Optical Switching and Networking 22 (2016) 77–85

Contents lists available at ScienceDirect

Optical Switching and Networking

http://d1573-42

n Corr

journal homepage: www.elsevier.com/locate/osn

STorus: A new topology for optical network-on-chip

Xiuhua Li a, Huaxi Gu a,n, Ke Chen a, Liang Song b, Qingfen Hao b

a State Key Lab of ISN, Xidian University, Xi'an, Chinab Huawei Technologies, Co. Ltd., Beijing, China

a r t i c l e i n f o

Article history:Received 1 February 2015Received in revised form6 November 2015Accepted 15 April 2016Available online 20 May 2016

Keywords:Optical interconnectNetwork on chipTopologySubnetOptical router

x.doi.org/10.1016/j.osn.2016.04.00477/& 2016 Elsevier B.V. All rights reserved.

esponding author.

a b s t r a c t

Optical Network on Chip (ONoC) architecture is emerging as a promising candidate in multiprocessorsystems-on-chip (MPSoC) because of ultrahigh communication bandwidth, low latency and powerconsumption. The topology and layout of ONoC has a great impact on the network performance. Wa-veguide crossing in the topology would influence power consumption and crosstalk noise, which couldinfluence the size of ONoC directly. In this paper, a screwy torus topology (STorus) and a new opticalrouter are proposed. We divide an optical network into two subnets, and connect them with gateway.The diameter of STorus is half of Torus-based ONoC and almost a quarter of Mesh-based ONoC. To reducethe number of crossing, waveguide hierarchy has been employed. The simulation results show that thenew topology can improve throughput by 59.1% compared with traditional mesh under hotspot traffic,and 52.3% compared with traditional mesh under uniform traffic.

& 2016 Elsevier B.V. All rights reserved.

1. Introduction

WITH the growth of demand in communications, processingsystem with high performance is highly needed. Multi-ProcessorSystems-on-Chip (MPSoC) integrate many processors in a chip,and can dispose huge information [1,2]. The improvement inperformance achieved by the use of a multi-core processor de-pends on not only how efficiently the cores dispose information,but also the efficiency that cores communicate with each other[3,4]. An efficient on-chip communication architecture and net-work layout can help take full advantage of the computation re-sources offered by multiply processors. Therefore, on-chip inter-connects in the chip multiprocessor gradually change from businterconnects to networks-on-chip (NoC) [5]. However, with thegrowth of communication demand, metallic-interconnect-basedNOC was challenged because of limited bandwidth, long latency,electromagnetic interference and many other problems.

Optical Network-on-Chip (ONoC) rise with the development ofsilicon nanophotonics and CMOS-compatible components, and ithas generally become the consensus that future on-chip inter-connection network should also exploit as many advantages thatoptics can provide as possible [6–8]. Compared with electrical in-terconnect, optical interconnect have the advantage of providinglow latency, high bandwidth, and energy-efficient communication,which makes ONoC more attractive for future energy-efficient andhigh-performance on-chip interconnection [9]. Since chip size is

limited, ONoC faces with expandability problem. Then Three-Di-mensional Integrated Circuit architectures are proposed and itbecomes the focus of researchers. Such architecture consists ofelectrical layers, whose main role is to configure resource, and atleast one optical layer stacking upon the electrical layers. It makespossible the realization of mixed-technology eletronic-controlledONoCs. Based on the three-dimension integration technologies,researchers designed many three-dimension NoC architecturesbase on three-dimension network topologies to improve perfor-mance of NoC [10,11]. In this paper, a new ONoC architecture isproposed, and it employs three-dimension integration technolo-gies, which obtains all the strengths of ONoC and three-dimensionintegration technologies.

The network topology of ONoC consist of the layout and in-terconnection relationship of the devices on-chip, for instance, IPcores, optical routers, and waveguides [14]. Taking the mesh to-pology as an example, all the optical routers are connected viawaveguides in a grid shape, and each IP core is connected to a localoptical router. It is widely applied because of regular layout andsimple architecture [15,16]. However, the mesh topology [34] has abarrier, high diameter, which would result in a high packet hopcount. Under this condition, torus topology is employed [17,18].Through applying long wraparound waveguides, torus can reducethe diameter. However, it is imperfect, because the added wave-guides would result in additional waveguide crossings, which ag-gravate the optical loss. In order to utilize the silicon photonicstechnology fully, several new architectures were designed. Forinstance, the 2D-HERT, proposed by Koohi et al. [19], built upon

Optical router

Waveguide in layer B

Waveguide in layer A

Fig. 2. Subnet 0 topology.

X. Li et al. / Optical Switching and Networking 22 (2016) 77–8578

clusters of processing cores locally connected by optical rings.Different clusters are interconnected by global optical 2D rings. LeBeux et al. proposed a contention-free new architecture based onoptical network on chip, called Optical Ring Network-on-Chip(ORNoC) [20]. In this architecture, a wavelength is reused formultiple communications on a single waveguide, to take into ac-count the design constraints. However, most of these innovativeoptical architectures are based on traditional mesh and torus, andthey still face with loss, diamond and end-to-end delay problems.The new topology proposed in this paper can deal with theseproblems in some degree.

Compact optical routers are important components for ONoC[21]. Several silicon on-chip optical routers have been proposed[22–24]. Among these optical routers, 5�5 optical routers arewidely employed [12,13,25]. Generally, on-chip optical routers areadopted in a 2D network architecture, which would introducemany waveguide crossings, and cause a great power loss as a re-sult. In order to decrease the waveguide crossings and power lossas many as possible, a new optical router is proposed in this paper.

In this paper, we propose a new ONoC topology, Screwy Torusarchitecture (STorus), which takes advantage of the strengths ofONoC and three-dimension integration technologies. STorus alsoremits the hop count, diameter, and optical loss problems.Through employing two types screwy ring, a regular architectureturns up. The optical network is divided into two subnets witheach subnet consisting of a STorus architecture. With the help ofgateway, each IP core is capable of serving two subnets simulta-neously. It is proved that STorus is a compact topology because itleads to shorter path and less optical loss. In addition, we design arelevant 5�5 optical router, which is appropriate for Three-Di-mension architecture. The routers base on silicon micro-ring re-sonators (MRs) and hierarchy waveguides. Last but not least,Three-Dimension layout is adopted and part of waveguides arerelocated outside the optical network, which result in less wave-guide crossing and less optical loss.

2. Screwy torus architecture (STorus)

To exploit the properties of optical interconnect further, STorusarchitecture (Fig. 1) is designed, which is devised by reconfiguringthe connectivity of IP cores and routers, and replacing the wave-guides. For a N�N network, the number of IP core is N2, gatewayis N2/2, and optical router is N2. In addition, N is even. An 8�8

Fig. 1. Logic STorus architecture. (For interpretation of the references to color inthis figure, the reader is referred to the web version of this article.)

architecture is shown in Fig. 1. STorus architecture divides oneoptical network into two subnets. All the optical routers that locatein odd columns of odd lines and even columns of even lines belongto the same subnet, and all the other optical routers that locate ineven columns of odd lines and odd columns of even lines belong tothe other subnet. For an optical network with size 8�8, as shownin Fig. 1, all the blue nodes situate in one subnet, named subnet 0;and all the green nodes situate the other subnet, named subnet 1.The two IP cores that connected by one gateway belong to twodifferent subnets. The two nodes that connected by one gatewayare Upper and lower adjacent. For example, as shown in Fig. 7, thefirst line node share gateway with its nether node in the secondline, and the third lines node share with its nether node in thefourth line. Because one subnet overlaps another after rotating itclockwise by 90 degrees, it is possible to understand the entirenetwork by analyzing one subnet. Here we choose subnet 0.STorus topology is comprised of 5�5 optical routers and a seriesof interconnection. The interconnection is shown in Fig. 2. Eightnodes that with the same color are connected by a screwy ring.The rings just like number 8, which are shown as blue ring. Eightdifferent color nodes that located in the same diagonal are con-nected with nodes, which are located in the fourth line belowthem by a screwy ring. The rings like number 8 also, which areshown as green ring shown in Fig. 2. Each ring includes eightnodes, no matter green rings or blue rings. After moving thesubnet 0 topology of STorus in a counterclockwise direction by 45degrees, and moving up the shaded area, we get a N� (N/2) logictopology. Nodes that locate in one horizontal rank belong to onehorizontal ring, and nodes that locate in column j belong to onevertical ring with nodes that locate in columns jþN. As to an 8�8network, the logic topology is shown as Fig. 3. Waveguide thatconnects eight horizontal nodes are located in a same layer, whichare shown as blue lines in Fig. 3. Waveguide connecting eightvertical nodes locate in another layer which are shown as greenlines in Fig. 3.

To make it clear, we establish a coordinate system for STorustopology as shown in Fig. 3. We define the direction from left toright is direction x and from up to down is direction y. Eight nodesthat locate in a vertical ring share the same x-coordinate. The left

Fig. 3. Logic topology. (For interpretation of the references to color in this figure,the reader is referred to the web version of this article.)

X. Li et al. / Optical Switching and Networking 22 (2016) 77–85 79

four nodes in a horizontal ring share the same y-coordinates, so dothe right four nodes, and y-coordinate of the right nodes is equalto the left plus four. For example, node (i, j) belong to the samevertical ring with node (i, k), and belong to the same horizontalring with node (k, j) and node (k, jþ4). In addition, direction z isused to distinguish the two subnets, with nodes in one subnetshold the same direction z, for instance, the direction z of nodes insubnet 0 are 0, and in subnet 1 are 1.

The diameter of STorus is half of torus and almost a quarter ofmesh [20]. Hence, another main character of STorus is shorteraverage hop counts. For any two nodes of STorus communicating,packets just need to go through four optical routers in the worstcase. However, packets in the torus topology need go througheight optical routers, and mesh topology need fourteen in thesame network size.

3. New optical routers

Optical routers are the basic switching units of ONoC. Variantrouters of different sizes are proposed in literature [1,26]. How-ever, these router architectures are not so suitable for STorus, be-cause the packet in STorus topology must go through the verticalring first and then the horizontal ring, unless the source anddestination located in the same horizontal ring.

In this section, a new 5�5 optical router (Fig. 4) is proposed. Itconsists of four subsystems, switching fabric, control unit, sendingunit, and receiving unit. The sending units consist of ring resonatormodulator and transmitter. It is responsible for converting datasignals from the electrical domain into the optical domain andsends them to the optical network. Each ring modulator resonateswith a type of wavelength, and it is designed to have a free

Sending unit

R11

R12

R10

R9R8 R6

R7

R5R4

R1

R3

R2

Light direction

down

up

rightleft

Waveguide located in layer A

Waveguide located in layer B

Receiving unit Control unit

Broadband micro ring

Fig. 4. 5�5 optical router.

spectrum range (FSR) that equal to the channel spacing. The microring colored red and purple responses to a type of wavelengthbeing modulated. Switching fabric is built from the basic elementssuch as micro-ring resonator (MR) and waveguide. Broadbandrings are used to simultaneously switch all the wavelengths of aWDM signal. The control unit uses electrical signals to configurethe switching fabric, and control the state of broadband microrings, on or off. The receiving units receive optical data andtranslate optical data into electrical data streams. A passive ringresonator filter is allotted to each wavelength channel of the sys-tem, which is able to divert a specific channel from a WDM signal.

The optical router is optimized both for straight routing andbending routing. For the straight routing, the signal travels alongthe path and will not be coupled into any MR before it reaches thedestination. As for the bending routing, the signal goes throughthe minimum numbers of MRs in the optical router before chan-ging its direction and then travels to its destination. As limited bythe routing algorithm, the optical paths from left to up and down,and from right to up and down are not allowed. In addition, thesame port communication is forbidden. Hence there are sixteenpossible optical cases exist, as shown in Table 1. Under certaindevice technologies, the power efficiency of an ONoC is dom-inantly caused by the optical power loss that is encountered by thelight signal along the path. Although waveguide crossings inONoCs do not affect the bandwidth, it certainly cause more opticalpower loss and power consumption in packet transmissions, forexample each waveguide crossing introduces about 0.12 dB in-sertion loss to the optical signals passing through. In order tominimize power loss, the router architectures are designed tominimize the number of waveguide crossings. As shown in Fig. 4,each three parallel waveguides in the same colors are located inthe same layers, there are two layers in total. By this mean, wa-veguide crossing can be decreased to 2 inside the router.

4. The communication mechanism

In order to fulfill inter-core communication, optical circuitswitching mechanism [27] is employed. Generally, circuit-swit-ched ONoC consist of a high speed photonic circuit-switchednetwork and an electrical packet-switched control network. Themain function of photonic circuit-switched network is datatransmission. Electrical control network is used for path reserva-tion and control of optical routers [17]. The topology of electricalpacket-switched control network is same as photonic circuit-switched network and is overlapped by the photonic network. Anoptical path is reserved before data transmission, and electricalpacket-switched control network is used for optical path config-uration and maintenance. A single-flit path-setup packet would berouted in the electrical packet-switched control network for pathreservation. After confirming that the optical path is reserved, thesource sends payload data along it. Without buffering in inter-mediate routers, high-speed optical transmission is possible.

Since packets generated by the source may route to one of two

Table 1Micro-ring resonators application under sixteen paths.

Output ports

left right up down ejection

Input ports left – None – – R7right None – – – R5up R2 R6 – None R10down R4 R9 None – R11injection R3 R8 R1 R12 –

Fig. 5. The pseudo-code of the routing algorithm.

X. Li et al. / Optical Switching and Networking 22 (2016) 77–8580

routers that belong to two different subnets at the first hop, agateway is employed between the two routers that lie in twodifferent subnets to control the direction of signals with the helpof destination address (Fig. 1). The packets choose the router thatlocate in the same subnet with its destination node at the first hop,and travel in the subnet. When transmitting in a subnet, a basicprinciple is adopted. The packet would go through the vertical ringfirst, after arriving at the node that locates in the same horizontalring with its destination node, it turned to the horizontal directionuntil arriving destination. In a way, the transmit path in subnetcorrespond to YX order routing algorithm.

To deliver the optical packets to the right destination withinguaranteed latency and power consumption, a routing algorithm is

developed, as presented in Fig. 5. In general, before communica-tion, the source node will judge which subnet it delivers thepacket to with the help of gateway, and transmit to destination inthe reserved subnet. The signals generated by the source, will gostraightly to gateway. Then be separated, by z-coordinate of des-tination, into the subnet that its destination node located; Arrivingat the subnet, it would be sent in the vertical ring first until arriveat the same horizontal ring as destination node, then travel in thehorizontal ring, unless the destination node lie on the same hor-izontal ring with the source. Considering this basic principle, theremay be two turning nodes where packet turns from vertical ring tothe horizontal ring, which is shown as gray nodes in Fig. 3. In otherwords, there are two corresponding routing paths from the source

Optical router in subnet 0Fig. 6. Waveguide layout of subnet 0.

Table 2Parameters used for optical router loss.

Parameter LB LC Loff Lon Lmod Ldet

Value 0.005 dB/90° 0.04 dB/90° 0.005 dB 0.5 dB 1.2 dB 0.1 dB

Fig. 8. Optical loss of optical router.

X. Li et al. / Optical Switching and Networking 22 (2016) 77–85 81

to turning nodes in the vertical ring, one is up, and the other isdown. It chooses the turning node going through which the path isshorter. If the path that adopted one turning node is as short as theother, choose one randomly. For a packet from the source (0,1,0) tothe destination (3,3,0), the z-coordinate of destination is 0, so itwould be first routed into subnet 0. After recognizing that thedestination is not in the same horizontal ring with the source, itwould be transmitted in the vertical ring. Assuming that thepacket selects turning node (0,3,0), there are two paths from theturning node to the destination, one is left direction and the otheris right. The path is seven hops and five hops respectively. If itchooses another turning node (0,7,0), there are two paths from theturning node to the destination as well, one is left direction andthe other is right with the hop is three and nine, respectively. So itchooses turning node (0,7,0) and moves to the shortest path.

Optical router in subnet 1 Fig. 7. Optimized hierarchical waveguides in optical layers. (a) Waveguide layout of opreferences to color in this figure, the reader is referred to the web version of this articl

Shortest path is achieved under the special routing algorithm. Fora N�N STorus-based ONoC, the hop count could be limited underN/2.

Because cyclic dependence exists in STorus, deadlock will occur.Consequently, latency and throughput performance will be se-verely impaired. It is clear that deadlock will even occur when asingle-flit path-setup packet transmits in network and it is im-possible that the deadlock occur in the transmission of ACK packetand payload data. To avoid deadlock, a novel switching mechan-ism, Timeout retransmission (TRM), is proposed. A time-to-effect(TTE) filed is inserted in the path-setup packet with an initial valuesuch as 0. In addition, a threshold is set. When a path-setup packetstays in a router longer than threshold value, TTE will be changedinto 1. After recognizing TTE is changed, the router will create atear-down packet and send it back to the source, release the re-source that reserved by the transmission and inform source ofsending the packet once again after a certain time. The tear-down

Optical router in subnet 1 tical layer A and (b) waveguide layout of optical layer B. (For interpretation of thee.)

Table 3The performance comparison between STorus and three topologies.

Topology Scale (IP core) Diameter Router type MRs Waveguide crossings

Mesh [30] N�N 2N-2 5�5 25N2 26 N2

mesh (XY routing) [31] N�N 2N-2 5�5 (XY) 12 N2 11 N2

Torus [32,33] N�N N 5�5 25 N2 Z26N2

torus(XY routing) [31] N�N N 5�5 (XY) 12 N2 Z11 N2

Dmesh [9] N�N N-2 4�4 26N2�32Nþ32 6 N2�10Nþ10STorus N�N N/2 5�5 12 N2 o3N2

Fig. 9. Waveguide crossings in four topologies.

Table 4Parameters used for simulation.

Parameter Value

Optical network Packet length 1024 bitModulator rate 50 GbpsDetector rate 31.2 GbpsDetector sensitivity �5.59 dBm

Electrical control network Clock frequency 1 GHzPath-setup length 32 bitACK length 32 bitTear-down packet length 32 bit

X. Li et al. / Optical Switching and Networking 22 (2016) 77–8582

packet is a single-flit, and it is electrical. It routes through the samerouters with path-setup packet in the opposite direction, fromblocked router to the source. When the path that the path-setuppacket wants to go is blocked, tear-down packet would be send. Itis used to release the path that path-setup packet booked in orderto improve resource utilization.

Fig. 10. Performance with packet s

5. Network layout of the screwy torus-based ONoCs

In order to fully utilize the space and minimize the area ofnetwork, a 3D architecture, consist of two optical layers and oneelectrical layer, is proposed as shown in Fig. 1. We present theelectrical layer with two layers for the sake of displaying theconnection relations of each unit better. The top layer is opticallayer A, it include optical routers and part of waveguides. Thesecond layer is optical layer B, and it consists of part waveguides.The under two layers are electrical layers, and they are one layer infact. We display the electrical layer as two layers with the purposeof expressing the connection relations of gateways and electricalrouters better.

When integrating all the optical routers and waveguides intoone single optical layer, waveguide intersect and crossings wouldbe very serious. Consequently, optical power loss and crosstalk inONoCs would be enlarged. Additionally, the diameter and hop inmesh and torus network increase linearly with network size. Inorder to minimize the total number of waveguide crossings andimprove the performance of network, three design methods areadopted.

The first method is dividing the network into two subnets,which are named subnet 0 and subnet 1. Subnet 0 consist of all theblue optical router and waveguides connecting them as shown inFig. 2, and all the green optical router and waveguides connectingthem locates in subnet 1 which are shown on the top layer inFig. 1. After rotating subnet 1 clockwise by 90 degrees, we getsubnet 0. So it is possible to understand the optical network byanalyzing one single subnet, and here we choose subnet 0. Theoptical router in one subnet would not connect with the other.However, the IP cores in two subnets could communicate witheach other with the help of gateway. In addition, all optical routerslocate in one optical layer, no matter which subnet they belong to.

There are 64 waveguide crossings in subnet 0, and 49 of themare generated by three waveguides as shown in Fig. 2. Given this,we present the second method, rearranging the optical waveguide.Fig. 6 shows the layout of a subnet in 8�8 STorus-based ONoCafter rearranging the optical waveguide. Consider that most

ize 1024 bit for uniform traffic.

Fig. 11. Performance with packet size 1024 bit for hotspot traffic.

Fig. 12. Performance with packet size 2048 bit for uniform traffic.

Fig. 13. Performance with packet size 2048 bit for hotspot traffic.

X. Li et al. / Optical Switching and Networking 22 (2016) 77–85 83

waveguide crossings are produced by waveguide that connectnonadjacent optical routers, we rearrange them at the outside ofthe subnet. The layout maintains the connection property ofSTorus.

We could see that waveguide crossings in a single subnet areserious, even after layout rearrangement as shown in Fig. 6. If allthe optical routers connect like this, and the two subnets intersectwith each other, waveguide crossings would be more serious. Inorder to solve this problem, we employ the third method, wave-guide hierarchy. Waveguides are arranged into two optical layers.All the waveguide that connect the optical routers in vertical ringslocate in layer A, and the all the other waveguides that connect theoptical routers in horizontal rings in layer B, which shown as greenline and blue line in Fig. 7, respectively. After doing these, wave-guide crossing that produce by the two types of ring could beavoided.

6. Experimental results

In this section, we make analysis on the performance of theproposed network, including optical loss, topology analysis andsimulation result. Firstly, we calculate the loss of optical router andoptical network, and obtain the power for the off chip laser. Sec-ondly, in order to understand the topology better, a topologycomparison among STorus, mesh, Dmesh, and torus is made.Thirdly, a network simulator for STorus is built based on OPNET.We present the end-to-end (ETE) delay and throughput perfor-mance of STorus, and compare it with traditional mesh and meshadopting TRM.

6.1. Optical loss

Energy consumption is an important evaluation parameter of

X. Li et al. / Optical Switching and Networking 22 (2016) 77–8584

ONoC. Low energy consumption could reduce the cost. Loss is asignificant part of energy consumption. In order to analyze energyconsumption of optical network, we study loss of STorus first.

As presented in Eq. (6.1), the loss of optical router includeswaveguide bending loss, waveguide cross loss [28], transmissionloss of through port, and transmission loss of drop port. NB, NC, Noff,Non present the number of bending waveguide, cross waveguide,micro-ring of close-state and micro-ring of open-state, respectively.The parameters that used are listed in Table 2 [27,29]. As shown inFig. 8, the max loss, 0.59 dB appears when the packet transmitsfrom port up to port right, because this path passes through twocrossing waveguides and needs a micro-ring opened. The mini-mum, 0.02 dB turns up when the paths is from port right to left anddown to up, because these paths are straight without micro-ringopened and bending waveguide. The average is 0.416 dB.

( ) = × + × + × + × ( )Loss dB L N L N L N L N 6.1router B B C C off off on on

Based on the analysis of optical router's loss, optical loss of theONoC is calculated. As presented in Eq. (6.2), the loss of ONoC in-cludes modulator loss (Lmod), optical loss in optical routers (Lrouter),the loss between two adjacent optical routers (Li,iþ1), and detectorloss (Ldet). Because the propagation loss is small, it is not take intoaccount. The maximum of ONoC loss is 4.425 dB. Eq. (6.3) displaysthe power of laser. It (Losslaser) includes maximum of ONoC loss(Lossmax), waveguide number adopted (n), and detector sensitivity(S). Detector sensitivity is present in Table 4. The power of laser is1.845 dBm.

( ) = + + + ( )+Loss dB L L L L 6.2mod router i i det, 1

( ) = + + ( )Loss mW Loss lgn S10 6.3laser max

6.2. Topology analysis

In order to analyze the topology deeper, a detailed comparisonamong STorus, mesh, torus, and Dmesh is made, as shown in Ta-ble 3. We analyze each topology in terms of the diameter, totalnumber of MRs, and waveguide crossings. It can be seen that thediameter of the STorus is minimal among these topologies, whichis half of torus and almost a quarter of mesh, because two types ofrings are adopted. The number of MRs employed in STorus is equalto that of that of mesh and torus topologies. Besides, because ofthe layout of hierarchical waveguide, the number of waveguidecrossings in the STorus topology is minimal, and it decreases thewaveguide crossings of the Dmesh topology by more than thirtypercent and almost ninety percent of the mesh topology when thenetwork scale is 8�8. Fig. 9 presents the waveguide crossings inmesh, torus, Dmesh and STorus when the network size is 4�4,6�6 and 8�8. Profit from its topology and algorithm, the net-work obtain the minimal average hop over mesh and torus thatadopting XY routing algorithm under 8�8 network size. When thenetwork scale is 8�8, it decreases the average hop of torus to-pology by more than thirty percent and approximatively fiftypercent.

6.3. End-to-end delay and throughput

A network simulator for STorus is built based on OPNET. Wepresent the performance of STorus, and compare it with traditionalmesh and mesh that adopting TRM. The specific information usedfor simulation is listed in Table 4. Based on these, performanceanalysis of end- to-end delay (ETE delay) and throughput aremade.

Fig. 10 illustrates the result for 8�8 STorus, mesh, and meshadopting TRM under packet size 1024 bit for uniform traffic. It is

clearly that STorus exhibits higher throughput and lower ETE delaythan mesh with the same scale, because of two layers layout andshortest path algorithm. STorus reduce ETE delay by 44.4% thanmesh that adopting TRM, and even 52.8% than traditional mesh.Fig. 11 illustrates the result under packet size 1024 bit for hotspottraffic. STorus perform well, too. It improves the throughput by22.7% compared with mesh adopting TRM, and 59.1% comparedwith traditional mesh. As optical packet transmits at the speed oflight, the time it cost is very small. However, path-setup packettransmit at the electric network, the time it cost is huge. Whenpacket size increased, the information it transmitted at every re-served path increased. Consequently, the ETE delay would de-crease, and throughput would increase. When packet size in-creased, the improvement of performance is obvious still, whichshown in Figs. 12 and 13. As shown in Fig. 12, the performance ofSTorus and mesh are improved, but STorus is better as well. Itperforms a throughput improvement of 22.7%, and 52.3% com-pared with mesh adopting TRM and traditional mesh respectively.For hotspot traffic under packet size 2048 bit, STorus improvethroughput by 21.4% than mesh employing TRM, and 52.4% thantraditional mesh.

7. Conclusion

STorus is proposed as an effective topology to decrease thediamond, power loss, waveguide crossing and promote the net-work performance. By using subnets, its diameter is a quarter ofmesh, and half of torus compared with the same scale network. Inaddition, Hierarchical waveguide layout gives STorus fewer cross-ing, which lead to lower optical loss, and the average loss is 0.4 dB.These effective methods working with its suitable routing algo-rithm makes good end-to-end delay and throughput. STorus im-proves throughput by 59.1% compared with traditional mesh un-der hotspot traffic, and 52.3% compared with traditional meshunder uniform traffic.

Acknowledgment

This work is supported by the National Natural Science Foun-dation of China Grant nos. 61472300 and 61334003, the 111 Pro-ject Grant no. B08038, and the Huawei Innovation ResearchProgram.

References

[1] P. Paulin, The inexorable progression of parallelism in systems-on-a chip, Syst.Des. Front. 3 (2) (2006).

[2] L. Benini, G. De Micheli, Networks on chip: a new paradigm for systems onchip design, Des. Autom. Test. Eur. Conf. Exhib. (2002) 418–419.

[3] R. Ji, L. Yang, L. Zhang, Y. Tian, J. Ding, H. Chen, Y. Lu, P. Zhou, W. Zhu, Five-portoptical router for photonic networks-on-chip, Opt. Express 19 (21) (2011)20258–20268.

[4] H. Hossain, M. Akbar, M. Islam, Extended-butterfly fat tree interconnection(EFTI) architecture for network on chip, IEEE Pac. Rim Conf. Commun. Comput.Signal Process. (2005) 613–616.

[5] S. Le Beux, I. O’Conner, G. Nicolescu, G. Bois, P. Paulin, Reduction methods foradapting optical network on chip topologies to 3D architectures, Mocropro-cessors Microsyst. 37 (1) (2013) 87–98.

[6] P. Ho, G. Shen, S. Subramaniam, H.T. Mouftah, C. Qiao, L. Wosinska, Guesteditorial energy-efficiency in optical networks, IEEE J. Sel. Areas Commun. 32(8) (2014) 1521–1523.

[7] A. Shacham, K. Bergman, L.P. Carloni, Photonic networks-on-chip for futuregenerations of chip multiprocessors, IEEE Trans. Comput. 57 (9) (2008)1246–1260.

[8] Ian O’Connor, Optical solutions for system-level interconnect, SLIP (2004)79–88.

[9] Z. Chen, H. Gu, Y. Yang, L. bai, H. Li, A power efficient and compact opticalinterconnect for network-on-chip, IEEE Comput. Soc. 13 (1) (2013) 5–8.

X. Li et al. / Optical Switching and Networking 22 (2016) 77–85 85

re

[10] Y. Ye, J. Xu, B. Huang, X. Wu, W. Zhang, X. Wang, N.M., Zh Wang, W. Liu,Zh Wang, 3-D mesh-based optical network-on-chip for multiprocessor sys-tem-on-chip, Comput.-Aided Des. Integr. Circuits Syst. 323 (4) (2013) 584–596.

[11] A. Biberman, L. Lira, K. Padmaraju, Broadband silicon photonic electroopticswitch for photonic interconnection networks, IEEE Photonics Technol. Lett. 23(8) (2011) 504–506.

[12] Y. Li, H. Dai, G. Shen, S.K. Bose, Adaptive FEC-based lightpath routing andwavelength assignment in WDM optical networks, Opt. Switch. Netw. 14 (3)(2014) 241–249.

[13] L. Liu, Y. Yang, Energy-aware routing in hybrid optical network-on-chip forfuture multi-processor system-on-chip, J. Parallel Distrib. Comput. 73 (2)(2013) 189–197.

[14] H. Gu, J. Xu, W. Zhang, A low-power fat tree-based optical network-on-chip formultiprocessor system-on-chip, in: Proceedings of the Conference on Design,Automation and Test in Europe, pp. 3–8, 2009.

[15] H. Gu, J. Xu, Zh Wang, A novel optical mesh network on-chip for giga scalesystems-on-chip, in: Proceedings of the IEEE Asia Pacific Conference on Cir-cuits and Systems, pp. 1728–1731, 2008.

[16] Y. Xie, M. Nikdast, J. Xu, W. Zhang, Q. Li, X. Wu, Y. Ye, X. Wang, W. Liu, Crosstalknoise and bit error rate analysis for optical network on-chip, in: Proceedings ofthe ACM/IEEE Design Autom. Conf., pp. 657–660, 2010.

[17] A. Shacham, B.G. Lee, A. Biberman, K. Bergman, L.P. Carloni, Photonic NoC forDMA communications in chip multiprocessors, 15th Annu. IEEE Symp. High.-Perform. Interconnects (2007) 29–36.

[18] Y. Ye, J. Xu, X. Wu, W. Zhang, W. Liu, Nikdast, A torus-based hierarchical op-tical-electronic network-on-chip for multiprocessor system-on-chip, ACM J.Emerg. Technol. Comput. Syst. 8 (1) (2012), Article 5.

[19] S. Koohi, M. Abdollahi, S. Hessabi, All-optical wavelength-routed NoC based ona novel hierarchical topology, Int. Symp. Netw. -Chip (NOCS) (2011) 97–104.

[20] S. Le Beux, J. Trajkovic, I. O’Connor, G. Nicolescu, P. Paulin, Optical ring net-work-on-chip (ORNoC): architecture and design methodology, Des. Autom.

[22] A.Biberman,G.Lee,B.Sherwood-Droz,Broadbandoperationofnanophotonic

Test. Eur. Conf. Exhib. (DATE) (2011) 1–6.[21] H. Shabani, A. Roohi, A. Reza, M. Reshadi, N. Bagherzadeh, R. F. Demara, Loss-Awa

Switch Design and Non-Blocking Detection Algorithm for Intra-Chip Scale Photonic Interconnection Networks, IEEE Transactions on Computers. 65 (6) (2016) 1789 - 1801.

(12) (2010) 926–928.[23] D. Ding, B. Yu, D.Z. Pan, GLOW: a global router for low power thermal-reliable

interconnect synthesis using photonic wavelength multiplexing, Proc. AsiaSouth Pac. Des. Autom. Conf. (2012) 621–626.

[24] R. Ji, L. Yang, L. Zhang, Y. Tian, J. Ding, H. Chen, Y. Lu, P. Zhou, W. Zhu, Five-portoptical router for photonic networks-on-chip, Opt. Express 19 (21) (2011) 20258–20 268.

[25] N. Sherwood-Droz, H. Wang, L. Chen, B.G. Lee, A. Biberman, K. Bergman,M. Lipson, Optical 4�4 hitless silicon router for optical networks-on-chip(NoC), Opt. Express 16 (20) (2008) 15915–15922.

[26] M.J. Cianchetti, J.C. Kerekes, D.H. Albonesi, Phastlane: a rapid transit opticalrouting network, SIGARCH Comput. Archit. News 37 (3) (2009) 441–450.

[27] K.I. Kitayama, Y. Huang, et al., Torus-topology data center network based onoptical packet/agile circuit switching with intelligent flowmanagement, Light.Technol. J. 33 (5) (2015) 1063–1071.

[28] W. Ding, D. Tang, Y. Liu, L. Chen, X. Sun, Compact and low crosstalk waveguidecrossing using impedance matched metamaterial, Appl. Phys. Lett. 96 (11)(2010) 111114-111114-3.

[29] A. Biberman, K. Preston, G. Hendry, N. Sherwood-droz, J. Chan, J.S. Levy,M. Lipson, K. Bergman, Photonic network-on-chip architectures using multi-layer deposited silicon materials for high-performance chip multiprocessors,ACM J. Emerg. Technol. Comput. Syst. (JETC) 7 (7) (2011) 7:1-7:25.

[30] H. Gu, J. Xu, Z. Wang, A novel optical mesh network on chip for gigascalesystems on chip, IEEE Asia Pac. Conf. Circuits Syst. (APCCAS) (2008)1728–1731.

[31] L. Ni, P. McKinley, A survey of wormhole routing techniques in direct net-works, Computer 26 (1993) 62–76.

[32] J. Chan, G. Hendry, A. Biberman, K. Bergman, Architectural exploration of chip-scale photonic interconnection network designs using physical-layer analysis,J. Light. Technol. 28 (9) (2010) 1305–1315.

[33] K. Feng, Y. Ye, J. Xu. A formal study on topology and floorplan characteristics ofmesh and torus-based optical networks-on-chip, Microprocess Microsyst.,http://dx.doi.org/10.1016/j.micpro.2012.

[34] J. Chan, G. Hendry, A. Biberman, K. Bergman, Architectural exploration of chip-scale photonic interconnection network designs using physical-layer analysis,J. Light. Technol. 28 (9) (2010) 1305–1315.

router for silicon photonic networks-on-chip, IEEE Photonics Technol. Lett. 22

ÿÿÿÿÿÿ ÿÿÿÿÿÿ ÿÿÿ ÿÿÿÿÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿÿ ÿÿÿ ÿÿÿÿÿÿÿÿÿÿ ÿÿÿÿÿ ÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿÿ