Low Polling Time TDM ONoC With Direction-Based Wavelength Assignment

Transcription

1 Zhang et al. VOL. 9, NO. 6/JUNE 2017/J. OPT. COMMUN. NETW. 479 Low Polling Time TDM ONoC With Direction-Based Wavelength Assignment Bowen Zhang, Huaxi Gu, Kun Wang, Yintang Yang, and Wei Tan Abstract Optical network-on-chip (ONoC) is a promising technology for future on-chip many-core communications. ONoC can achieve high bandwidth density and low communication latency under low network cost, which leads to an energy-efficient network. Time-division multiplexing (TDM)-based ONoC aims to solve the congestion problem faced by many optical circuit switching (OCS)-based mesh ONoCs. But TDM-based ONoC still suffers from high network polling time, which causes high average network latency, especially under low network load. In this paper, we propose anoveltdmonocusingtorustopology and direction-based wavelength assignment, named TTWA. TTWA ONoC combines TDM communication strategy and torus network topology to solve the congestion problem under the heavy network load faced by OCS-based ONoCs. With the utilization of direction-based wavelength assignment and torus topology, we further propose a high-efficiency communication strategy to improve the adjacent inter-cluster parallel communications and decrease the total TDM slot number, which contributes to a smaller network polling time and lower average network latency. A novel optical router is simultaneously designed to meet the communication strategy in TTWA. The analyses and simulation results indicate that TTWA ONoC improves the polling time compared with other TDM-based ONoCs and achieves better network performance compared with the equivalent OCS-based ONoCs under various synthetic traffic patterns and realistic scientific applications. After the calculation, TTWA ONoC reduces the network insertion loss of OCS-based mesh ONoC 46.92% and that of OCS-based torus ONoC 31.75% and reduces the required laser power of OCS-based mesh ONoC 62.90% and that of OCSbased torus ONoC 40.66% under the worst-case situation. Index Terms ONoC; Optical interconnects; Realistic scientific applications; Silicon nanophotonics; TDM; Wavelength assignment. I. INTRODUCTION W ith the continuously increasing number of cores on the chip, future many-core processors require a high-performance yet energy-efficient and low-cost on-chip Manuscript received September 30, 2016; revised April 13, 2017; accepted April 26, 2017; published May 22, 2017 (Doc. ID ). B. Zhang, H. Gu ( hxgu@xidian.edu.cn), and W. Tan are with the State Key Laboratory of Integrated Service Networks, Xidian University, Xi an , China. K. Wang is with the School of Computer Science and Technology, Xidian University, Xi an , China. Y. Yang is with the Institute of Microelectronics, Xidian University, Xi an , China. network to support high-efficiency communication requirements [1]. Advancements in silicon-photonic technology have brought optical network-on-chip (ONoC) a bright future to solve the problems of high network delay, limited bandwidth, low scalability, and high power consumption faced by the electronic network-on-chip (ENoC) [2,3]. Many ONoCs have been proposed in recent years, such as Firefly [4], Phastlane [5], Corona [6], and ATAC [7]; they achieve huge bandwidth density with the use of wavelength-division multiplexing (WDM), which is an important feature of optical interconnection. Some researchers have made some improvements to achieve higher power efficiency, lower cost, and better network performance in the ONoC, such as silicon-photonic Clos networks [8], LumiNoC [9], FONoC [10], TorusNX [11], Square Root [11], etc. [12 16]. Mesh topology is preferred in ONoC designs because of its regular structure, simplicity, and divinable scalability in terms of network performance and cost. Torus topology [15,17] offers a lower network diameter and a higher network path diversity at the expense of having longer links. Torus is the better choice compared with mesh since the transmission power on photonic links is independent of the propagation length. The optical circuit switching (OCS) [18] mechanism is widely used in many mesh- and torus-based ONoCs on account of the lack of optical buffer in the ONoC. However, OCS-based mesh ONoC often suffers from high network congestion and low network utilization due to the blocking problem in the path setup stage and poor network path diversity. OCS-based torus ONoC solves several problems in OCS-based mesh ONoC, but is still faced with the deadlock trouble, which leads to the rapid deterioration of the network condition and limits the utilization of torus topology in ONoC. Time-division multiplexing (TDM) ONoC [19] is proposed to solve the congestion problem in OCS, especially under heavy network load. TDM avoids link competition by distributing different communication requirements into different timeslots. However, the network delay and put performance of TDM ONoC cannot satisfy the communication due to a long arbitration polling time and low network bandwidth. TDM is usually used and effective for long-distance traffic, but for short-distance traffic [20], the overhead limits the communication efficiency. The total number of timeslots /17/ Journal 2017 Optical Society of America

2 480 J. OPT. COMMUN. NETW./VOL. 9, NO. 6/JUNE 2017 Zhang et al. needs to be further reduced, as it has a serious impact on the performance of average network latency. In this paper, a novel TDM ONoC using torus topology and a direction-based wavelength assignment named TTWA is proposed to solve the congestion problem in previous OCS-based ONoCs and decrease the polling time in the previous TDM ONoC. The network diameter is decreased and the potential deadlock problem in torus ONoC is simultaneously avoided by combing the TDM communication method and torus topology. The polling time in the TDM communication is further decreased with the utilization of direction-based wavelength assignment between the neighbor cluster communications. In the optical network layer, a novel router is designed to meet the communication demands of TTWA. The network performance of an example TTWA ONoC is evaluated with OPNET and compared with the equivalent OCS-based ONoCs. TTWA ONoC can achieve better network latency and put performance, while the optical insertion loss and required laser power for TTWA ONoC are also optimized when compared with other ONoCs. The main contributions of this paper are summarized as follows: 1) The TDM communication method and torus topology are combined to decrease the network diameter, lower the total number of timeslots, avoid deadlock in the torus-based ONoC, and solve the congestion problem in the OCS-based ONoCs. 2) A high-efficiency communication strategy is proposed to improve the neighbor inter-cluster communication, decrease the polling time in TDM arbitration, and enhance the parallel performance of the network. 3) A novel optical router is designed to meet the communication demands in TTWA ONoC. With the utilization of an active broadband microring resonator (BMR) and a passive BMR, the optical router can transmit different optical signals automatically that do not influence each other. 4) The proposed ONoC is analyzed, evaluated, and compared with other equivalent ONoCs using several traffics. The insertion loss and required laser power are also considered in the paper. The rest of the paper is organized as follows. Section II introduces the related work. Section III presents the network architecture design and the novel router design for TTWA. Section IV introduces the communication strategy, gives the communication flow diagram, and shows the TDM arbitration result. Section V evaluates and analyzes the network performance between different ONoCs. The conclusion is given in Section VI. II. RELATED WORK transmission [21], etc. Network-on-chip (NoC) [22,23] has been proposed as a promising solution to satisfy the performance and design demands of many-core systems. Compared with SoC, NoC has better network communication performance and scalability and more effective parallelization. However, future many-core chips need to integrate a mass of components and be larger scale to satisfy the system performance requirements. Some factors, such as power consumption, signal interference, crosstalk, latency, and bandwidth, will limit the further development of NoC. The NoC component has gradually become the bottleneck of the future improvement for many-core processors. Many researchers have focused on the optical interconnect to solve the problem faced by NoC and come up with the ONoC design. Benefiting from WDM technology [24], ONoC can use optical interconnection to transfer multiple parallel optical streams of data signals via a single waveguide simultaneously, which leads to huge network bandwidth, yet lower latency. ONoC is suitable for large-scale system design because the signal transmission in the optical waveguide is distance-independent power consumption [25,26]. Furthermore, with the use of various other multiplexing technologies, such as TDM [15], space-division multiplexing (SDM) [27], and mode-division multiplexing (MDM) [28], ONoC has plenty of design and improvement potential. In recent years, researchers have come up with kinds of high bandwidth, low latency, low power consumption, high scalability, and reduced signal crosstalk ONoCs [4 11]. More and more ONoC architectures will be designed and utilized in the near future. Due to the lack of optical buffer on-chip, OCS is widely used in the designs of ONoC [18]. Figure 1 shows the communication process of OCS-based ONoC. A path from the source node to the destination node is reserved by the setup packet before the transmission of payload, which contributes to the non-blocking communication. Mesh topology is often used in ONoC design for its simplicity and predictable scalability in terms of performance and power consumption. However, OCS-based mesh ONoC suffers from long latency and low put due to the high path-setup payload pathteardown electronic control packet ACK optical signal The number of IP cores on a single chip is continuously increasing, and the traditional bus-based system-on-chip (SoC) will face the problems of high link congestion, clock skew, system out of synchronization, unreliable signal Fig. 1. Communication process diagram of OCS-based ONoC.

3 Zhang et al. VOL. 9, NO. 6/JUNE 2017/J. OPT. COMMUN. NETW. 481 contention probability under heavy offered load and low path diversity. The torus topology [15,17] utilizes a ring instead of the straight line in each row and column in the mesh topology, which leads to a smaller network diameter. Compared with OCS-based mesh ONoC, OCS-based torus ONoC has admirable network latency and put performance, because of fewer setup packets blocking the communication for OCS-based torus ONoC. However, torus-topology-based ONoC often faces the deadlock problem due to the special ring architecture. If the deadlock happens, the entire network would come into the paralyzed state rapidly. A lot of work focused on the utilization of TDM technology in optical communications has been performed for manycore processor networks [29 31]. Some researchers use TDM to arbitrate optical communication channels and propose a TDM ONoC. Each communication requirement in the ONoC is scheduled into different TDM slots by a genetic algorithm, which eliminates the setup stage in OCSbased ONoC and leads to non-blocking communication. However, the long polling arbitration waiting time leads to high latency and low network utilization especially under low network load. The total number of timeslots needs to be further reduced to obtain better network performance. III. TTWA ONOC DESIGN TTWA is a hybrid optoelectronic NoC that consists of an electronic layer and an optical network layer. n n clusters are distributed in the electronic layer uniformly. Each cluster has N c intellectual property (IP) cores, a waiting queue, an arbitration unit, and an electronic crossbar. As a result, the TTWA ONoC has N c n n IP cores in the electronic layer design. The optical network layer consists of n n optical routers. All the optical routers are connected with each other by optical waveguides using torus topology. Each cluster in the electronic layer connects with the corresponding optical router in the optical network layer via silicon via (TSV). The electronic layer and optical network layer can be designed, respectively, which contributes to the good network flexibility and the further development. Figure 2 shows the architecture of an example IP core TTWA ONoC. In the electronic layer, cluster architecture is designed to realize TDM arbitration and intra-cluster packet electrical transmission. All IP cores are placed in each cluster regularly; they are responsible for the generation, receiving, and processing of data packets. The arbitration unit can determine whether the communication request for the packet is intra-cluster or inter-cluster. It can also judge whether the inter-cluster communication is between the neighbor clusters or not. Furthermore, the arbitration unit can determine whether the inter-cluster communication packet is in the appropriate timeslot according to the TDM arbitration table, which will be described in more detail in Section IV. The electronic crossbar is employed to realize the communication between different IP cores in the same cluster. Inter-cluster communication packets will be transmitted and stored in the waiting queue by the electronic crossbar if they are not in the appropriate timeslot. The packets in the waiting queue will also be arbitrated by the arbitration unit to determine the timeslots that can be transmitted. Fig. 2. Architecture of an example IP core TTWA ONoC.

4 482 J. OPT. COMMUN. NETW./VOL. 9, NO. 6/JUNE 2017 Zhang et al. The optical network layer is responsible for inter-cluster communications. A novel router is designed to support the communication of TTWA. All the optical routers are interconnected by optical waveguides and form the torus topology, which increases the communication path, reduces network diameter, and contributes to the minimization of total timeslots. As Fig. 2 shows, the optical router consists of three parts: an optical switch, a control unit (CU), and an electronic-to-optical (E/O) converter and optical-toelectronic (O/E) converter. The optical switch has five ports namely, East, West, South, North, and Local. Each port has an input and an output. For the Local port, the ejection and the injection are regarded as the input and output, respectively. The design of the optical switching unit is non-blocking, which means the optical router can always find a path from one free input to another free output. It should be noted that the system does not have the communication requirement from the input to the output in the same port. Furthermore, the number of microring resonators (MRs) is optimized with the utilization of an X Y routing algorithm. A BMR [32] is employed in the router design and used to transmit the optical data signals with different wavelengths at the same time. As shown in Fig. 2, two types of BMRs are used inside of the router; one is an active element [25,33,34] (with green color, numbered 0), and the others are passive elements [32,35,36] (with other colors, numbered from 1 to 8). Active BMR needs to be controlled to change the resonant state for transmitting an optical signal. In TTWA ONoC, the active BMRs are controlled by the control unit (CU) in the optical router. By tracking the TDM table, the control unit is aware of the current TDM timeslot. At the beginning of each timeslot, the control unit sets the state of each active BMR in each optical switch according to the TDM arbitration table, which confirms the entire optical network state and contributes to a non-blocking communication. Figure 3 shows the parallel switching element and crossing switching element for active BMR. Passive BMRs are used to realize direction-based wavelength assignment. Passive BMR can couple the optical signal with the corresponding resonant wavelength when the signals pass by it. In the communications, packets with different neighbor destination clusters will be modulated with different wavelength groups adapted to different directions. Packets with the destination cluster east neighboring input drop Optical waveguide BMR ON state and resonant wavelength optical path add The rest condition optical path drop BMR input Fig. 3. Parallel switching element and crossing switching element for active BMR. input drop Optical waveguide BMR resonant wavelength optical path add Non-resonant wavelength optical path BMR to the source cluster are modulated with the wavelength group adapted to the RED BMR, south with the YELLOW one, west with the BLUE one, and north with the PURPLE one. In the direction-based wavelength assignment communication, a packet will be modulated to the corresponding wavelength group according to its destination, then injected into the optical network layer and transmitted by the optical routers to its destination. In this way, TTWA ONoC can achieve an active transmission between neighbor clusters, which means it does not require additional control actions. Figure 4 shows the parallel switching element and crossing switching element for the passive BMR. Each optical router has 12 active BMRs to realize TDM communication and 8 passive BMRs to realize the direction-based wavelength assignment between neighbor clusters. For an N c n n scale TTWA ONoC, the total required number of MRs is 12 8 n n 20n 2. The E/O O/E unit is used to convert the signal between electrical and optical. It connects with the electronic crossbar via TSV and the optical switch unit to realize the communication between the electronic layer and the optical network layer. In the design of TTWA ONoC, benefiting from combining the TDM method and torus topology, the optical network will have a smaller diameter, which means a smaller number of total timeslots, and contributes to the decrease of the polling time in the TDM arbitration. Meanwhile, the optical network layer in the TDM ONoC avoids the deadlock problem in the torus ONoC. The novel optical router can support the TDM method and the direction-based wavelength assignment communications at the same time. The structure of the optical router is simple and clear with few BMRs. With the utilization of BMRs, the optical router can support a huge network bandwidth. By using directionbased wavelength assignment, the TTWA ONoC can use an active transmission between neighbor clusters instead of the TDM method, which further reduces the total number of timeslots in TDM communication. IV. COMMUNICATION STRATEGY AND PROCESS In the TTWA ONoC, communications are divided into two parts, intra-cluster communication strategy and inter-cluster communication strategy. drop input Fig. 4. Parallel switching element and crossing switching element for passive BMR.

5 Zhang et al. VOL. 9, NO. 6/JUNE 2017/J. OPT. COMMUN. NETW. 483 In the intra-cluster communications, an electronic crossbar is used to transmit the packets between different IP cores in the same cluster. Packets generated from one source IP core will be directly exchanged to the destination IP core with no additional steps. In inter-cluster communications, the direction-based wavelength assignment is used in the communications between the neighbor clusters. Packets generated from the source IP core will be modulated by the corresponding wavelength group according to the direction of its destination cluster. The modulated optical signal is transmitted the optical router, demodulated, and transmitted to the destination cluster. The TDM communication is used in other conditions. At the beginning of each timeslot, the control unit sets the state of each active BMR in each optical switch according to the TDM arbitration table, which confirms the entire optical network state. Packets generated from the source IP core will be arbitrated by the arbitration unit in the cluster. If it is the correct timeslot, the packet will be modulated by the corresponding wavelength groups and sent to the destination cluster by the TDM communication method. If it is not the correct timeslot, the packet will be sent to the waiting queue and wait for the next arbitration until the arrival of the appropriate timeslot. With WDM technology, each IP core s data in one cluster uses diverse wavelength groups to be modulated and demodulated when communicating with the IP cores in one other cluster in the TDM communication method. Thus, each IP core in one source cluster can communicate with IP cores in one other destination cluster simultaneously in an appropriate timeslot. Packets will not interfere with each other when transmitted in the waveguide in the optical network layer. As a result, a superior modulation rate and tremendous network bandwidth can be achieved. Furthermore, the intra-cluster communication strategy, the direction-based wavelength assignment communication strategy, and the TDM communication strategy are parallelized. Packets in the same cluster, between the neighbor, and in other conditions can be transmitted at the same time, which contributes to a high-efficiency communication strategy. Figure 5 shows the flow diagram of the TTWA ONoC communication process. Several examples are given to further indicate the communication strategy and process of TTWA ONoC: IP1 has a communication requirement with IP2; the packet from IP1 is transmitted to IP2 via an electronic crossbar. IP0 has a communication requirement with IP5. First, the packet is transmitted to the OR0, modulated by the wavelength group adapted to the RED one, and injected to the optical network layer. Second, the signal is coupled by the RED 1 BMR in OR0, then transmitted to OR1. Third, the signal is coupled when passing by the RED 2 BMR in OR1, then ejected out from optical network layer. Finally, the signal is demodulated by the O/E unit and transmitted to IP5 successfully. IP3 has a communication requirement with IP63. In the appropriate timeslot, each control unit sets the necessary 0 BMRs to the coupled state in OR0, OR3, and OR15; then the packet from IP3 is sent to OR0, modulated by the E/O unit, and transmitted to IP63. The genetic algorithm is used to allocate each communication requirement into different timeslots appropriately. An example core TTWA ONoC is employed to show how it works. There are 4 4 clusters in the electronic layer, and each cluster has four neighbor clusters. Each cluster may have the communication requirement with the remaining 15 clusters. Among all the inter-cluster requirements, 16 4 communication requirements are between the neighbor clusters. The remaining inter-cluster communications will use the TDM communication strategy. By using the genetic algorithm, we get the minimum timeslot number N slot 12, and the result of the genetic algorithm includes all 176 TDM communication requirements correctly. Compared with the number of timeslots in the previous work [19], the number of timeslots is reduced from 18 to 12, which means TDM ONoC can reduce one-third of the polling wait time in the TDM arbitration. Table I shows parts of the TDM arbitration for the core TTWA ONoC. The complete table includes START No Extract the source and the destination location information of the packet Are the source and destination IP cores of the packet in the same cluster? Yes No Is the destination cluster in the neighborhood of the source cluster? Yes Transmit the packet to the destination IP core with TDM method Transmit the packet to the destination IP core with wavelength assignment method Transmit the packet to the destination IP core via electronic crossbar END Fig. 5. Flow diagram of the TTWA ONoC communication process.

6 484 J. OPT. COMMUN. NETW./VOL. 9, NO. 6/JUNE 2017 Zhang et al. TABLE I TDM ARBITRATION TABLE FOR A 4 4 TTWA OPTICAL NETWORK LAYER SC TS NULL 1 12 NULL 15 NULL 13 0 NULL the communication request and distribution of all 16 clusters in different timeslots. Each row represents a different timeslot, and each column represents a source cluster in the electronic layer. Assume that a timeslot starts from 1 to 12 and a cluster starts from 0 to 15. With a row and a column, TDM ONoC can determine in which timeslot the source cluster can communicate with which destination cluster. For example, the value of row 3 and column 2 is 5. This means in timeslot 3, source cluster 2 can communicate with destination cluster 5. NULL in Table I means there is no arranged communication between the source and destination clusters in this timeslot. The arbitration table is stored in the arbitration unit in the cluster and the control unit in the optical router. According to the arbitration table, the arbitration unit can determine whether the packet can be transmitted with the TDM communication method in a certain timeslot. At the beginning of each timeslot, the control unit sets the state of each active BMR in each optical switch by tracking the TDM arbitration table, which confirms the entire optical network state. The scalability of an N c n n TTWA is decided by two factors: 1) the number of cores in each cluster (N c ) and 2) the number of clusters (n n). With the increasing of N c, each cluster in TTWA can include more IP cores. With the increasing of n, TTWA can be extended as the optical network layer will connect more clusters, which means including more IP cores. With the changing of n, the TDM arbitration table needs to be updated at the same time. The new TDM arbitration result for different network scale can be obtained easily by using the genetic algorithm. The scalability and flexibility of an N c n n TTWA can satisfy the communication requirements. TABLE II SIMULATION PARAMETER CONFIGURATION Parameter Value Clock frequency (GHz) 1 Ack packet length (bit) 32 Path setup packet length (bit) 32 Optical fixed packet length (bit) 1024, 576 Timeslot length (ns) 16, 8 wavelength is 12.5 Gbps [37]. With the utilization of WDM technology, the modulation rate can reach 100 Gbps. End-to-end (ETE) delay and saturation put are utilized as the evaluation metrics for network performance. Figures 6 9 show the ETE delay versus offered load and saturation put comparison for TTWA ONoC, OCS-based mesh, and torus ONoCs under a uniform traffic pattern and hotspot traffic pattern under 1024 bit 16 ns timeslot length, respectively. Fig. 6. ETE delay versus offered load comparison under uniform traffic (1024 bit). V. EVALUATION AND DISCUSSION In this section, an example core TTWA ONoC performance is evaluated with OPNET and compared to the similarly configured 64-core OCS-based mesh and torus ONoCs. Uniform and hotspot traffic patterns are adopted in the simulation. Table II shows the parameter configuration for TTWA ONoC and OCS-based ONoCs. In TTWA ONoC communications, we use 16 ns for the timeslot under 1024 bit packet size. Sixty-four wavelengths are used in the ONoCs. The modulation speed for a single Fig. 7. Saturation put comparison under uniform traffic (1024 bit).

7 Zhang et al. VOL. 9, NO. 6/JUNE 2017/J. OPT. COMMUN. NETW. 485 Fig. 8. ETE delay versus offered load comparison under hotspot traffic (1024 bit). Figures 6 and 7 show that TTWA ONoC achieves lower latency and a higher network saturation point under heavy loads in a uniform traffic pattern. This is because TTWA ONoC combines direction-based wavelength assignment, torus topology, and the TDM method to decrease the network diameter, optimize the TDM slot, and lead to a non-blocking communication for the entire network. For the same network load, there will be more packet contentions under the hotspot traffic pattern. However, TTWA ONoC can still keep good network performance, but OCS-based mesh and torus ONoCs are badly influenced by the fierce contention, as Figs. 8 and 9 show. In Figs. 7 and 9, the saturation put of TTWA ONoC is compared with that of the OCS-based mesh and torus ONoC uniform traffic pattern and hotspot traffic pattern. Under uniform traffic, the saturation put for TTWA, OCS-based mesh, and torus ONoC is 0.66, 0.28, and 0.15, respectively. Under hotspot traffic, TTWA ONoC can keep an acceptable saturation put compared with other ONoCs. To further compare the network performance between different ONoCs, a smaller packet size (576 bit) is used in the simulation. Figures 10 and 11 show the ETE delay versus offered load comparison between different ONoCs under uniform and hotspot traffic patterns, respectively. The timeslot length for a 576 bit packet size is 8 ns. The simulation results show that compared with other ONoCs, TTWA ONoC can still have better network performance under a smaller packet size. To further deeply analyze the network performance of the TTWA ONoC, Netrace [38] is used in network performance evaluation. Netrace is a trace-based NoC evaluation methodology, which captures and enforces the dependencies between messages in the network from a full system. Netrace is usually utilized to evaluate 64-core systems. Nine realistic scientific application benchmarks from the PARSEC benchmark suite are used in the simulation. Latency (ns) TTWA ONoC OCS-torus OCS-meshe Offered Load Fig. 10. ETE delay versus offered load comparison under uniform traffic (576 bit) TTWA ONoC OCS-torus OCS-mesh 700 Latency (ns) Offered Load Fig. 9. Saturation put comparison under hotspot traffic (1024 bit). Fig. 11. ETE delay versus offered load comparison under hotspot traffic (576 bit).

8 486 J. OPT. COMMUN. NETW./VOL. 9, NO. 6/JUNE 2017 Zhang et al. The normalized execution speed and ETE delay are chosen to compare and analyze the network performance of the example (64) core TTWA ONoC, the similarly configured OCS-mesh, and OCS-torus ONoCs. Different scientific benchmarks are running in the network simulation model. The execution time and ETE delay for different ONoCs under different benchmarks are recorded. The execution time and ETE delay for OCS mesh are set to 1, and the execution time and ETE delay for TTWA ONoC and OCS torus are also normalized at the same time. The simulation result in Fig. 12 shows that TTWA ONoC has faster execution speed performance under almost all benchmarks compared with other ONoCs except blackscholes, vips, and x264. This is because the offered load of different applications in Netrace is quite low, and there is little network congestion in the network. Figure 13 shows the normalized ETE delay comparison between different ONoCs. TTWA achieves better network latency performance generally, while TTWA ONoC uses several methods to achieve a low polling time. Comparison results show that TTWA ONoC can solve the path setup congestion problem in the OCS-based mesh ONoC and the deadlock problem in the OCS-based torus ONoC, and optimize the network performance. Optical insertion loss and optical energy consumption are two other important issues in the ONoC designs. Optical insertion loss has an important influence on the scalability and reliability in the design of ONoCs. Meanwhile, optical insertion loss also affects the energy performance of an ONoC. In the following, the optical insertion loss and the required laser power are calculated for TTWA ONoC and the equivalent OCS-based mesh and torus ONoCs. The optical insertion loss for a data transmission from a source node to a destination node P Loss;Signal is computed as the following formula [39]: P Loss;Signal P MR;DP N D P MR;TP N T P B N B P W L l P IL;WC N WC P CR ; (1) where N D and N T represent the number of resonators passed by the optical signal in the drop and states, respectively. N B is the number of bending waveguides. L l is longest optical link length, which is approximately 4.5 mm in the optical network layer for each row and column. N IL;WC is the number of crossing waveguides in the optical network layer. Therefore, the required laser power for a signal transmission P Laser;Signal is calculated as follows: P Laser;Signal N w 10 P RE_ min P Loss;Data 10 P LE P CW : (2) Fig. 12. Normalized execution speed comparison under various PARSEC benchmarks using Netrace. N w is the total number of wavelengths used in the ONoC, which can be 64 [40,41]. P RE_ min is the minimum power required by the receiver, which is 22.3 dbm. P LE is the laser efficiency, which can be 30% [42]. P CW is the coupling coefficient, which is 90% [42]. Other parameters, their implications, and their values are shown in Table III. In the calculation of optical insertion loss and the required laser power, we consider the worst communication cases to meet the maximum communication requirements in each ONoC. After calculation, the P Loss;Signal and P Laser;Signal for the 64-core TTWA ONoC are approximately db and mw, while the P Loss;Data and P Laser;Signal for the equivalent OCS-based mesh and torus ONoCs are 9.18 db, mw and 7.14 db, mw, respectively, as Figs. 14 and 15 show. Compared with the OCS-based mesh and torus ONoCs, TTWA ONoC reduces the network insertion loss of OCS-based mesh ONoC 46.92% and that of OCS-based torus ONoC 31.75% and reduces the required laser power of OCS-based mesh ONoC 62.90% and that TABLE III PARAMETERS FOR OPTICAL INSERTION LOSS AND POWER Parameter Implication Value Fig. 13. Normalized ETE delay comparison under various PARSEC benchmarks using Netrace. P MR;DP Drop-port insertion loss (BMR) 1.3 db [43] P MR;TP Through-port insertion loss 0.01 db [4] P B Waveguide bending loss db [42] P W Waveguide propagation loss 0.5 db cm [44] P IL;WC Waveguide crossing insertion loss 0.12 db [42] P CR Coupling loss 0.6 db [45]

9 Zhang et al. VOL. 9, NO. 6/JUNE 2017/J. OPT. COMMUN. NETW. 487 Our future work will focus the power of the whole system and extend application scenarios for direction-based wavelength assignment, especially for the three-dimensional ONoC. ACKNOWLEDGMENT This work was supported by the National Natural Science Foundation of China (NSFC) Grant Nos , , and and the 111 Project under Grant No. B REFERENCES Fig. 15. Fig. 14. Insertion loss comparison between ONoCs. Required laser power comparison between ONoCs. of OCS-based torus ONoC 40.66% under the worst-case situation. TTWA ONoC has admirable insertion loss and laser power requirement performance. VI. CONCLUSION In this paper, TDM technology and torus topology are combined to solve the deadlock problem faced by previous OCS-based torus ONoCs and decrease the total number of TDM timeslots in previous TDM ONoCs. A high-efficiency communication method combines intra-cluster, directionbased wavelength assignment, and the TDM method is proposed to further optimize the polling time and enhance the parallel performance of the network. Active and passive broadband MRs are employed to design a novel optical router, which supports the TDM method and directionbased wavelength assignment simultaneously and offers a huge network bandwidth. TTWA ONoC achieves admirable network ETE delay, put, optical insertion loss, and optical energy consumption performance compared with the similarly configured OCS-based mesh and torus ONoCs in the network simulation. [1] L. Zheng, Y. Lu, M. Guo, S. Guo, and C. Z. Xu, Architecturebased design and optimization of genetic algorithms on multiand many-core systems, Future Generation Comput. Syst., vol. 38, no. 3, pp , [2] I. O Connor, Optical solutions for system-level interconnect, in Int. Workshop on System Level Interconnect Prediction, 2004, pp [3] G. Chen, H. Chen, M. Haurylau, N. A. Nelson, D. H. Albonesi, P. M. Fauchet, and E. G. Friedman, On-chip copper-based vs. optical interconnects: Delay uncertainty, latency, power, and bandwidth density comparative predictions, in IEEE Int. Interconnect Technology Conf., 2006, pp [4] Y. Pan, P. Kumar, J. Kim, G. Memik, Y. Zhang, and A. Choudhary, Firefly: Illuminating future network-on-chip with nanophotonics, in 36th Annu. Int. Symp. Computer Architecture, 2009, pp [5] M. J. Cianchetti, J. C. Kerekes, and D. H. Albonesi, Phastlane: A rapid transit optical routing network, ACM SIGARCH Computer Architecture News, vol. 37, no. 3, pp , [6] D. Vantrease, R. Schreiber, M. Monchiero, M. McLaren, N. P. Jouppi, M. Fiorentino, A. Davis, N. Binkert, R. G. Beausoleil, and J. H. Ahn, Corona: System implications of emerging nanophotonic technology, in Int. Symp. Computer Architecture, 2008, pp [7] G. Kurian, J. E. Miller, J. Psota, J. Eastep, J. Liu, J. Michel, L. C. Kimerling, and A. Agarwal, ATAC: A 1000-core cachecoherent processor with on-chip optical network, in 19th Int. Conf. Parallel Architectures and Compilation Techniques., 2010, pp [8] A. Joshi, C. Batten, Y. J. Kwon, S. Beamer, I. Shamim, K. Asanovic, and V. Stojanovic, Silicon-photonic Clos networks for global on-chip communication, in 3rd ACM/IEEE Int. Symp. on Networks-on-Chip (NoCS), 2009, pp [9] C. Li, M. Browning, P. V. Gratz, and S. Palermo, LumiNOC: A power-efficient, high-performance, photonic network-onchip, IEEE Trans. CAD Integr. Circ. Syst., vol. 33, no. 6, pp , [10] H. Gu, J. Xu, and W. Zhang, A low-power fat tree-based optical network-on-chip for multiprocessor system-on-chip, in Design, Automation & Test in Europe Conf., 2009, pp [11] J. Chan, G. Hendry, A. Biberman, and K. Bergman, Architectural exploration of chip-scale photonic interconnection network designs using physical-layer analysis, J. Lightwave Technol., vol. 28, no. 9, pp , [12] S. Liu, T. Chen, L. Li, X. Feng, Z. Xu, H. Chen, F. Chong, and Y. Chen, IMR: High-performance low-cost multi-ring NoCs, IEEE Trans. Parallel Distrib. Syst., vol. 27, no. 6, pp , 2016.

10 488 J. OPT. COMMUN. NETW./VOL. 9, NO. 6/JUNE 2017 Zhang et al. [13] C. Zhang, S. Zhang, J. D. Peters, and J. E. Bowers, Gbps fully integrated silicon photonic network on chip, Optica, vol. 3, no. 7, pp , [14] W. Tan, H. Gu, Y. Yang, M. Fadhel, and B. Zhang, Network condition-aware communication mechanism for circuitswitched optical networks-on-chips, J. Opt. Commun. Netw., vol. 8, no. 10, pp , [15] Y. Xie, W. Xu, W. Zhao, Y. Huang, T. Song, and M. Guo, Performance optimization and evaluation for torus-based optical networks-on-chip, J. Lightwave Technol., vol. 33, no. 18, pp , [16] A. B. Ahmed and A. B. Abdallah, PHENIC: Silicon photonic 3D-network-on-chip architecture for high-performance heterogeneous many-core system-on-chip, in 14th Int. Conf. Sciences and Techniques of Automatic Control and Computer Engineering (STA), 2013, pp [17] A. Shacham, B. G. Lee, A. Biberman, K. Bergman, and L. P. Carloni, Photonic NoC for DMA communications in chip multiprocessors, in 15th Annu. IEEE Symp. High- Performance Interconnects, 2007, pp [18] A. Shacham, K. Bergman, and L. P. Carloni, Photonic networks-on-chip for future generations of chip multiprocessors, IEEE Trans. Comput., vol. 57, no. 9, pp , [19] G. Hendry, J. Chan, S. Kamil, L. Oliker, J. Shalf, L. Carloni, and K. Bergman, Silicon nanophotonic network-on-chip using TDM arbitration, in 18th Annu. IEEE Symp. High Performance Interconnects (HOTI), 2010, pp [20] J. Kim, D. Park, T. Theocharides, N. Vijaykrishan, and C. R. Das, A low latency router supporting adaptivity for on-chip interconnects, in 42nd Annu. Design Automation Conf., [21] L. Benini and G. De Micheli, Networks on chip: A new paradigm for systems on chip design, in IEEE Design, Automation and Test in Europe Conf. and Exhibition, 2002, pp [22] W. J. Dally and B. Towles, Route packets, not wires: On-chip interconnection networks, 38th Annu. Design Automation Conf., Las Vegas, NV, June 18 22, [23] A. Ben Abdallah and M. Sowa, Basic network-on-chip interconnection for future gigascale MCSoCs applications: Communication and computation orthogonalization, in Joint Symp. Science, Society and Technology (JASSST), Dec. 4 9, 2006, pp [24] A. Banerjee, Y. Park, F. Clarke, H. Song, S. Yang, G. Kramer, K. Kim, and B. Mukherjee, Wavelength-division-multiplexed passive optical network (WDM-PON) technologies for broadband access: A review [Invited], J. Opt. Netw., vol. 4, no. 11, pp , [25] D. A. B. Miller, Rationale and challenges for optical interconnects to electronic chips, Proc. IEEE, vol. 88, no. 6, pp , [26] W. Bogaerts, P. Dumon, and D. V. Thourhout, Low-loss, lowcross-talk crossings for silicon-on-insulator nanophotonic waveguides, Opt. Lett., vol. 32, no. 19, pp , [27] K. Chen, H. Gu, Y. Yang, and D. Fan, A novel two-layer passive optical interconnection network for on-chip communication, J. Lightwave Technol., vol. 32, no. 9, pp , [28] L. W. Luo, N. Ophir, C. P. Chen, L. H. Gabrielli, C. B. Poitras, and K. Bergmen, WDM-compatible mode-division multiplexing on a silicon chip, Nat. Commun., vol. 5, 3069, [29] C. Qiao and R. Melhem, Reducing communication latency with path multiplexing in optically interconnected multiprocessor systems, IEEE Trans. Parallel Distrib. Syst., vol.8,pp , [30] X. Yuan, R. Melhem, and R. Gupta, Distributed path reservation algorithms for multiplexed all-optical interconnection networks, IEEE Trans. Comput., vol. 48, no. 12, pp , [31] C. Qiao and R. Melhem, Reconfiguration with time division multiplexed mins for multiprocessor communications, IEEE Trans. Parallel Distrib. Syst., vol. 5, pp , [32] A. Biberman, H. L. Lira, K. Padmaraju, N. Ophir, J. Chan, M. Lipson, and K. Bergman, Broadband silicon photonic electrooptic switch for photonic interconnection networks, IEEE Photon. Technol. Lett., vol. 23, no. 8, pp , [33] Z. Li, M. Mohamed, X. Chen, H. Zhou, A. Mickelson, L. Shang, and M. Vachharajani, Iris: A hybrid nanophotonic network design for high performance and low-power on-chip communication, J. Emerg. Technol. Comput. Syst., vol. 7, no. 2, article no. 8, [34] H. Gu, K. Mo, J. Xu, and W. Zhang, A low-power low-cost optical router for optical networks-on-chip in multiprocessor systems-on-chip, in IEEE Computer Society Annu. Symp. on VLSI, 2009, pp [35] A. Biberman, B. G. Lee, N. Sherwood-Droz, M. Lipson, and K. Bergman, Broadband operation of nanophotonic router for silicon photonic networks-on-chip, IEEE Photon. Technol. Lett, vol. 22, no. 12, pp , [36] X. Tan, M. Yang, L. Zhang, Y. Jiang, and J. Yang, A generic optical router design for photonic network-on-chips, J. Lightwave Technol., vol. 30, no. 3, pp , [37] Q. Xu, S. Manipatruni, B. Schmidt, J. Shakya, and M. Lipson, 12.5 Gbit/s carrier-injection-based silicon microring silicon modulators, Opt. Express, vol. 15, no. 2, pp , [38] J. Hestness and S. W. Keckler, Netrace: Dependency-tracking traces for efficient network-on-chip experimentation, in 3rd Int. Workshop on Network on Chip Architectures, 2011, pp [39] S. Koohi and S. Hessabi, All-optical wavelength-routed architecture for a power-efficient network on chip, IEEE Trans. Comput., vol. 63, no. 3, pp , [40] N. Kirman and J. Martínez, A power-efficient all-optical onchip interconnect using wavelength-based oblivious routing, in 15th Int. Conf. on Architectural Support for Programming Languages and Operating Systems, 2010, pp [41] K. Preston, N. Sherwood-Droz, J. S. Levy, and M. Lipson, Performance guidelines for WDM interconnects based on silicon microring resonators, in Conf. on Lasers and Electro- Optics (CLEO), May 2011, paper CThP4. [42] D. M. Vantrease, Optical tokens in many-core processors, Ph.D. dissertation, University of Wisconsin Madison, [43] A. Biberman, P. Dong, B. G. Lee, J. D. Foster, M. Lipson, and K. Bergman, Silicon microring resonator-based broadband comb switch for wavelength-parallel message routing, in 20th Annu. Meeting of the IEEE Lasers and Electro-Optics Society (LEOS), 2007, pp [44] M. Lipson, Guiding, modulating, and emitting light on silicon Challenges and opportunities, J. Lightwave Technol., vol. 23, no. 12, pp , [45] N. Kirman, M. Kirman, R. K. Dokania, J. F. Martinez, A. B. Apsel, M. A. Watkins, and D. H. Albonesi, Leveraging optical technology in future bus-based chip multiprocessors, in 39th Annu. IEEE/ACM Int. Symp. on Microarchitecture, 2006, pp