Lei Guo,Weigang Hou, Pengxing Guo

Key Laboratory of Medical Image Computing of Northeastern University, Ministry of Education, Shenyang 110819, China School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China

* The corresponding author, email: houweigang@cse.neu.edu.cn

I. INTRODUCTION

System-on-Chip (SoC) using multi-processor is a promising solution for chip that replaces the traditional baseband architecture style, and multiple cores are implemented on a SoC for control and data processing with increased power eddiciency[1]. However, the future System-on-Chip (SoC) will integrate a list of function blocks, and it will have a powerful computing ability [2-4]. For example, With the rapid development of 5G[5-6], Internet of things [7] and big data [8]. The demand for on-chip processing will be greatly increased.The traditional bus-based on-chip architecture causes severe communication conflicts, which becomes the bottleneck of developing SoCs.A new architecture is urgently constructed by decoupling Intellectual Property (IP) cores from a bus pipeline. Thus, industry and academia communities proposed the concept of Network-on-Chip (NoC) that applies computer networking into chip design [5]. NoC implements the separation of computing and communication resources at the low complexity of designing the interconnection architecture inside a multi-core processor [9-11].

While in NoCs, a very high local clock frequency leads to serious capacitance delay and signal crosstalk. According to the forecast report of ITRS, the local clock frequency of NoCs will rise up to 73GHz in 2020 [12]. In addition, with the increasing number of IP cores, both electrical switching rate and bandwidth provisioning become very limited, and the power consumption proportional to bit rate inevitably grows. Owing to bit-rate transparency, optical waveguides are power-efficient and have extremely high transmission capacity, which makes Optical NoC (ONoC) become a hot research topic.

There have been various types of 2-dimensional (2D) ONoC topologies, mainly including mesh [15], fat tree [16, 17] and so forth.However, with the increasing number of IP cores on a single chip, the wavelength conflict1A communication pipeline will be influenced by another if they share the same wavelength.will become worse in 2D ONoCs, thus leading to a poor network throughput. For this end, 3D ONoC was proposed [18-21] by using multi-chip integration. Aside from supporting intra-layer packet forwarding, 3D ONoC focuses on the communication among IP cores from different chip layers, i.e., inter-layer packet forwarding. It means that, when the wavelength conflict occurs in a chip layer, the relevant packet can be transferred to another layer for the purpose of avoiding this conflict.In addition, the 3D integration reduces the physical connection length between a pair of chip lines, resulting in short transmission distance and latency.

Most existing works took 3D mesh ONoC into account. It seems simple, but the advantage of other topology structures is underutilized. Compared with 3D mesh, 3D torus is the better option of mitigating wavelength conflicts due to its loopback characteristic.Moreover, researchers always deployed the same type of optical router in a 3D ONoC topology, which results in the waste of chip resources such as micro resonator and crossbar.

In this paper, we first propose 3D mesh and torus ONoC topology structures assorted with optical routers. Next, the routing algorithm is simplified so long as the packet integrity can be ensured, meanwhile, the connectivity, flexibility and deadlock-free are also considered. In our switching mechanism, the electrical layer controls the establishment and demolition of lightpaths, while the optical layer forwards packets at a considerable transmission speed.Finally, OPNET is used to simulate 3D mesh and torus ONoCs. Simulation results demonstrate that 3D integration has the advantages of reducing average delay and packet loss rate, and 3D torus ONoC has the better performance compared with 3D mesh scheme.

The authors design mesh and torus ON-oCs from the perspectives of topology,router, and routing module, with the help of 3D integration. The authors also summarize some future challenges with possible solutions, including microcosmic routing inside optical routers and highly-efficient traffic grooming in this paper.

Our contributions are summarized as follows.

● We design a novel 3D torus topology along with simplified inter-layer and vertical optical routers, in order to achieve a better performance.

● We establish 3D ONoC platforms based on the three-layer modeling of OPNET.

● We give new ideas and possible solutions for future works on microcosmic routing algorithm and traffic grooming.

The rest of this paper is organized as follows. Section 2 summarizes the existing ONoC researches. The structures of 3D mesh and torus topologies and the corresponding optical routers are designed in Section 3. The communication protocol and routing algorithm are in Section 4. Simulation results are shown in Section 5, and we summarize challenging issues with possible solutions in Section 6.

II. RELATED WORK

In this section, we make a general description of current ONoC researches in three aspects:topology, optical router and routing algorithm.

2.1 ONoC topologies

An ONoC topology consists of optical routers and waveguides. Each waveguide connects with a pair of routers. Every router is connected with the IP core that completes the packet forwarding between optical and electrical layers via Optical-Electrical-Optical (OEO)conversions. Here, one packet generates at an IP core. In optical layer, utilizing different topology structures produces various levels of wavelength conflicts. Classic 2D ONoC topol-ogies include mesh [15] and fat tree [16, 17].In 2D mesh ONoC topology shown in Fig.1(a), except from the routers in edges, each of others has four neighbours and an IP core.But this topology structure has a large average number of routing hops between two communication nodes. On the other hand, the relevant researchers proposed a Butterfly Fat-Tree(BFT) ONoC topology with two dimensions.In this topology, the waveguide close to the root has a larger bandwidth capacity, which mitigates wavelength conflicts and improves network throughput. However, a severe power loss will be caused by intersection waveguides. Since mesh and fat tree have their own advantages, it may overcome shortages above through integrating mesh with fat tree. As a result, Fat tree on Mesh (FoM) ONoC topology was presented [16]. As shown in Fig. 1(b), every four routers constitute a cluster, and each of them electrically communicates with others.Two roots per cluster have a larger number of transmission directions to select compared with ordinary ones. Thus, FoM has the fattree characteristic that mitigates wavelength conflicts and improves network throughput. It also has the lower power loss caused by intersection waveguides than that of a pure fat-tree topology, because it still has mesh property.However, FoM has a huge power consumption of electrical communications in clusters.In addition, there were some research works on 3D ONoC topology. At present, the most common 3D topologies include 3D mesh,3D cilium mesh and 3D BFT. Among which,the 3D mesh achieves the interconnection of layers based on TSV (Through Silicon Vias)technology, and it has been widely studied due to its well extendibility and symmetry. The authors in [19] proposed a 3D cubic-mesh ONoC together with an optimized partial crossbar-based optical router which saved 24.5%energy compared to 2D ONoCs. However, 3D torus is the better option of mitigating wavelength conflicts due to its loopback characteristic.

2.2 On-chip optical router

On-chip optical router is often abbreviated as optical router. An optical router includes modulator, optical waveguide, micro resonator and filter. The modulator is mainly responsible for loading electrical information on the optical signal at the transmitter side. The optical waveguide forwards packets between a pair of components inside the router. The component can be a micro resonator. When the micro resonator is in open state, the packets with the same resonance wavelength change their transmission direction; otherwise, they pass through this micro resonator. The micro filter is mainly used to filter out the local packet that has the same resonant wavelength at the receiver side. In view of design principles, the optical router should be adaptive to the ONoC topology structure. As for the 2D mesh ONoC topology shown in Fig. 1(a), a non-blocking 4×4 router was presented in [20], and it has four ports including east, south, west and north, which can be seen in Fig. 2(a). Each kind of port has an input and an output. In addition, this optical router totally has eight micro resonators and four optical waveguides.However, the communication pipeline between a pair of ports will be influenced by another if they share the same wavelength, i.e., wavelength conflict. As shown in Figs. 2(b) and 2(c), the non-blocking 4×4 and 6×6 routers were proposed in [22], and they were adaptive to the FoM topology shown in Fig. 1(b). As mentioned above, there exist root routers and ordinary routers in FoM. A 4×4 ordinary router is shown in Fig. 2(b) and it has four ports(two east ports, one south port, and one north port), each of them has an input and an output.The east port is connected with the IP core locating at the same cluster, but south and north ports connect adjacent routers. In Fig. 2(c), the root router has six ports including two north ports, two south ports, one east port and one west port. Similarly, each kind of port has an input and an output. However, the design of optical routers tailored to 3D ONoCs is still in its initial stage. The same type of router is uti-lized in existing 3D ONoCs [19] [23], which results in the waste of chip resources such as micro resonator and crossbar.

Fig. 1 Structures of classic 2D ONoC topologies

Fig. 2 Structures of classic optical routers

Fig. 3 Structures of 3D mesh ONoC topology and corresponding optical router

2.3 ONoC routing

An ONoC routing algorithm is used to determine the lightpath from a source router to a destination router, and the relative packets will be transmitted along the selected lightpath.As a common ONoC routing algorithm, the deterministic routing decides the lightpath based on pre-determined router addresses. In other words, the selected lightpath remains unchanged, even if the current network state changes. Deterministic routing algorithms mainly include dimension order routing and source node routing [24, 25]. The most extensively used one is dimension order routing,because it has a routing logic so that the packet is strictly in accordance with the sequence of dimension variation, i.e., free deadlock.According to the visual range, dimension order routing is further divided into macroscopic routing and microscopic routing. The macroscopic one finds the lightpath between a pair of routers, but it neglects the computation of the microcosmic path between a pair of components inside the router. Currently,the microcosmic routing has not been solved.Though routing algorithms have been extensively implemented in 2D ONoCs, the routing algorithm tailored to 3D ONoCs is the better option of mitigating wavelength conflicts,which is our focus.

III. DESIGN OF TOPOLOGY AND OPTICAL ROUTER

With the topology dimension conversion from 2D to 3D, the coordinate of optical routers will be changed from (X, Y) to (X, Y, Z). Our 3D mesh ONoC shown in Fig. 3(a) has three vertical chip layers, each of which is a standard 2D mesh with 3 rows and 3 columns. Each row and each column has three optical routers,in order to implement the intra-layer packet forwarding. The IP core, where the packet is generated, connects with the optical router via OEO interfaces. Here, the OEO interface converts an optical signal into an electrical one and vice versa. Every optical waveguide between a pair of routers is bidirectional.

As shown in Fig. 3(b), we adopt a 7×7 full-connected crossbar as the optical router of 3D mesh ONoCs. There totally have 49 micro resonators, 14 optical waveguides, and 7 ports including (e)-injection, east, south,west, north, up and down. Among which, up and down ports support the inter-layer packet forwarding. We also propose a 3D torus ONoC topology structure in Fig. 4(a). The first and last optical routers at a row/column are connected by using an optical waveguide, and the second-layer optical router is additionally equipped with a vertical optical router for the inter-layer packet transmission. Most importantly, different from our 3D mesh ONoC that includes homogeneous routers, in this 3D torus ONoC, the vertical optical router is not required to connect with IP cores. Thus, compared with the intra-layer router shown in Fig.4(b), we design the simpler structure for the vertical router in Fig. 4(c).

Fig. 4 Structures of 3D torus ONoC topology and corresponding optical routers

Our 3D mesh and torus ONoC topologies both have two partially overlapped networks.One is the electrical network controlling switching fabrics and forwarding control signals. This electrical network includes electrical links, a cache unit and a control unit.Another is the optical network that has optical waveguides and routers. This optical network is mainly responsible for the packet transmission. Two networks share the same interconnection but use different wavelengths, in order to distinguish packet information and control signaling.

IV. COMMUNICATION PROTOCAL AND ROUTING ALGORITHM

Figure 5 explains the communication protocol tailored to our ONoC topologies based on the optoelectronic interconnection. In essence,this protocol represents a collaborative working process of optical and electrical networks.At a source router, the IP core generates a communication demand, and then it sends a setup packet to the destination router. The setup packet includes destination router address and routing-related control signaling, and this packet will be transmitted along the path determined by our routing algorithm presented later. Next, the destination router returns an ACK packet back to the source router along the reverse path in the electrical network. Finally, the data packet is transmitted along the lightpath in the optical network, once ACK packet arrives at the source router. When the data packet forwarding is completed, the reserved path will be released.

According to the dimensionality of 3D ONoC topologies, XYZ order routing is described as follows.

The packet walks first in X dimension, and we compare the X coordinate values of the current and destination routers. If these two coordinate values are not equal, forward this packet in the direction that the absolute value of X coordinate is decreasing.

Once the current and destination routers have the same X coordinate, we compare the Y coordinate values of these two routers. If these two coordinate values are not equal, forward this packet in the direction that the absolute value of Y coordinate is decreasing. Once the current and destination routers have the same X coordinate and the same Y coordinate,we compare the Z coordinate values of these two routers. If these two coordinate values are not equal, forward this packet in the direction that the absolute value of Z coordinate is decreasing. Otherwise, this packet arrives at the destination router.

V. SIMULATION RESULTS AND ANALYSIS

We establish the simulation platform of 3D mesh and torus ONoCs by using the three-layer modeling mechanism in OPNET [26]. More specifically, we first establish a network model according to the topology structure, then design a node model according to port number and router function, and finally construct a process model for each function module in the node model.

Fig. 5 Communication protocol for ONoC routing

Fig. 6 Network models for 3D mesh and torus ONoCs

Fig. 7 Node model for 3D mesh and torus ONoCs

5.1 Network model

Figures 6(a) and 6(b) demonstrate 3×3×3 mesh and torus ONoC models, where the blue polygon represents the node including an optical router and an IP core. In Fig. 6(b), light blue polygons are inter-layer optical routers in our 3D torus, while dark blue ones are vertical optical routers. The 3D coordinate (X, Y, Z)is used to identify nodes. After establishing a complete network model, the processor generates packets independently, and the time interval obeys an exponential distribution. We strictly define the sending rate of packets, the length of control signaling, the number of IP cores, and the capacity of optical waveguides.On the basis of this, node and process models are constructed in sequence.

5.2 Node model

Figure 7 describes the implementation of the node model by using OPNET node editor.The source module generates packets and the sink module destroys received packets after executing the data processing. There are six pairs of receiver and transmitter in the node model, i.e., bidirectional connections with surrounding nodes in six directions (east, west,north, south, up and down). As an example of the node_19 in Fig. 6(a), the eastern surrounding node is node_20 and the up surrounding node is node_22. For the vertical router node in Fig. 6(b), only three pairs of receiver and transmitter are used to connect with up, down and eastern/western surrounding routers while others are idle.

5.3 Process model

Figure 8 describes our process model. Among which, PROC denotes a packet processing state. A control flow will enter into this state when a self-interruption occurs. In the routing process, we determine a self-interruption interval according to the simulation time, and we start to route all packets in sub-queues. If more than one packet has the same routing result, the same port is required to send these packets. As a result, only one packet is sent every time, and the rest of packets are temporarily stored in a transmitter buffer.

The process function will poll seven subqueues (i.e., east, west, south, north, up, down and locality) during every period. Note that only three sub-queues are used for vertical routers. If a sub-queue has its own packets, it checks whether these packets are received or transmitted to the next-hop, according to our routing algorithm and the address of destination and current nodes. In addition, the packet priority depends on the polling order of seven sub-queues from sub0 to sub6. The detailed working process of PROC can be described as follows. If a sub-queue is not empty, we capture the packet locating at the head of ithsubqueue by using the sub-queue packet capture function op_ subq _ pk _ access. Next, we acquire the destination address coordinate (dest_x, dest_y, dest_z) of this packet. Through the coordinate comparison of the current and destination addresses, determine the next-hop port followed by XYZ dimension order routing principle. Finally, this packet is transferred from the current port to the next-hop port using the packet sending function op_pk_send(pksw, i). Here, i maybe east, west, south,north, up, down or locality.

Fig. 8 Process model for 3D mesh and torus ONoCs

5.4 Simulation results

In this paper, we consider two performance metrics: the average delay which is the average experience time of packets from generation to correctly received, and the normalized packet loss rate which refers to the ratio of the total number of packets over the number of misrepresented packets in the network.

Fig. 9 Simulation results of 2D and 3D mesh ONoCs when the link transmission rate is up to 1.28Gbit/s

Fig. 10 Simulation results of 2D and 3D mesh ON-oCs when the data packet generation time interval is 0.1

Due to the limitation of chip area, the cache capacity cannot be unlimitedly expanded. We assume all optical routers have the same cache size 200Mbit. Figures 9(a) and 9(b) show the simulation results of two performance metrics for 2D and 3D mesh ONoCs with homogeneous optical routers shown in Fig. 3(b) under different settings of the packet generation time interval, when the link transmission rate is up to 1.28 Gbit/s. In Fig. 9(a), when the packet generation time interval is large, the average delay of two mechanisms are both small, but the average delay of the 2D mesh ONoC is longer than that of the 3D mesh ONoC. When the packet generation time interval becomes small, the average delay of two mechanisms rises. Due to the advantage of 3D integration,the average network hop count is much less,hence the average delay of the 3D mesh ONoC is lower than that of the 2D mesh ONoC. In Fig. 9(b), we observe that with the reduction of the packet generation time interval, the packet loss rate has a rising trend. When the time interval becomes larger than 0.25, the packet loss rate of two mechanisms is roughly equal. In the 3D mesh ONoC, the node degree is very big, which means that a large set of solutions are in the routing mode for reducing the network congestion.

Therefore, the 3D mesh ONoC has the slightly lower packet loss rate compared with the 2D mesh ONoC. With the further increasing packet generation time interval, the packet loss rate of two mechanisms becomes stable without any noticeable fluctuations. In general,compared to 2D mesh ONoC results, the 3D mesh ONoC enhances up to 59.8% average delay and 22.5% packet loss rate.

Figures 10(a) and 10(b) demonstrate the simulation results of two performance metrics for 2D and 3D mesh ONoCs with homogeneous optical routers shown in Fig. 3(b) under different settings of the link transmission rate,when the packet generation time interval is 0.1. With the increasing link transmission rate,the performance metrics of two mechanisms degrades. The two performance metrics of the 3D mesh ONoC have a slight fluctuation.While for the 2D mesh ONoC, these metrics have an excessive fluctuation. With the increasing link transmission rate, the network congestion of the 2D mesh ONoC decreases,which means that the waiting time of packets become gradually reduced so that the average delay is changed into the transmission delay.Compared to 2D mesh ONoC results, the 3D mesh ONoC enhances up to 45.9% average delay and 34.8% packet loss rate.

To compare the performance of two 3D ONoC architectures as well as to demonstrate the superiority of our novel vertical routers,we implement 3D mesh with homogenous routers, 3D torus with homogenous routers and 3D rorus with vertical routers, respectively. Figures 11(a) and 11(b) show the analytical results of two performance metrics for 3D mesh and torus ONoCs under different settings of the packet generation time interval, when the link transmission rate is up to 1.28Gbit/s. In Fig. 11(a), we observe that the best performance of average delay is achieved for the 3D torus architecture with vertical optical routers, which keeps a low average delay under 0.02ms, and the average delay decreases to 26.1% and 48.6%, compared to 3D mesh and 3D torus with homogeneous optical routers, respectively. Moreover, when the packet generation time interval is large, the average delay of three mechanisms are small, but the average delay of 3D mesh and the 3D torus with homogeneous optical routers has a rapid growth, while the packet generation time interval is below 0.15. Due to the loopback feature, the average network hop count of the 3D torus ONoC is much less compared with the 3D mesh structure. Therefore, the average delay of the 3D torus ONoC is always lower.In addition, the utilization of vertical routers relieves the congestion between layers, thus leading to a low average delay. In Fig. 11(b),we observe that, three mechanisms have almost same packet loss rate when the packet generation time is greater than 0.3, the packet loss rate of two mechanisms with homogeneous optical routers sharply rises when the packet generation time is less than 0.2, and the packet loss rate of the 3D torus with vertical routers decreases up to 89.4% and 9.7%,compared with the 3D mesh and 3D torus with homogeneous optical routers. The 3D torus with vertical routers has the lower packet loss rate because its loopback characteristic further mitigates wavelength conflicts.

Figures 12(a) and 12(b) demonstrate the simulation results of two performance metrics for 3D mesh with homogeneous optical routers, 3D torus with homogeneous optical routers and 3D torus with vertical routers under different settings of the link transmission rate, when the packet generation time interval is 0.1. We can see that the 3D torus with vertical routers achieves the best performance: the average delay decreases to 18.8% and 16.6%,and the packet loss rate decreases to 55.5%and 24.9%, compared with 3D mesh and 3D torus architectures with homogeneous optical routers. The vertical optical router simplifies the structure and reduces the delay caused by routers. The results of Fig. 11(b) and Fig.12(b) also suggest that the use of vertical routers can reduce the packet loss rate.

Fig. 11 Simulation results of 3D mesh and torus ONoCs when the link transmission rate is up to 1.28Gbit/s

Fig. 12 Simulation results of 3D mesh and torus ONoCs when the data packet generation time interval is 0.1

Fig. 13 Directed graph model of a 2×2 crossbar

Fig. 14 (a) Directed graph model of optical router; (b) Entire directed graph model

VI. CHALLENGING ISSUES WITH POSSIBLE SOLUTIONS

6.1 Challenging issue 1

Our routing algorithm merely has the macroscopic visual range of finding a lightpath between a pair of routers, but it neglects the microcosmic path along the components inside the router. Due to the lack of precise information inside the router, the proposed routing strategy cannot well realize the optimization of ONoC performance. Hence, we have to realize a joint routing from perspectives of macro network and microchip. The possible solution is the design of directed graph models. As an example of Fig. 13(a), a 2×2 crossbar has two micro resonators, which makes the corresponding graph model have four vertexes X-, X+, Y- and Y+. A 2×2 crossbar is closed only when its micro resonators are in close state. In this case, the packet passes through these micro resonators without any transmission direction variations. Thus in Fig.13(b), we construct the graph model with the edges from X- to X+ and from Y- to Y+. If we open the micro resonator, the relative packet will change its transmission direction when it gets through this micro resonator. Thus in Fig. 13(c), the corresponding graph model with the edges from Y- to X+ and from X- to Y+ is determined. Next, if the router has an IP core (PE) and three 2×2 crossbars (SF, SX and SY), the corresponding graph model is shown in Fig. 14(a). Since SF, SX and SY have the same model, for simplicity, SX and SY are represented by two blocks. Finally, the entire directed graph model is obtained in Fig. 14(b),according to the information of our 3D ONoC topologies.

6.2 Challenging issue 2

The waveguide utilization rate will substantially drop when we transmit a small-size packet along a high-capacity lightpath. This motivates us to forward a group of small packets along a single lightpath, i.e., traffic grooming. In the future, with the objective of minimizing wavelength conflicts, we will generate the 3D wavelength-assignment matrix transformed from the 2D one obtained by row/column or source/destination-based strategy of wavelength assignment. Next, to implement traffic grooming, the entire graph model will be extended by adding virtual edges. In addition, the wavelength information will be assigned as edge weights according to the result of 3D wavelength-assignment matrix. Then,a novel and extensional graph model will be constructed to support traffic grooming. As an example of Fig. 14(b), a virtual edge will be added from the source router to the destination router after completing the joint routing of the current packet. If this virtual edge has enough residual bandwidth, the following packets can be further groomed into this virtual edge, thus leading to the improvement of waveguide utilization.

6.3 Challenging issue 3

The advantage and limitation of 3D IC and optical interconnection technologies make the emphasis of 3D ONoCs different to the traditional electrical NoC. Thermal sensitivity is the intrinsic characteristic of photonic devices,and it is the potential issue in ONoCs. Besides,silicon waveguide crossing and microresonator-based photonic switching elements are extensively used in ONoCs. The manufacturing of those optical on-chip components, and the heat generated by semiconductors will cause the reise in temperature and ultimately result in the deterioration of ONoC performance.The phenomena above can be attributed to the reliability issue of 3D ONoCs. In the future,we plan to take it into account.

ACKNOWLEDGEMENTS

This work was supported in part by the National Nat-ural Science Foundation of China(Grant Nos. 61401082, 61471109, 61502075,61672123, 91438110, U1301253), the Fundamental Research Funds for Central Universities (Grant Nos. N161604004, N161608001,N150401002, DUT15RC(3)009), Liaoning BaiQianWan Talents Program, and National High-Level Personnel Special Support Program for Youth Top-Notch Talent.

[1] D. Liu, Z. Cai, W. Wang, "Trends of Communication Processors", China Communications, vol.13,no.1, pp.1-16,2016.

[2] L. Benini, G. D. Micheli, “Networks on chips: a new SoC paradigm”, IEEE Computer Society, vol.35, no. 1, pp. 70-78, 2002.

[3] M. Ali, M. Welzl, M. Zwicknagl, “Networks-onchips: scalable interconnects for future systems on chips”, Proc. ECCSC, pp. 240-245, 2008.

[4] T. Schonwald, J. Zimmermann, O. Bringmann, et al, “Network-on-chip architecture exploration framework”, Proc. DSD, pp. 375-382, 2009.

[5] Y. Tao, L. Liu, S. Liu, Z. Zhang,"A Survey: Several Technologies of Non-Orthogonal Transmission for 5G",China Communications,vol.13,-no.10,pp.1-15,2015.

[6] C.Yu, W. Hou, Y. Guan, Y.Zong, P. Guo,"Virtual 5G Network Embedding in a Heterogeneous and Multi-Domain Network Infrastructure",China Communications,vol.13,no.10,pp.29-43,2016.

[7] X. Chen, R. Woogeun,Z. Wang,"Low Power Sensor Design for IoT and Mobile Healthcare Applications",China Communications,vol.13,-no.5,pp.42-54,2015.

[8] M.Yang, X. Zhou, J. Zeng, J. Xu,"Challenges and Solutions of Information Security Issues in the Age of Big Data",China Communications,vol.13,-no.3,pp.193-202,2015.

[9] J. Hu, R. Marculescu, “Energy-and-performance-aware mapping for regular NoC architectures”, Proc. TCAD, pp. 551-562, 2005.

[10] U. Y. Ogras, R. Marculescu, “Energy-and-performance-driven NoC communication architecture synthesis using a decomposition approach”,Proc. DATE, pp. 352-357, 2005.

[11] U. Y. Ogras, J. Hu, R. Marculescu, “Key research problems in NOC design: a holistic perspective”,Proc. H/SCSS, pp. 69-74, 2005.

[12] E. Rijpkema, K. Goossens, et al, “Tradeoffs in the design of a router with both guaranteed and best-effort services for networks on chip”, Proc.CDT, pp. 294-302, 2003.

[13] H. Gu, K. Chen, Y. Yang, Z. Chen, B. Zhang,"MRONoC: A Low Latency and Energy Efficient On Chip Optical Interconnect Architecture", IEEE Photonics Journal, vol.9, no.1, 2017.

[14] K. Chen, H. Gu, Y. Yang, D. Fan, "A Novel Two-layer Passive Optical Interconnection Network for On-Chip Communication",IEEE Journal of Lightwave Technology, vol.32, no.9,pp. 1770-1776,2014.

[15] V. D. Ngo, H. N. Nguyen, H. W. Choi, “Analyzing the performance of mesh and fat-tree topologies for network on chip design,” Proc. EUC, pp.300-310, 2005.

[16] P. P. Pande, C. Grecu, A. Ivanov, R. Saleh, “Design of a switch for network on chip applications,”Proc. ISCAS, pp. 217-220, 2003.

[17] S. L. Beux, I. O. Connor, G. Nicolescu, G. Bois, P.Paulin, “Reduction methods for adapting optical network on chip topologies to 3D architectures,” Microprocessors and Microsystems, vol.37, no. 1, pp. 87-98, 2013.

[18] H. Gu, J. Xu, “Design of 3D optical network on chip,” Proc. SOPO, pp. 1-4, 2009.

[19] M. Yang, W. Green, S. Assefa, et al, “Non-block-ing 4x4 electro-optic silicon switch for on-chip photonic networks,” Optics Express, pp. 47-54,2011.

[20] B. G. Lee, A. Biberman, K. Bergman, et al,“Multi-wavelength message routing in a non-blocking four-port bidirectional switch fabric for silicon photonic networks-on-chip,” Proc.OFC, pp. 1-3, 2009.

[21] P. Guo, W. Hou, L. Guo, "Designs of low insertion loss optical router and reliable routing for 3D optical network-on-chip", Science China Information Sciences, vol. 59, no.10, pp.102302,2016.

[22] Y. Ye, J. Xu, B. Huang, et al, “3-D mesh-based optical network-on-chip for multiprocessor system-on-chip,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 32, no. 4, pp. 584-596, 2013.

[23] J. Zhang, et al, “An improved distributed routing algorithm for Benes based optical NoC,” Proc.IPPR, 2011, pp. 1-6.

[24] Q. Cai, W. Hou, C. Yu, “Design and OPNET implementation of routing algorithm in 3D optical network on chip,” Proc. ICCC, 2014, pp. 1-4.

[25] J. Wang, H. Gu, Y. Yang, K. Wang, "Energy and Buffer Aware Adaptive Routing Algorithm for Network-on-Chip", Microelectronics Journal,vol.44, no.2, pp.137-144, 2013.

[26] http://www.bjkp.gov.cn/art/2009/7/15/art_6547_160038.html.