Meta's data centers underpin a vast array of Internet services and have faced unprecedented demand due to the rapid expansion of AI workloads. The traditional approach of building standardized data centers is increasingly challenged by the exponential growth in required capacity that is now sourced in a variety of non-standard physical environments and data center designs. This shift introduces a complex challenge: how to rapidly and repeatably design custom data center networks that balance multiple, often conflicting, objectives across diverse engineering disciplines.

Meta's rapid expansion in users, business operations, and AI workloads is straining our backbone network, while physical constraints—such as fiber, space, and power—limit the speed of capacity growth. To address these challenges, we present a service-aware network capacity planning suite that systematically improves network efficiency with a service collaboration approach. We propose the "safe capacity" abstraction which enables services to incorporate current and projected network conditions into their compute and storage allocation decisions. We introduce a hose-carving method that efficiently translates service-level traffic demands into detailed traffic matrices, allowing for more precise bandwidth allocation. To promote responsible network usage, we design a network rate card which attributes network consumption to individual services, incentivizing optimization and resource trade-offs. Additionally, new enforcement features at the end-host layer dynamically adjust resource allocations and traffic flows at runtime to maximize utilization. This paper is the first to detail a collaborative, service-aware approach to backbone network efficiency at Meta scale. Based on years of operational experience, we share practical insights and highlight new directions for research in network efficiency.

While recent work shows ML-based approaches are a promising alternative to conventional optimization methods for Traffic Engineering (TE), existing research is limited to a single traffic class. In this paper, we present Hattrick, the first ML-based approach for handling multiple traffic classes, a key requirement of cloud and ISP WANs. As part of Hattrick we have developed (i) a novel neural architecture aligned with the sequence of optimization problems in multiclass TE; and (ii) a variant of classical multitask learning methods to deal with the unique challenge of optimizing multiple metrics that have a precedence relationship. Evaluations on a large private WAN and other public datasets show Hattrick outperforms state-of-the-art optimization-based multiclass TE methods by better coping with prediction error - e.g., for GEANT, Hattrick outperforms SWAN by 5.48% to 19.3% across classes when considering the traffic that can be supported 99% of the time.

Recently, researchers have proposed ML-driven traffic engineering (TE) schemes where a neural network model is used to produce TE decisions in lieu of conventional optimization solvers. Unfortunately existing ML-based TE schemes are not explicitly designed to be robust to topology changes that may occur due to WAN evolution, failures or planned maintenance. In this paper, we present HARP, a neural model for TE explicitly capable of handling variations in topology including those not observed in training. HARP is designed with two principles in mind: (i) ensure invariances to natural input transformations (e.g., permutations of node ids, tunnel reordering); and (ii) align neural architecture to the optimization model. Evaluations on a multi-week dataset of a large private WAN show HARP achieves an MLU at most 11% higher than optimal over 98% of the time despite encountering significantly different topologies in testing relative to training data. Further, comparisons with state-of-the-art ML-based TE schemes indicate the importance of the mechanisms introduced by HARP to handle topology variability. Finally, when predicted traffic matrices are provided, HARP outperforms classic optimization solvers achieving a median reduction in MLU of 5 to 10% on the true traffic matrix.

Network design is the process of dimensioning IP capacity over an optical network infrastructure to satisfy a given set of demands and reliability constraints. Specifically, we consider the problem of hose-based cross-layer network design, which seeks to find a minimum cost design that is able to route demand for all hose traffic matrices under all specified failure states. While most network design problems are solved as Mixed Integer Programs, a commercial solver can become intractable due to the scale of today's networks. We demonstrate how the classic Benders decomposition algorithm can be applied and improved for this problem and discuss practical implementation aspects. We showcase a horizontally scalable distributed framework to leverage the decomposable problem structure and solve millions of linear programs in a distributed manner, thereby making the network design problem tractable. In contrast to the conventional approach where failure states and traffic matrices are planned sequentially, the Benders algorithm finds global optimal designs across all traffic matrices and failure states. This leads to network designs with improved solution quality and reliability, with 20--30% less IP capacity and spectrum consumption, 50% less link augments and up to 20x faster runtime that enables design for hyper scale networks in a matter of hours.

This paper presents Meta's Production Wide Area Network (WAN) Entitlement solution used by thousands of Meta's services to share the network safely and efficiently. We first introduce the Network Entitlement problem, i.e., how to share WAN bandwidth across services with flexibility and SLO guarantees. We present a new abstraction entitlement contract, which is stable, simple, and operationally friendly. The contract defines services' network quota and is set up between the network team and services teams to govern their obligations. Our framework includes two key parts: (1) an entitlement granting system that establishes an agile contract while achieving network efficiency and meeting long-term SLO guarantees, and (2) a large-scale distributed run-time enforcement system that enforces the contract on the production traffic. We demonstrate its effectiveness through extensive simulations and real-world end-to-end tests. The system has been deployed and operated for over two years in production. We hope that our years of experience provide a new angle to viewing WAN network sharing in production and will inspire follow-up research.

This paper presents Facebook's design and operational experience of a Hose-based backbone network planning system. This initial adoption of the Hose model in network planning is driven by the capacity and demand uncertainty pressure of backbone expansion. Since the Hose model abstracts the aggregated traffic demand per site, peak traffic flows at different times can be multiplexed to save capacity and buffer traffic spikes. Our core design involves heuristic algorithms to select Hose-compliant traffic matrices and cross-layer optimization between the optical and IP networks. We evaluate the system performance in production and share insights from years of production experience. Hose-based network planning can save 17.4% capacity and drops 75% less traffic under fiber cuts. As the first study of Hose in network planning, our work has the potential to inspire follow-up research.

Network planning is critical to the performance, reliability and cost of web services. This problem is typically formulated as an Integer Linear Programming (ILP) problem. Today's practice relies on hand-tuned heuristics from human experts to address the scalability challenge of ILP solvers.

We introduce the concepts of monitoring paths (MPs) and monitoring cycles (MCs) for unique localization of shared risk linked group (SRLG) failures in all-optical networks. An SRLG failure causes multiple links to break simultaneously due to the failure of a common resource. MCs (MPs) start and end at the same (distinct) monitoring location(s). They are constructed such that any SRLG failure results in the failure of a unique combination of paths and cycles. We derive necessary and sufficient conditions on the set of MCs and MPs needed for localizing any single SRLG failure in an arbitrary graph. When a single monitoring location is employed, we show that a network must be (k+2)-edge connected for localizing all SRLG failures, each involving up toklinks. For networks that are less than (k+2)-edge connected, we derive necessary and sufficient conditions on the placement of monitoring locations for unique localization of any single SRLG failure of up toklinks. We use these conditions to develop an algorithm for determining monitoring locations. We show a graph transformation technique that converts the problem of identifying MCs and MPs with multiple monitoring locations to a problem of identifying MCs with a single monitoring location. We provide an integer linear program and a heuristic to identify MCs for networks with one monitoring location. We then consider the monitoring problem for networks with no dedicated bandwidth for monitoring purposes. For such networks, we use passive probing of lightpaths by employing optical splitters at various intermediate nodes. Through an integer linear programming formulation, we identify the minimum number of optical splitters that are required to monitor all SRLG failures in the network. Extensive simulations are used to demonstrate the effectiveness of the proposed monitoring technique.

Multiple description coding (MDC) is a powerful source coding technique that involves encoding a media stream into r independently decodeable substreams. With every successful reception of a substream, decoded signal quality improves. We consider the problem of placing a set of servers in the network such that a desired quality of service can be provided to a community of clients. We formulate the server placement (SP) problem, whose goal is to identify the minimum number of server locations that can provide r descriptions to a set of clients such that the delay associated with each path from a chosen server location to a given client is bounded by a given delay constraint and the total "unreliability" associated with the group of paths to a given client is also upper bounded. We show that the SP problem is NP-complete. We propose a mixed-integer linear programming (MILP) formulation and heuristic solution for the SP problem. Simulations are conducted to evaluate the performance of the proposed algorithm and to compare it with the MILP solution.

We present a novel routing approach for multichannel cognitive radio networks (CRNs). Our approach is based on probabilistically estimating the available capacity of every channel over every CR-to-CR link, while taking into account primary radio (PR). Our routing design consists of two main phases. In the first phase, the source node attempts to compute the most probable path (MPP) to the destination (including the channel assignment along that path) whose bandwidth has the highest probability of satisfying a required demand D. In the second phase, we verify whether the capacity of the MPP is indeed sufficient to meet the demand at confidence level delta. If that is not the case, we judiciously add channels to the links of the MPP such that the augmented MPP satisfies the demand D at the confidence level delta. We show through simulations that our protocol always finds the best path to the destination, achieving in some cases up to 200% improvement in connection acceptance rate compared to the traditional Dijkstra.

We introduce the concepts of monitoring paths (MPs) and monitoring cycles (MCs) for unique localization of shared risk linked group (SRLG) failures in all-optical networks. An SRLG failure is a failure of multiple links due to a failure of a common resource. MCs (MPs) start and end at same (distinct) monitoring location(s). They are constructed such that any SRLG failure results in the failure of a unique combination of paths and cycles. We derive necessary and sufficient conditions on the set of MCs and MPs needed for localizing an SRLG failure in an arbitrary graph. When a single monitoring location is employed, we show that a network must be (k + 2)-edge connected for localizing all SRLG failures with up to k links. For networks that are less than (k + 2)-edge connected, we derive necessary and sufficient condition on the placement of monitoring locations for unique localization of any SRLG failure of up to k links. We use these conditions to develop an algorithm for the placement of monitoring locations. We show a graph transformation technique that converts the problem of identifying MCs and MPs with multiple monitoring locations to a problem of identifying MCs with single monitoring location. We provide an integer linear program and a heuristic to identify MCs for networks with one monitoring location. Through extensive simulations, we demonstrate the effectiveness of the proposed monitoring technique.

Multiple description coding (MDC) has emerged as a powerful technique for reliable real-time communications over lossy packet networks. In its basic form, it involves encoding a media stream intorsubstreams that are sent independently from a source to a destination. Each substream (or description) can be decoded independent of the otherr-1 substreams. With every successful reception of a substream, the quality of the decoded signal improves. In this paper, we consider the problem of placing a set of servers in the network such that a desired quality of service can be provided to a community of clients that request MDC-coded traffic. We formulate the server placement (SP) problem, with the goal of identifying the minimum number of server locations that can providerdescriptions to a set of clients such that the delay associated with each path from a chosen server location to a given client is bounded by a given delay constraint and the total ldquounreliabilityrdquo associated with the group of paths to a given client is also upper bounded. We show that the SP problem belongs to the class of NP-complete problems. We propose a mixed-integer linear programming (MILP) formulation and an efficient heuristic solution for the SP problem. Simulations are conducted to evaluate the performance of the proposed algorithm and compare it with the optimal solution provided by the MILP solution.

Efficient routing and wavelength assignment (RWA) in wavelength-routed all-optical networks is critical for achieving high efficiency over the backbone links. Extensive research has been conducted to find strategies to solve the RWA problem with exact network state available at the time of path selection. We consider the problem of minimizing the blocking probability in an all-optical network with partial wavelength conversion and with imprecise network state information. We model imprecision in link-state information (wavelength availability) probabilistically and use Markovian analysis to predict the availability of each link based on its previous advertisement and the estimated average traffic over the link. The estimated probabilities are then used to find the most probable path between a source destination pair. We also consider the problem of lightpath establishment with 1+1 protection. We use the same Markovian model to predict the link availability and then use the probabilistic estimate and a modified version of flow algorithm to find two link-disjoint paths between the source and destination. Simulations are conducted to compare the performance of random-fit and the proposed wavelength assignment schemes. We have also performed extensive simulations to study the performance of the proposed flow algorithm. It is observed that the proposed wavelength assignment scheme performs significantly better in terms of blocking probability than conventional wavelength assignment schemes.

Summary form only given. This paper introduces an integrated source/channel rate-control scheme for video broadcast over wireless channels. The scheme aims at sustaining continuity of playback process at each receiver under varying channel conditions while gracefully degrading the rendered quality. An adaptive playback was used to bound probability of starvation at playback buffers and guarantee a bound on the average delay introduced

Multiple description coding (MDC) has emerged as a powerful technique for reliable real-time communications over lossy packet networks. In its basic form, it involves encoding media into m substreams that are routed independently towards a given destination. Each substream can be decoded independently and with every successful reception of a substream, the overall quality of the decoded signal is improved. In this paper, we consider the problem of placing a set of servers in the network such that a desired QoS can be provided to a community of clients that request MDC coded traffic. Specifically, we consider the server placement (SP) problem where the goal is to identify the "optimal" server positions and associated set of client-server paths such that if MDC content is placed at these servers a cost function that is a linear combination of average delay and path disjointness is minimized. We propose an MILP formulation and a highly efficient heuristic to solve the SP problem. Simulations are conducted to evaluate the performance of the proposed algorithm and compare it with the optimal solution obtained by using the MILP solution.

Ethernet over SONET (EoS) is a popular approach for interconnecting geographically distant Ethernet segments using a SONET transport infrastructure. It typically uses virtual concatenation (VC) for dynamic bandwidth management. The aggregate SONET bandwidth that supports a given EoS system is obtained by "concatenating" a number of virtual channels (VCs), which together form a virtually concatenated group (VCG). This aggregate bandwidth can be increased on demand by adding one or more VCs to the existing VCG. The new VC must be selected such that its end-to-end delay is within a certain range that reflects the delays of all existing VCs in the VCG and the available memory buffer of the EoS system. Algorithmically, the problem of selecting such a VC becomes that of finding a path in a graph network that is bounded by an upper and lower bounds. In this paper, we first prove that the TSCP problem is NP-complete. We then propose a new solution for it based on the "backward-forward" search approach. We show that this solution is much more efficient than the previously proposed MLW-KSP algorithm. We then consider the TSCP problem under inaccurate link information, in which the delay and available bandwidth for each link are taken as random variables. The problem is now formulated as that of finding the most probable path that satisfies the upper and lower delay constraints. We consider two cases. In the first case, we assume that the link delays are random but the link bandwidths are exact. We then consider the more general case where both the link delays and bandwidths are random. Heuristic solutions are presented for both cases. Simulations are conducted to evaluate the performance of the proposed algorithms and to demonstrate the advantages of the probabilistic path selection approach over the classic trigger-based approach.

We consider the problem of minimizing the differential delay in a virtually concatenated Ethernet over SONET (EoS) system by suitable path selection. The link capacity adjustment scheme (LCAS) enables network service providers to dynamically add STS-n channels to or drop them from a virtually concatenated group (VCG). A new STS-n channel can be added to the VCG provided that the differential delay between the new STS-n channel and the existing STS-n channels in the VCG is within a certain bound that reflects the available memory buffer supported by the EoS system. We model the problem of finding such a STS-n channel as a constrained path selection problem where the cost of the required (feasible) path is constrained not only by an upper bound but also by a lower bound. We propose two algorithms to find such a path. Algorithm I uses the well-known k-shortest-path algorithm. Algorithm II is based on a modified link metric that linearly combines the original link weight (the link delay) and the inverse of that weight. The theoretical properties of such a metric are studied and used to develop a highly efficient heuristic for path selection. Simulations are conducted to evaluate the performance of both algorithms in terms of the miss rate and the execution time (average computational complexity).

We consider the problem of optimal power allocation to a user transmitting on a fading channel. The our Is using some specific, convolutional or turbo code. The cost function to be minimized is the bit error rate. The user needs to satisfy an average power constraint We And that the optimal power policy is code and fading distribution dependent and can be significantly different from the commonly considered policies e.g, water filling and (truncated) channel inversion. The gain In BER can be substantial.