To adapt to the intensive, bursty, and latency-constrained traffic from large language model (LLM) training and inference, scale-up networks are now facing tremendous challenges. Existing electrical packet switching (EPS) fabrics provide packet-level flexibility at the cost of high power consumption and long latency. Introducing optical circuit switching (OCS) in scale-up networks offers direct optical connections that can effectively reduce power consumption and latency, but OCS lacks the packet-level flexibility required by LLM inference (especially for mixture-of-experts (MoE) inference). To address these dilemmas, this work presents xSwitch, an optical-electrical-integrated (O-E-integrated) interconnect for scale-up networks, and demonstrates its effectiveness experimentally. Unlike the traditional hybrid-optical-electrical interconnects that usually place EPS and OCS in parallel, xSwitch integrates one port-count-reduced (defined by the O/E port ratio) EPS layer (ESL) on top of an OCS layer (OSL). Then, OSL can establish optical connections for xPU pairs with stable and intensive traffic demands, while bypassing the ESL for energy and latency reduction, and the dynamic and unpredictable traffic between xPUs can be provisioned by letting OSL forward it to the ESL. We prototype xSwitch with off-the-shelf components, validate its effectiveness with real-world MoE inference tasks, and also confirm its scalability with large-scale simulations. Our results indicate that for MoE inference workloads, xSwitch with a 2:1 O/E port ratio limits the average gaps to the full-EPS baseline to 1.96% in TTFT and 0.59% in TPOT.

Driven by emerging distributed computing and cloud-edge collaborative applications, metro and regional networks have experienced continuous surges in hub-and-spoke (H&S) traffic, posing great challenges to existing point-to-point network infrastructures. While coherent point-to-multipoint optical transceivers (P2MP-TRXs) present a more cost-effective solution for accommodating H&S traffic, existing P2MP networking paradigms fail to address the dynamic and asymmetric nature of traffic prevalent in such networks, and consequently, could lead to subpar resource utilization. In this paper, we fill this gap by investigating dynamic asymmetric subcarrier allocation and reconfiguration in drop-and-continue (D&C) optical networks. In particular, our approach aims at minimizing the operational cost of service provisioning by invoking reactive and coordinated connection consolidation as a remedy for the inability to accommodate traffic demands with in-service or newly activated P2MP-TRXs. We first devise an integer linear programming (ILP) model to solve the multi-objective optimization problem exactly. As the problem is proved to beNP-hard, we further develop a column generation (CG)-based approximation algorithm that can offer guaranteed optimality bounds within reasonable time, as well as a polynomial-time heuristic framework employing priority-queue-based progressive search. Extensive simulations verify the effectiveness of our proposal, demonstrating up to 43.3% reduction in bandwidth blocking ratio and 2.9% improvement in spectrum utilization compared with the state of the art.

In remote direct memory access (RDMA)-based remote procedure call (RPC) systems, using reliable connection (RC) transport mode with one-sided verbs might face challenges from excessive queue pairs (QPs) and per-client request regions when there is high client concurrency. This can cause cache thrashing on the RDMA network interface card (RNIC) and CPU and restrict server-side scalability. To address this issue, this paper proposesRdmax, a highly-scalable RC-based RPC system incorporating two innovations. First, it leverages QP multiplexity to serve multiple independent clients simultaneously with one RC QP, breaking the one-to-one connection constraint of RC. This also allows to post responses for different clients in one batch. Second, it uses in-network request dispatching to consolidate per-client request regions into a shared one, while still ensuring conflict-free client writes.Rdmaxrealizes these innovations with a programmable network, which manages the states of RC QPs and the request region and modifies relevant fields in RDMA packets accordingly. In a dummy RPC benchmark,Rdmaxachieves up to$10.4\times $higher throughput. For online transaction processing,Rdmaxoutperforms two state-of-the-arts based on client grouping and unreliable datagram schemes, improving throughput by up to 44.3% and 45.7%, while reducing 99% tail latency by up to 66.8% and 40.0%, respectively.

Recently, the growing of deep space explorations has attracted notable interests on interplanetary network (IPN), which is the key infrastructure for communications across vast distances in the solar system. However, the unique characteristics of IPN pose numerous unexplored challenges for interplanetary data transfers (IP-DTs), i.e., the challenges that existing schemes developed for Earth-based networks are ill-equipped to handle. To address these challenges, we first propose a novel distributed algorithm that leverages the Lyapunov optimization to jointly optimize the routing, scheduling and rate control of IP-DTs at each node. Specifically, our proposal adaptively optimizes the data-rate and bundle scheduling at each output port of a node, significantly improving the end-to-end (E2E) latency and delivery ratio of IP-DTs under a long-term energy constraint. Then, we further explore the heterogeneity of IPN to introduce limited state information exchange among nodes, and devise mechanisms for generating and disseminating state messages to facilitate timely adjustments of routing and scheduling schemes in response to unexpected link disruptions and traffic surges. Simulations verify the advantages of our proposal over the state-of-the-arts.

As deep space exploration missions grow in complexity, efficient data transfer in interplanetary networks (IPNs) becomes paramount. However, the vast distances, limited bandwidth, and dynamic nature of IPNs pose significant challenges for the routing and data scheduling of interplanetary data transfers (IP-DTs). To address these challenges, this work proposes a novel distributed, graph neural network (GNN) based multiagent deep reinforcement learning (DRL) approach that can jointly optimize the routing and scheduling of IP-DTs. Our proposal is based on the proximal policy optimization (PPO) framework along with the graph attention networks (GATs). We make the DRL agents for IPN nodes in each subnetwork around a celestial body learn and operate independently, for making intelligent routing and scheduling decisions to properly tradeoff between average end-to-end (E2E) latency and delivery ratio of IP-DTs while ensuring good scalability. Extensive simulations confirm that our proposal handles the routing and scheduling of IP-DTs much better than existing benchmarks. Further, by modifying the interplanetary overlay network (ION) software platform developed by NASA, we build a semi-physical IPN emulator based on Raspberry Pi boards, implement our proposal in it, and conduct experiments with real data transfers between IPN nodes. Experimental results verify that our proposal can work for practical IPNs without causing excessive overheads and prove its advantages.

Nowadays, the performance of data-center networks (DCNs) has become crucial for advancing large-scale computing applications. Hence, to improve the throughput, energy-efficiency and latency of DCNs, people are trying to replace the electrical packet switching (EPS) based spine switches with optical circuit switching (OCS) based ones. In an OCS-based DCN, topology engineering (TPE) is the key operation to dynamically reconfigure its inter-pod topology for accommodating traffic with optimized resource utilization. TPE consists of two highly-correlated steps, i.e., optimizing the target physical inter-pod topology of the DCN based on a traffic matrix, and planning the procedure of OCS reconfiguration such that hitless transition can be achieved. In this paper, we study how to optimize the two steps jointly to efficiently accelerate the hitless reconfiguration of an OCS-based DCN. We formulate a mixed linear programming model (MILP) to solve the joint optimization exactly. Then, to solve the problem time-efficiently, we propose an approach that optimizes TPE design greedily according to various metrics to minimize the number of stages required in hitless reconfiguration for TPE. Extensive simulations verify the effectiveness of our proposals and demonstrate their benefits over existing benchmark.

The dynamic traffic changes in an optical circuit switching based data-center network (ODCN) can make the utilization of its optical connections fragmented, degrading the efficiency of service provisioning in the ODCN. Consequently, connection defragmentation becomes imperative. However, consolidating traffic onto fewer connections may unilaterally add the risk of bandwidth contention and undermine ODCNs’ robustness against unexpected traffic bursts. To address this issue, we propose risk-aware connection defragmentation (RA-cDF), which explores the topology flexibility of ODCN to consolidate traffic such that the active optical connections through optical circuit switching (OCS) switches can be minimized together with the risk of future bandwidth contention on remaining connections. We formulate a mixed integer linear programming (MILP) model to address the RA-cDF problem exactly, followed by a heuristic to solve it time-efficiently. Extensive simulations confirm the effectiveness of our proposals.

Increasing demands for distributed machine learning (DML) have posed significant pressure on data-center networks (DCNs). This promotes the adoption of reconfigurable all-optical interconnects (AOI) in DCNs leveraging optical circuit switching (OCS) for better performance on throughput, energy efficiency, and data transfer latency. Despite their benefits, these OCS-based DCNs (ODCNs) still bear limited flexibility due to the larger switching granularity and longer reconfiguration latency of OCS. To address this issue, this work introduces in-network computing (INC) in an ODCN to realize P4INC-AOI, which can orchestrate INC and AOI to explore their mutual benefits for accelerating the training of DML jobs with less AOI reconfigurations. In the control plane of P4INC-AOI, we address the scheduling of concurrent DML jobs by formulating a mixed integer linear programming (MILP) model and proposing a time-efficient heuristic, to allocate multi-dimensional resources and configure AOI for minimizing the longest job completion time (JCT) across workloads. For the data plane, we extend existing in-network gradient aggregation schemes to accelerate DML jobs more efficiently. We first implement P4INC-AOI and verify its performance in a small-scale ODCN testbed, and further justify its effectiveness with large-scale simulations. Our experimental results demonstrate that compared with an ODCN without INC, P4INC-AOI not only cuts down AOI reconfigurations effectively but also reduces the average JCT of DML jobs in ResNet50 and VGG16 by$46.66\%$and$56.34\%$, respectively.

Deep reinforcement learning (DRL) enables autonomic optical networking by allowing the network control and management systems to self-learn successful networking policies from operational experiences. This paper proposes a transfer learning approach for effective and scalable DRL in optical networks. We first present a modular DRL agent design to facilitate retrieving and transferring relevant knowledge between tasks requiring different dimensions of network state data. In particular, we partition network state data into common states, which contain generic information critical to multiple tasks (e.g., the spectrum utilization on fiber links), and task-specific states that are only used by a specific task (e.g., the utilization of virtual network functions). Separate neural network blocks are employed to process different state data. Based on the modular agent design, a multi-task learning (MTL) aided knowledge transferring scheme is proposed. The proposed scheme trains an MTL agent that can master multiple tasks simultaneously and thus enables to learn and transfer better-generalized knowledge across tasks. We perform case studies on the proposed transfer DRL approach considering two scenarios, namely, knowledge transferring between routing, modulation and spectrum assignment (RMSA) tasks for different networks and knowledge transferring from RMSA tasks to anycast service provisioning tasks. The DRL designs for RMSA and anycast service provisioning, including the state, action, and reward formulations and the training mechanisms, are also elaborated. Performance evaluations under both scenarios show that the proposed approach can effectively expedite the training processes of the target tasks and improve the ultimate service throughput.

Accurate delay estimation is one of the enablers of future network connectivity services, as it facilitates the application layer to anticipate network performance. If such connectivity services require isolation (slicing), such delay estimation should not be limited to a maximum value defined in the Service Level Agreement, but to a finer-grained description of the expected delay in the form of, e.g., a continuous function of the load. Obtaining accurate end-to-end (e2e) delay modeling is even more challenging in a multi-operator (Multi-AS) scenario, where the provisioning of e2e connectivity services is provided across heterogeneous multi-operator (Multi-AS or just domains) networks. In this work, we propose a collaborative environment, where each domain Software Defined Networking (SDN) controller models intra-domain delay components of inter-domain paths and share those models with a broker system providing the e2e connectivity services. The broker, in turn, models the delay of inter-domain links based on e2e monitoring and the received intra-domain models. Exhaustive simulation results show that composing e2e models as the summation of intra-domain network and inter-domain link delay models provides many benefits and increasing performance over the models obtained from e2e measurements.

In this paper, we propose an incentive-driven virtual network function service chaining (VNF-SC) framework for optimizing the cross-stratum resource provisioning in multi-broker orchestrated inter-datacenter elastic optical networks (IDC-EONs). The proposed framework employs a non-cooperative hierarchical game-theoretic mechanism, where the resource brokers and the VNF-SC users play the leader and the follower games, respectively. In the leader game, the brokers calculate VNF-SC service schemes for users and compete for the provisioning tasks. While in the follower game, the users compete for VNF-SC services for jointly optimizing the resource cost and the received quality-of-service. We first elaborate on the modeling of the follower game, discuss the existence of Nash equilibrium and propose a mixed-strategy gaming approach enabled by an auxiliary graph-based algorithm to facilitate users selecting the most appropriate service schemes. Then, under the assumption that the brokers are aware of the principle of the follower game, we present the model for the leader game and develop a time-efficient heuristic algorithm for brokers to compete for the provisioning tasks. Simulations show that the proposed incentive-driven VNF-SC framework significantly improves the network throughput (i.e., $>4.8 \times$ blocking reduction) while assisting users and brokers in achieving higher utilities compared with existing solutions.

Software-defined elastic optical networks (SD-EONs) provide operators more flexibility to customize their optical infrastructure dynamically and adaptively, and network virtualization, i.e., infrastructure-as-a-service (IaaS), enables multiple tenants to share the substrate infrastructure efficiently. In this paper, we study how to provision virtual SD-EONs (vSD-EONs) with the correlated data and control plane embedding (χ-VNE) that considers the quality-of-service (QoS) of virtual control plane (vCP), i.e., availability and control channel latency. We propose a χ-VNE algorithm to solve the problem, design the network system to realize it, and accomplish proof-of-concept experimental demonstrations in an OpenFlow-based network testbed. Numerical and experimental results indicate that the proposed algorithm and system function well and can realize on-demand and reliable vSD-EON provisioning.

As multi-domain elastic optical networks (EONs) can enhance network scalability, extend service reach, and accommodate the inter-operability issues, it is very relevant to consider them in practical network operations. In this work, we study the problem of how to achieve energy-aware service provisioning in a multi-domain EON, where the optoelectronic regenerators only exist in border nodes. We consider dynamic lightpath requests and propose two algorithms, i.e., the greedy regenerator allocation (GRA) and the set-cover based regenerator allocation (STC), to realize the joint optimization of routing, modulation and spectrum assignment (RMSA) and regenerator allocation. The algorithms are evaluated with extensive simulations that use multi-domain EONs built with different regenerator placement strategies. The results verify that GRA and STC outperform the existing algorithm for energy-aware multi-domain service provisioning, and STC achieves the best performance in terms of both blocking probability and power efficiency.

Software defined networking (SDN) makes networks programmable and application-aware by decoupling network control and management (NC&M) from data forwarding and leveraging centralized NC&M to facilitate user-customized routing and switching. Inspired by these, this paper investigates how to realize the OpenFlow-controlled (OF-controlled) network orchestration that can facilitate efficient scalable video coding (SVC) streaming to heterogeneous clients. Specifically, we consider real-time SVC streaming and address the situation in which video sources reside in geographically- distributed servers and clients can join and leave the streaming services dynamically. We formulate this as a multi-source multi-destination manycast problem and realize the networking system with an OF-controlled SDN architecture. We first design the OF controller to enable efficient network operations. Then, we focus on solving the multi-source multi-destination SVC video manycast problem and design several algorithms. Initially, an integer linear programming (ILP) model is formulated to obtain the optimal solutions for small-scale problems. Next, we try to make the manycast algorithm suitable for practical implementation, and design two time-efficient heuristics. Simulation results indicate that the heuristics can provide close-to-optimal solutions. Finally, we build an OF network testbed that consists of OF switches, SVC video servers and clients, and perform SVC streaming experiments to demonstrate our design. Experimental results verify that the proposed scheme can allocate bandwidth intelligently and ensure high-quality video streaming. To the best of our knowledge, this is the first work that accomplishes experimental demonstration of OF-controlled network orchestration for adaptive SVC video manycast.

We propose three algorithms for dynamic pre-configured-cycle (p-cycle) configuration in the spectrum-sliced elastic optical networks (EONs) based on the optical orthogonal frequency-division multiplexing (O-OFDM) technology, aiming to provide 100% restoration against single-link failures. The first one (PE-p-cycle) configures the working path and p-cycles of a request together according to the protection efficiencies of the p-cycles. In order to reduce request blocking probability, we then propose to use the spectrum planning technique to regulate the spectra of working and protection resources and design two algorithms based on the protected working capacity envelope (PWCE) cycles (PWCE-p-cycle-SP) and Hamiltonian cycles (Ham-p-cycle-SP). The three algorithms are then evaluated in two mesh network topologies, using dynamic Poisson traffic. The simulation results indicate that with the spectrum planning, PWCE-p-cycle-SP and Ham-p-cycle-SP reduce the blocking probability effectively, when comparing with the one without it, i.e., PE-p-cycle. To the best of our knowledge, this is the first attempt to address dynamic p-cycle configuration in EONs.

We layout a network infrastructure that leverages the storage and computing power of a cloud residing in the core for collecting network status and computing multi-path scalable video coding (SVC) streaming provisioning strategies. Therefore, in addition to its conventional tasks in the application layer, the cloud also gets involved in the network layer for the optimization of routing and forwarding. We call this scheme as cloud-assisted SVC streaming, and use it to further improve the performance of SVC streaming by using close cooperation between cloud and network. Compared to source-routing based provisioning, the cloud-assisted scheme can provide more cost-effective provisioning strategies by utilizing better knowledge of network environment together with more powerful computation power. We then propose several multi-path provisioning algorithms for cloud-assisted SVC streaming in heterogeneous networks. To the best of our knowledge, these are the first proposals to work on the problem of adaptive multi-path SVC streaming under the bandwidth, delay and differential delay constraints. Our design of the provisioning algorithms starts from an approach that is based on Max Flow and an Auxiliary Graph. Several extensions are then made based on this approach to address the situations such as provisioning from multiple sources and provisioning in dynamic network environments with rapid background traffic fluctuations. Simulations in both static and dynamic network environments show that the proposed algorithms can achieve effective performance improvements in terms of request blocking probability, bandwidth utilization, packet delay, packet loss rate, and video playback quality.