Optical Switching and Networking最新文献

The availability and reliability of optical backbone links are very important to ensure the efficient operation of the Internet. To address the issue of data loss due to optical link failures, there is a need for an optimal recovery strategy so that the traffic can be rerouted on a backup path to the destination. This paper builds on top of our prior research efforts (Yavary Mehr et al., 2022; Zhou et al., 2017) which introduced the concept of Resource Delayed Release (RDR) by adding a new state called ”idle state” which begins when the channel has completed carrying its services so that the next request can be carried immediately instead of waiting for a new channel to be established. While RDR improves the network performance by reducing the service provisioning time and blocking probability, it does not handle link failures which are quite common in optical networks. Therefore, enhancing RDR with protection strategies will make the network more reliable and thus we investigate this topic in this work.

In this paper, we evaluate four different protection methods for single link failure recovery in WDM networks (Path Protection (PP), Partial Path Protection (PPP), Segment Protection (SegP) and Link Protection (LP)) with two different routing approaches namely Shortest Path (SPath) and Greedy (G) algorithm under uniform and non-uniform traffic generated using real traffic traces collected from a local Internet Service Provider (ISP). Special attention while evaluating these protection strategies was paid to the optimization of the amount of remaining bandwidth. The performance evaluation of the network under uniform and non-uniform traffic was done over the NSFNet and COST239 topologies by employing the metrics of link and network utilization, Blocking Probability (BP), Bandwidth Blocking Probability (BBP), Recovery Time (RT) and Service Provisioning Time (SPT). Our results show that the PPP method performs the best in terms of reducing BP, BBP, and SPT compared with PP, LP, and SegP in all three topologies while utilizing RDR.

Distributed deep learning (DDL) systems strongly depend on network performance. Current electronic packet switched (EPS) network architectures and technologies suffer from variable diameter topologies, low-bisection bandwidth and over-subscription affecting completion time of communication and collective operations. We introduce a near-exascale, full-bisection bandwidth, all-to-all, single-hop, all-optical network architecture with nanosecond reconfiguration called RAMP, which supports large-scale distributed and parallel computing systems (12.8 Tbps per node for up to 65,536 nodes). For the first time, a custom RAMP-x MPI strategy and a network transcoder is proposed to run MPI collective operations across the optical circuit switched (OCS) network in a schedule-less and contention-less manner. RAMP achieves 7.6-171$\times$ speed-up in completion time across all MPI operations compared to realistic EPS and OCS counterparts. It can also deliver a 1.3-16$\times$ and 7.8-58$\times$ reduction in Megatron and DLRM training time respectively while offering 38-47$\times$ and 6.4-26.5$\times$ improvement in energy consumption and cost respectively.

Hybrid light-fidelity (LiFi) and wireless-fidelity (WiFi) networks (HLWNets) provide a promising solution for the future indoor wireless communications. This network structure faces the challenge of traffic congestion since LiFi links are prone to be blocked due to angular misalignment and path obstruction while WiFi has limited capacity. In this paper, a novel network structure that enables device-to-device (D2D) technology in HLWNets is considered. Then, traffic grooming (TG) for D2D-enabled HLWNets with massive light-path blockages is researched. By jointly handling mode selection, user pairing, and resource allocation, TG is formulated as a joint optimization problem. This can efficiently groom low-speed connections from WiFi onto high-capacity LiFi when massive light-path blockages occur, thus increasing network throughput. Next, a three-stage heuristic TG algorithm is developed to reduce the computational complexity required to solve the optimization problem. Finally, by simulation, the effectiveness of the proposed algorithm has been demonstrated. The simulation results indicate that the network throughput can be increased by up to 20% with the proposed algorithm. Besides, the proposed algorithm also has significant advantages in terms of Jain's fairness index and user satisfaction.

As Ethernet has a large bandwidth capacity, it is commonly proposed as a backbone for future intra-vehicle networks. However, satisfying the severe hardware reliability requirements of intra-vehicle networks while providing high-bandwidth and low latency by Ethernet may be costly. As a solution, we propose a novel optical intra-vehicle backbone network architecture that may have a lower cost and higher reliability in terms of hardware when compared to Ethernet. However, unlike traditional optical Ethernet architectures, only a single master node has transmitter laser diodes in the backbone of our architecture, so the gateway nodes cannot generate and send packets to the backbone links directly. As the gateways cannot inform the master node and request a slot when they have a new packet to send, a slot scheduling algorithm with polling is necessary to detect and transfer the new packets in the gateways, which may cause higher transmission delays compared to Ethernet. In this paper, we present our optical intra-vehicle backbone network architecture and propose two slot scheduling algorithms. We show that using a dynamic slot scheduling algorithm decreases packet delays when compared to fixed periodic slot scheduling in our architecture. We also evaluate the total delays including traffic shaping and processing delays in an optical TSN Ethernet backbone architecture as a reference. We show that the extra delays due to slot scheduling in our architecture may be negligibly low when compared with traffic shaping and processing delays.

Elastic optical networks (EONs) have been introduced to meet the demands of the rapidly growing Internet. These networks can efficiently keep up with the emerging bandwidth-hungry and highly dynamic services, and can support multicast services using techniques like the path, tree or subtree methods. A multicast wavelength-space-wavelength (M-WSW) network is a switching node architecture for EONs, which adopts the subtree method to support multicast connections. An M-WSW network consists of three node stages in which wavelength, space and wavelength switches are used, respectively. A nonblocking M-WSW network guarantees that any connection between a free input and a free output can always be realized, and studying the nonblockingness of a network has attracted much attention from researchers. Sufficient conditions, in terms of the number of middle space switches, for an M-WSW network to be strict-sense nonblocking (SNB) or wide-sense nonblocking (WSNB) were examined in an earlier study. It is known that SNB networks usually incur a higher hardware cost, for instance, the number of middle space switches, compared to WSNB, rearrangeably nonblocking (RNB), or repackably nonblocking (RPNB) networks. This paper studies the rearrangeability and repackability of M-WSW networks, and derives the sufficient and necessary conditions for an M-WSW network to be RNB (or RPNB). The results show that the derived sufficient conditions for being RNB (or RPNB) require significantly fewer middle switches for SNB and WSNB networks, and the RPNB results require fewer middle switches than those for RNB in most cases.

Dynamic multiple multicasts widely exist in several applications of optical network-on-chip. However, there is no good solution for routing and wavelength assignment for multiple multicasts in the mesh-based network. This paper proposes a new routing strategy based on a modified artificial fish swarm algorithm. The modified artificial fish model can support unicast and multicast in the mesh-based network. The routing and wavelength assignment for multiple multicasts can be solved based on this model. Then, we design a layer-based algorithm to assign wavelength for multiple multicasts, which can utilize wavelength and area resources more effectively. Simulation results show that our scheme works better than the other tree-based schemes regarding average communication latency and power consumption. In general, our modified artificial fish swarm algorithm provides a universal platform to study different aspects of routing and wavelength assignment in mesh-based ONoC.

We revisit spectrum allocation (SA), a fundamental problem in optical network design, and we explain that it can be modeled as a permutation problem. This new model eliminates spectrum symmetry, a property that presents a significant challenge to conventional spectrum allocation solutions. Accordingly, we develop parameterized first-fit (PFF), a new symmetry-free heuristic for the SA problem that has several desirable features: it explores a pre-defined subset of the solution space whose size is tailored to the available computational budget; it constructs this subset by sampling from diverse areas of the solution space rather than from the neighborhood of an initial solution; it finds solutions by applying the well-known first-fit (FF) algorithm and thus it can be deployed readily; its execution can be easily parallelized; and it is effective and efficient in finding good quality solutions.

We consider a joint optimization problem of optical network resources and virtual network function (VNF) placement for efficient content distribution. Current content distribution networks (CDNs) are tightly coupled with network function virtualization (NFV) technologies. A virtual CDN (vCDN) has been intensively investigated to efficiently cope with unpredictable traffic demand. In vCDN, CDN functions are virtually provided as VNF, and we can provide sufficient flexibility regarding the usage of compute resources under traffic demand changes. For the cost-effective CDN services, we must reduce the redundant usage of network resources while improving the efficiency of compute resources. However, a conventional optimal VNF placement technique in vCDN was focused on the efficiency of compute resources, and this can lead to an increase in network cost. To address this issue, we formulate the optimization problem as mixed integer linear programming and then propose a heuristic algorithm called a light-weight VNF placement algorithm in vCDN (LP-vCDN) for reducing network cost while effectively utilizing compute resources. We conducted intensive numerical experiments and demonstrated that LP-vCDN always found solutions of sufficient quality with practical computational overhead.

Cloud computing and web applications have further highlighted the need for powerful data centers with large bandwidth and low power consumption. In order to provide such a large bandwidth, today's electronic data centers require high power levels. Furthermore, communications and devices are not always used to their full potential. Optical networks can be a proper option for data centers, as they can offer large and variable bandwidth with lower power consumption compared to their electronic counterparts. In this regard, a dynamic all-optical network architecture is proposed in this paper, with bandwidth reconfiguration capability and low latency to reduce energy consumption. This proposed architecture, called Flat Ball, offers dynamic bandwidth utilizing passive optical devices with low power consumption and latency. Under realistic data center traffic scenarios, the latency of this network is up to 50% lower than its electrical counterpart, while providing a much higher throughput and consuming up to 60% less power.