AU2019100750A4

AU2019100750A4 - A low laser output power consumption double-layer structure ONoCs

Info

Publication number: AU2019100750A4
Application number: AU2019100750A
Authority: AU
Inventors: Ye SU
Original assignee: Southwest University
Current assignee: Southwest University
Priority date: 2019-07-10
Filing date: 2019-07-10
Publication date: 2019-08-15
Anticipated expiration: 2027-07-10

Abstract

Abstract A novel double-layer structure optical networks-on-chip (ONoCs) based on Virtural Clusters, named VCmesh. The structure mainly contains two communication modes : inter-cluster communication and non-cluster communication. Moreover, The Length Optimize Routing Protocol (LORP) is proposed, which is based on XY dimension routing protocol. Its function is to pick communication mode that shorter communication link to transmit optical packets. The structure and routing protocol in the present invention organically combines inter-cluster communication with non cluster communication, which not only improves the connectivity of network, but also shortens the communication length of long Manhattan distance communication task of the network. Compared with typical mesh-based ONoCs, the ONoCs applying present invention has better network performance in terms of end-to-end (ETE) delay and througput. Meanwhile, it can reduce the output power consumption, results in effectively saving the cost related with packaging, cooling solution, and system integration. (4,4 (1,) / 4 Fig. 1 4 - - 4x4 mesh -- 4x4 VCmesh 3,5 -4- 5x5 mesh -A-5x5 VCmesh -4-6x6 mesh 25 -A- 6x6 VCmesh 0.5 - - 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.02 Injection Rate Fig. 2 0.22 mesh 0.2 VCmcsh 4x4 5x5 6x6 Network scale Fig. 3

Description

EDITORIAL NOTE

- There are 4 pages of Description

2019100750 10 Jul 2019

ALOW LASER OUTPUT POWER CONSUMPTION DOUBLE-LAYER STRUCTURE ONoCs

TECHNICAL FIELD

The invention mainly relates to the field of on-chip optical network and multi-core processors.

BACKGROUND

With the explosive growth of demand for data processing capabilities, it is inevitable tendency that multi-processors systems-on-chip (MPSoCs) replaces singlecore chip architecture. Designing effective interconnection among processing cores to make full use of computing resources is one of the key factors determining the performance of MPSoCs. Optical networks-on-chip (ONoCs), which is an emerge multi-processors interconnection technology, has attracted more and more attention owing to its low power consumption, low lantency, and high bandwidth compared with traditional electronic interconnects. In addition, the wavelength division multiplexing (WDM) technology can be applied in ONoCs for further improving bandwidth and network throughput. At present, with the development of silicon-based photonics technology and the significant improve in photonic integration of commercial CMOS chips, ONoCs has great application prospects in the field of multi-core interconnection.

In the process of designing ONoCs, a large number of waveguide crossing and microresonator are employed to establish and maintain optical paths from the source to the destination for an optical signal. However, due to their material properties and manufacturing process, optical signal inevitably suffer from power loss. In order to guarantee the reliability of communication, the laser source has to output enough optical power to ensure that the signal can still meet the sensitivity of the photodetector after suffering a series of losses. But this situation will increase the on-chip power consumption, resulting in raising the financial burden in packaging, cooling solution, and system integration.

In order to solve this thorny problem, some effective solutions have been proposed at present. However, the most research focuses on innovation and optimization of routing algorithms, router architecture, and the manufacturing process of silicon-based optical device. Few people consider the design and optimization of network structure to reduce the insertion loss by shorten link length of ONoCs. Therefore, it is very significance that explore and design network structure for promoting the application prospect of ONoCs.

SUMMARY OF THE INVENTION

The present invention proposes a double-layer structure optical networks-on-chip

2019100750 10 Jul 2019 called VCmesh. The lower-layer network of the structure is composed of several Virtual-Clusters. Correspondingly, an all-pass optical router is plcaed in upper-layer, and its ports acount equal the number of Virtual-Clusters. Therefore, the optical signal can directly use the XY routing protocol for non-cluster communication in lower-layer network, or the inter-cluster communication through the upper-layer routing network. In order to better combine the two communication modes, the LORP is proposed to control the data transmission in the network. Its function is to select the communication mode with fewer hops according to the Manhattan distance between the source node and the destination node. As a result, the communication hop count that long Manhatten Distance communication task can be greatly decreased, thereby reducing insertion loss and saving the laser output power consumption. In addition, the invention has better network performance in terms of end-to-end (ETE) delay and througput.

SPECIFIC IMPLEMENTATION METHODS

In the following sections, the implementation of the present invention is further illustrated in detail with reference to the attached figures:

The present invention introduce an low laser output power consumption doublelayer structure ONoCs called VCmesh. The implementation of present invention mainly includes the Method of dividing Virtual-Clusters, the construction of network structure, and the design of LORP.

It is very key that division of Virtual-Clusters in lower-layer network of VCmesh ONoCs. Due to the number of Virtual-Clusters is equal to the number of ports of allpass router used in upper-layer, and the communication environment of routers will be worse with the increase of the number of ports. Therefore, in the process of dividing Virtual-Clusters, two conditions need to be satisfied: 1) Each node in lower-layer can reach the interface node within one hop. 2) Make sure that the interface node as least as possible. Next, an integer linear programming (ILP) method can realize the two suggestions.

In m x n mesh, some node will be pick as interface node, and assume that they belong to set S. For node (x,y) that belongs to S, it is recored as T(_xy) = 1. Otherwise, T^_xy^ = 0. Which can be expressed as:

_(1, (x,y)ES ^(x'^y) (ο, (x,y)tS (1)

In order to ensure that all node can research the interface node with one hop, it can be written as:

|% — x'l + |y — y’\ < 1 (2)

Σ T(x,y) - ¹ (3)

Further, requirement that reduce the interface node as least as possible can be expressed

2019100750 10 Jul 2019 as equation (4).

N = τηϊηΣΣΊ\_χ,_γ) (4)

Then, calculating coordinate position of interface node by Lingo as shown in Table 1.

Table 1 The results for optical interface distribution

Scale	coordinate
4x4	(1,2)(2,4)(3,1)(4,3)
4x5	(1,4)(2,1)(2,2)(3,4)(3,5)(4,2)
5x5	(1,3)(2,1)(2,5)(3,2)(4,4)(5,1)(5,4)
5x6	(1,2)(1,4)(2,6)(3,1)(3,3)(4,4)(5,1)(5,4)
6x6	(1,5)(2,1)(2,2)(2,3)(3,6)(4,4)(5,1)(5,2)(5,6)(6,4)
6x7	(1,3)(1,6)(2,1)(2,5)(3,2)(3,7)(4,4)(5,1)(5,6)(6,3)(6,6)
7x7	(1,3)(1,5)(2,1)(2,7)(3,4)(4,2)(4,6)(5,4)(6,1)(6,7)(7,3)(7,5)

As shown in Figure 1, this is a 4x4 VCmesh ONoCs, its underlying network is divided into 4 Virtual-Clusters surrounded by dotted lines, and all node can reach the interface node, which is indicated with the yellow circle, within one hop. Correspondingly, 4x4 all-pass optical router is configured in upper-layer network. Since these clusters are virtual, the architecture can implement routing data directly in lower-layer network using non-cluster communication. In addition, the network can also perform inter-cluster communication, that is, the signal first reaches the upperlayer routing network through the interface node, then is forwarded to the destination cluster, finally it is transmitted to the destination node. For example, in Fig. 1, the signal is sent from the source node (1, 1) to the destination node (4, 4). If non-cluster communication is selected, its communication link is (1,1)-(2,1)-(3,1)-(4,1)-(4,2)-(4,3)(4,4). Conversely, if the signal performs inter-cluster communication, the communication link is (1,1)-( 1,2)-(/V₂, N₃)-(4,3)-(4,4), where (N₂, N₃) indicates that the signal is travelled from 1-port to 3-port in upper-layer all-pass router. It can be clearly seen that the number of hops for performing inter-cluster communication is less than non-cluster communication. Similarly, if node (1, 1) communicate with node (1, 2), non-cluster communication has fewer hops, so it is a better choice.

Therefore, in order to reduce the transmission delay and insertion loss. In the present invention, the LORP is proposed to select a communication mode with a smaller number of hops. The pseudo code of the routing protocol is as follows:

Routing algorithm:

Input: the position of current node and its interface node : (x_c, y_c), (x'_c, y'_c); the position of destination node and its interface node: (χ^,γ^), (x'd>y'd)’ Arbitration:

Ijd-Jdl) + 2) output the path: (x_c, y_c) -> (x_d, y_d) with xyroutingalgorithm else output the path: (x_c, y_c) -> (x'_c, y'_c) with xy routing algorithm;

then (xj, y'c) (.^x'd> y'd) by upper-layer all-pass routing netwrok;

(x_d, y'd) (^xd> Jd) 'th xy routing algorithm;

Next, we verify the superiority of the invention by simulation.

When studying the structure of ONoCs, end-to-end (ETE) delay and network throughput are important indicators for measuring network performance. In this section, we use the network simulation software OPNET based on discrete event scheduling task to obtain the delay and throughput results under different network scales. The simulation results are shown in Fig. 2 and 3. VCmesh outperforms typical mesh in terms of latency and throughput. This is because VCMesh combined with LORP can efficiently select the primary communication mode based on the location of the source node and the destination node, which not only improves the connectivity of the network, but also reduces the latency of set-up link.

In addition, a large number of waveguide crossing and microresonator are integrated in ONoCs, which results in serious insertion loss of the optical signals during transmission. Therefore, it is necessary to ensure that optical signal can still meet the sensitivity of photodetector when it is received at the destination node. In the present invention, considering all feasible communication in ONoCs, we only guarantee the usefulness of the detector under the worst-case routing path. So the relationship between the minimum output power of a laser source (P;_aser), the sensitivity of the detector (Psensitivity) and the worst-case routing path loss (PL_WOrst-case) ^can be expressed as (5).

Plaser — P^worst-case 3” Psensitivity (6)

Fig. 4 shows the comparison of the required minimum output power of a laser source between the mesh-based ONoCs and VCmesh-based ONoCs at different network scales. It is clearly seen that VCmesh-based ONoCs has a huge advantage in laser power consumption compared with mesh-based ONoCs. And with the expansion of network scale, the minimum output power of a laser for mesh-based ONoCs grows sharply, but VCmesh based ONoCs is increase slowly.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is the 4x4 VCmesh.

Fig. 2 is the ETE-delay versus injection rate under different network scale.

Fig. 3 is the saturation throughput comparsion under different network scale.

Fig. 4 is the minimum laser output power required for communication.

2019100750 10 Jul 2019

EDITORIAL NOTE

- There is 1 page of Claims only

2019100750 10 Jul 2019

Claims

Claim

The claims defining the invention are as follows:

1. A low laser output power consumption double-layer optical networks-on-chip (ONoCs) called VCmesh. The double-layer optical network includes two communication modes of inter-cluster communication and non-cluster communication, which strengthens the connectivity of network results in improving network performance. Moreover, it can utilize Length-Optimize-Routing-Protocol (LORP) to reduce the insertion loss of optical signals in long Manhattan distance communication, thus saving the output power consumption of lasers.
2. Double-layer network structure (subject to claim 1)

In lower-layer network of VCmesh, mesh topology is dvided into several VirtualClusters through an integer linear programming (ILP) method. Each cluster contains an interface node, and they are connected by a all-pass optical router in upper-layer. Since these clusters are imaginary, the communicate ability is completely preserved in lowerlayer network. Moreover, the structure can also rely on the upper-layer routing network to transmit information.
3. Partition of Virtual-Clusters (subject to claim 1 and 2)

An integer linear programming (ILP) method for partitioning Virtual-Clusters in mesh. In m x n mesh, setting the set of interface nodes as S. If node (x,y) belongs to interface node, it is recorded as 7(_xy) = 1, otherwise 7(_xy) = 0, as shown in equation (1).

(1, (x,y)ES to, (x,y)eS

Assume that the location of the current node is (%', y'), and ensure all node can reach the interface node within one hop. Which can be express as equation (2) and (3).

|% -x'| + |y -y'l < 1(2)

ZT_(x._y)>l(3)

Eventually, decrease the number of Virtual-Clusters to reduce the complexity of alloptical routers used in upper-layer network. Which can be express as equation (4).

IV = minZZT(_X,y)(4)

Then, use Lingo to get the location of the interface node. Finally, with the interface nodes as the center, the nodes within one hop are grouped into a Virtual-Cluster.
4. Length-Optimize-Routing-Protocol (LORP) (subject to claim 1)

A routing protocol is designed that combines inter-cluster communication with noncluster communication. Its function is to select the communication mode with fewer hops from the two candidate communication modes for signal transmission. If the signal is routed using the inter-cluster communication, it is first transmitted to the upper-layer network through the interface node, and then reaches the destination cluster, finally to the destination node. If non-cluster communication is performed, the signal is routed directly to the destination node by XY dimension-routing in lower-layer network.

2019100750 10 Jul 2019

EDITORIAL NOTE

- There are 2 pages of Drawings only

2019100750 10 Jul 2019