divided into small sub-networks and uses wireless links to communicate between sub-networks through hubs located in each sub-networks while wired links are used to communicate between the components within the sub-network. They showed that the performance, in terms of throughput, latency, and power consumption improved compared to a general wired network. In particular, more improvement can be seen in non-uniform traffic distributions compared to uniform traffic distributions by using low power and high speed long-range wireless links, which enable a single hop communication between distant nodes. Although the RF/wireless interconnects have many advantages, this new technology is still in an early stage of evolution. It needs more research to solve open problems, such as packaging and interference issues, ultra-high frequency requirements, efficient on-chip antennas, and power overhead.
Each of the above mentioned emerging interconnect technologies has several issues and problems that need to be investigated. However, in the past few years,
a few researchers have studied and shown some promising initial results that they may be suitable candidates for global on-chip interconnects in the future large-scale multi-core systems. While many authors have introduced new solutions to improve the network performance, they did not consider balancing all of the important performance metrics of the network, such as the wiring cost, the average shortest path length, the latency, the throughput, and the power consumption. Also, they only compared the results with a regular 2D mesh network and did not in general compare the networks with each other. In addition, their networks are based on regular mesh. Teuscher [154, 155] showed that unstructured nature-inspired NoCs can have benefits for performance and scalability over traditional structured NoCs.