arXiv subjects

Beichuan Zhang

Publications and source records attributed to Beichuan Zhang.

15 recordsLinked to original sources

arXiv2026

SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching

Modeling long-term user behavior is central to sequential recommendation and billion-scale industrial recommender systems, yet production ranking models operate under strict latency, memory, communication, and training-throughput constraints. At the 100K scale, the challenge extends beyond attention complexity: raw sequence features must be stored, transferred, and repeatedly processed during training and online serving. Existing approaches based on history truncation, multi-stage behavior retrieval, compressed lifelong histories, or train-short/infer-long extrapolation either weaken end-to-end optimization or retain substantial length-dependent cost. We present SequenceO1, an end-to-end framework for ultra-long user behavior sequence modeling, deployed at full traffic on Douyin with histories of up to 100K interactions. SequenceO1 follows a compress-then-reason design. Its Sketch Attention (SA) uses learnable prototypes and prototype-wise normalization to compress the raw history into a fixed-size, target-agnostic user representation. Target-conditioned Stacked Target-to-History Cross Attention (STCA) then models complementary time scales: a recent 10K suffix for short-term interests and the compact sketch for long-term preferences. To make training and inference practical, SequenceO1 combines low-rank user representation caching, multi-request user-level batching, pipeline lift, and a fused FlashSA kernel to amortize feature storage, communication, and computation across targets, training instances, and consecutive requests. Production experiments show consistent offline and online gains, while the compact cached sketch retains most of the benefit of directly scaling end-to-end sequence ranking to 100K. These results provide a practical model-system approach to efficient attention, sequence compression, and scalable long-sequence and long-context recommendation systems.

cs.IR↗

arXiv2026

From Map-and-Encap to BIER: Observations on Network Routing Scalability

The TCP/IP protocol stack uses IP addresses for two distinct roles: identifying hosts and locating their attachment points in the network topology. This dual purpose creates a fundamental tension that has led to routing and forwarding scalability challenges throughout the history of the Internet in unicast packet delivery and, more notably, in multicast delivery. This paper reviews the evolution of routing scalability solutions over the years and makes four observations. First, map-and-encap is a recurring architectural solution shared by all scalable unicast and multicast delivery methods, developed independently across different problem contexts. Second, a new solution tends to succeed when it can bring immediate local gains to early adopters without requiring coordination across administrative domains. Third, network routing and forwarding designs that depend on external factors, such as the number of distinct end sites or even application-specific deliveries, inherently preclude an upper bound on their scalability. Fourth, today's inter-domain routing protocol, BGP, lacks a topological abstraction equivalent to an egress router within a routing domain, thereby inherently preventing a map-and-encap solution for scalability. These observations offer insights into the design of future scalable routing system architectures.

cs.NI↗

arXiv2026

Starlink Constellation: Deployment, Configuration, and Dynamics

Starlink has rapidly emerged as the world's largest satellite constellation and the de facto reference system for low Earth orbit (LEO) networking research. Existing literature predominantly models Starlink as a static, symmetric, and fully deployed structure with uniformly distributed satellites. However, we reveal that Starlink's actual deployment, orbital configurations, and operational dynamics fundamentally deviate from these idealized assumptions. Leveraging satellite observation data spanning 2019 to 2025, we demonstrate that the constellation is highly dynamic across multiple temporal and spatial scales. Macroscopically, Starlink comprises multiple orbital shells undergoing continuous active deployment and reconfiguration. Microscopically, individual satellites exhibit high mobility, frequently executing collision-avoidance maneuvers, altitude adjustments, and intra-orbital relocations. We discover that while the majority of satellites form a relatively stable structure with near-uniform spacing, other satellites tend to cluster as twins or triads as in-orbit backups. Furthermore, empirical survival analysis indicates an operational lifespan of 4-6 years and an average daily failure probability of 0.0128%. Ultimately, our data-driven characterization exposes Starlink as a highly heterogeneous and continuously evolving network. We provide critical empirical insights that challenge prevailing simulation models, offering a more accurate foundation for future LEO topology design, routing protocols, and performance evaluations.

cs.NI↗

arXiv2026

LEO Topology Design Under Real-World Deployment Constraints

The performance of large-scale Low-Earth-Orbit (LEO) networks, which consist of thousands of satellites interconnected by optical links, is dependent on its network topology. Existing topology designs often assume idealized conditions and do not account for real-world deployment dynamics, such as partial constellation deployment, daily node turnovers, and varying link availability, making them inapplicable to real LEO networks. In this paper, we develop two topology design methods that explicitly operate under real-world deployment constraints: the Long--Short Links (LSL) method, which systematically combines long-distance shortcut links with short-distance local links, and the Simulated Annealing (SA) method, which constructs topologies via stochastic optimization. Evaluated under both full deployment and partial deployment scenarios using 3-months of Starlink data, our methods achieve up to 45% lower average end-to-end delay, 65% fewer hops, and up to $2.3\times$ higher network capacity compared to +Grid. Both methods are designed to handle daily node turnovers by incrementally updating the topology, maintaining good network performance while avoiding costly full reconstruction of the topology.

cs.NI↗

arXiv2025

Make It Long, Keep It Fast: End-to-End 10K Long User Behavior Sequence Modeling for Billion-Scale Douyin Recommendation

Short-video recommenders such as Douyin must exploit extremely long user behavior histories without breaking latency or cost budgets. We present an end-to-end industrial recommender system that scales long-sequence recommendation modeling to 10K-length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training over long user behavior sequences. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy -- train on shorter windows, infer on much longer ones -- so the model generalizes to 10K-scale histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end ultra-long sequence recommendation to the 10K regime.

cs.LG↗

arXiv2024

Full Stage Learning to Rank: A Unified Framework for Multi-Stage Systems

The Probability Ranking Principle (PRP) has been considered as the foundational standard in the design of information retrieval (IR) systems. The principle requires an IR module's returned list of results to be ranked with respect to the underlying user interests, so as to maximize the results' utility. Nevertheless, we point out that it is inappropriate to indiscriminately apply PRP through every stage of a contemporary IR system. Such systems contain multiple stages (e.g., retrieval, pre-ranking, ranking, and re-ranking stages, as examined in this paper). The \emph{selection bias} inherent in the model of each stage significantly influences the results that are ultimately presented to users. To address this issue, we propose an improved ranking principle for multi-stage systems, namely the Generalized Probability Ranking Principle (GPRP), to emphasize both the selection bias in each stage of the system pipeline as well as the underlying interest of users. We realize GPRP via a unified algorithmic framework named Full Stage Learning to Rank. Our core idea is to first estimate the selection bias in the subsequent stages and then learn a ranking model that best complies with the downstream modules' selection bias so as to deliver its top ranked results to the final ranked list in the system's output. We performed extensive experiment evaluations of our developed Full Stage Learning to Rank solution, using both simulations and online A/B tests in one of the leading short-video recommendation platforms. The algorithm is proved to be effective in both retrieval and ranking stages. Since deployed, the algorithm has brought consistent and significant performance gain to the platform.

cs.IR↗

arXiv2023

SHARK: A Lightweight Model Compression Approach for Large-scale Recommender Systems

Increasing the size of embedding layers has shown to be effective in improving the performance of recommendation models, yet gradually causing their sizes to exceed terabytes in industrial recommender systems, and hence the increase of computing and storage costs. To save resources while maintaining model performances, we propose SHARK, the model compression practice we have summarized in the recommender system of industrial scenarios. SHARK consists of two main components. First, we use the novel first-order component of Taylor expansion as importance scores to prune the number of embedding tables (feature fields). Second, we introduce a new row-wise quantization method to apply different quantization strategies to each embedding. We conduct extensive experiments on both public and industrial datasets, demonstrating that each component of our proposed SHARK framework outperforms previous approaches. We conduct A/B tests in multiple models on Kuaishou, such as short video, e-commerce, and advertising recommendation models. The results of the online A/B test showed SHARK can effectively reduce the memory footprint of the embedded layer. For the short-video scenarios, the compressed model without any performance drop significantly saves 70% storage and thousands of machines, improves 30\% queries per second (QPS), and has been deployed to serve hundreds of millions of users and process tens of billions of requests every day.

cs.IR↗

arXiv2021

BLEnD: Improving NDN Performance Over Wireless Links Using Interest Bundling

Named Data Networking (NDN) employs small-sized Interest packets to retrieve large-sized Data packets. Given the half-duplex nature of wireless links, Interest packets frequently contend for the channel with Data packets, leading to throughput degradation. In this work, we present a novel idea called BLEnD, an Interest-bundling technique that encodes multiple Interests into one at the sender and decodes at the receiver. The major design challenges are to reduce the number of Interest transmissions without impacting the one-Interest one-Data principle embedded everywhere in NDN architecture and implementation, and support flow/congestion control mechanisms that usually use Interest packets as signals. BLEnD achieves these by bundling/unbundling Interests at the link adaptation layer, keeping all NDN components unaware and unaffected. Over a one-hop WiFi link, BLEnD improves application throughput by 30%. It may also be used over multiple hops and be improved in a number of ways.

cs.NI↗

arXiv2021

On the Analysis of Adaptive-Rate Applications in Data-Centric Wireless Ad-Hoc Networks

Adapting applications' data rates in multi-hop wireless ad-hoc networks is inherently challenging. Packet collision, channel contention, and queue buildup contribute to packet loss but are difficult to manage in conventional TCP/IP architecture. This work explores a data-centric approach based on Name Data Networking (NDN) architecture, which is considered more suitable for wireless ad-hoc networks. We show that the default NDN transport offers better performance in linear topologies but struggles in more extensive networks due to high collision and contention caused by excessive Interests from out-of-order data retrieval and redundant data transmission from improper Interest lifetime setting as well as in-network caching. To fix these, we use round-trip hop count to limit Interest rate and Dynamic Interest Lifetime to minimize the negative effect of improper Interest lifetime. Finally, we analyze the effect of in-network caching on transport performance and which scenarios may benefit or suffer from it.

cs.NI↗

arXiv2021

On Data-centric Forwarding in Mobile Ad-hoc Networks: Baseline Design and Simulation Analysis

IP networking deals with end-to-end communication where the network layer routing protocols maintain the reachability from one address to another. However, challenging environments, such as mobile ad-hoc networks or MANETs, lead to frequent path failures and changes between the sender and receiver, incurring higher packet loss. The obligatory route setup and maintenance of a device-to-device stable path in MANETs incur significant data retrieval delay and transmission overhead. Such overhead exaggerates the packet loss manifold. Named Data Networking (NDN) can avoid such delays and overhead and significantly improve the overall network performance. It does so with direct application-controlled named-data retrieval from any node in a network instead of reaching a specific IP address with protocol message exchange. However, existing works lack any explicit or systematic analysis to justify such claims. Our work analyzes the core NDN and IP architectures in a MANET at a baseline level. The extensive simulations show that NDN, when applied correctly, yields much lower data retrieval latency than IP and can lower the network transmission overhead in most cases. As a result, NDN's stateful forwarder can significantly increase the retrieval rate, offering a better trade-off at the network layer. Such performance comes from its caching, built-in multicast, and request aggregation without requiring an IP-like separate routing control plane.

cs.NI↗

arXiv2021

Smart Name Lookup for NDN Forwarding Plane via Neural Networks

Name lookup is a key technology for the forwarding plane of content router in Named Data Networking (NDN). To realize the efficient name lookup, what counts is deploying a highperformance index in content routers. So far, the proposed indexes have shown good performance, most of which are optimized for or evaluated with URLs collected from the current Internet, as the large-scale NDN names are not available yet. Unfortunately, the performance of these indexes is always impacted in terms of lookup speed, memory consumption and false positive probability, as the distributions of URLs retrieved in memory may differ from those of real NDN names independently generated by content-centric applications online. Focusing on this gap, a smart mapping model named Pyramid-NN via neural networks is proposed to build an index called LNI for NDN forwarding plane. Through learning the distributions of the names retrieved in the static memory, LNI can not only reduce the memory consumption and the probability of false positive, but also ensure the performance of real NDN name lookup. Experimental results show that LNI-based FIB can reduce the memory consumption to 58.258 MB for 2 million names. Moreover, as it can be deployed on SRAMs, the throughput is about 177 MSPS, which well meets the current network requirement for fast packet processing.

cs.NI↗

arXiv2020

The Case for Hop-by-Hop Traffic Engineering

State-of-the-art Internet traffic engineering uses source-based explicit routing via MPLS or Segment Routing. Though widely adopted in practice, source routing can face certain inefficiencies and operational issues, caused by its use of bandwidth reservations. In this work, we make the case for Hop-by-Hop (HBH) Traffic Engineering: splitting traffic among nexthops at every router, rather than splitting traffic among paths only at edge routers. We show that HBH traffic engineering can achieve the original goals of MPLS (i.e., efficient use of network resources), with a much simpler design that does not need bandwidth reservations or predictions of traffic demand. We implement a prototype in the ns-3 network simulator, to investigate the cost imposed by 1) the restricted path choice of loop-free HBH multipath routing, and 2) the distributed decisions of each router, based on its local network view. We show that the former is more important than the latter, but that, other than a few outliers, our design shows a performance (= aggregate user utility) close to the theoretical optimum.

cs.NI↗

arXiv2020

Internet-Scale Video Streaming over NDN

Research in Information-Centric Networking (ICN) and Named Data Networking (NDN) has produced many protocol designs and software prototypes, but they need to be validated and evaluated by real usage on the Internet, which is also critical to the realization of the ICN/NDN vision in the long-run. This paper presents the first Internet-scale adaptive video streaming service over NDN, which allows regular users to watch videos using NDN technology with no software installation or manual configuration. Since mid-2019, the official NDN website has started using our service to deliver its videos to Internet users over the global NDN testbed, showcasing the feasibility of NDN. We conduct real-world experiments on the NDN testbed to validate the proper implementation of the client, network, and server sides of the proposed system and also evaluate the performance of our video streaming service.

cs.NI↗

arXiv2019

Hop-by-Hop Multipath Routing: Choosing the Right Nexthop Set

The Internet can be made more efficient and robust with hop-by-hop multipath routing: Each router on the path can split packets between multiple nexthops in order to 1) avoid failed links and 2) reduce traffic on congested links. Before deciding how to split traffic, one first needs to decide which nexthops to allow at each step. In this paper, we investigate the requirements and trade-offs for making this choice. Most related work chooses the viable nexthops by applying the "Downward Criterion", i.e., only adding nexthops that lead closer to the destination; or more generally by creating a Directed Acyclic Graph (DAG) for each destination. We show that a DAG's nexthop options are necessarily limited, and that, by using certain links in both directions (per destination), we can add further nexthops while still avoiding loops. Our solution LFID (Loop-Free Inport-Dependent) routing, though having a slightly higher time complexity, leads to both a higher number of and shorter potential paths than related work. LFID thus protects against a higher percentage of single and multiple failures (or congestions) and comes close to the performance of arbitrary source routing.

cs.NI↗

arXiv2016

An Experimental Investigation of Hyperbolic Routing with a Smart Forwarding Plane in NDN

Routing in NDN networks must scale in terms of forwarding table size and routing protocol overhead. Hyperbolic routing (HR) presents a potential solution to address the routing scalability problem, because it does not use traditional forwarding tables or exchange routing updates upon changes in network topologies. Although HR has the drawbacks of producing sub-optimal routes or local minima for some destinations, these issues can be mitigated by NDN's intelligent data forwarding plane. However, HR's viability still depends on both the quality of the routes HR provides and the overhead incurred at the forwarding plane due to HR's sub-optimal behavior. We designed a new forwarding strategy called Adaptive Smoothed RTT-based Forwarding (ASF) to mitigate HR's sub-optimal path selection. This paper describes our experimental investigation into the packet delivery delay and overhead under HR as compared with Named-Data Link State Routing (NLSR), which calculates shortest paths. We run emulation experiments using various topologies with different failure scenarios, probing intervals, and maximum number of next hops for a name prefix. Our results show that HR's delay stretch has a median close to 1 and a 95th-percentile around or below 2, which does not grow with the network size. HR's message overhead in dynamic topologies is nearly independent of the network size, while NLSR's overhead grows polynomially at least. These results suggest that HR offers a more scalable routing solution with little impact on the optimality of routing paths.

cs.NI↗