INFOCOM 2026 论文列表

Edge computing platforms are increasingly deployed to support delay-sensitive and resource-intensive applications. However, current task scheduling strategies, which rely on user-requested resources, often lead to low resource utilization and reduced platform profit due to users’ tendency to over-request resources. Overbooking, widely adopted in industries such as airlines and hotels, can improve utilization but introduce risk under uncertain task resource usage. This paper applies resource overbooking to edge clouds, focusing on task scheduling optimization under uncertainty. We formulate the problem as a stochastic mixed integer program, which is proven to be NP-hard. To this end, a risk evaluation scheme is proposed to accurately quantify the risk of overload without assuming specific resource usage distributions, which has an additive error guarantee. Based on this scheme, we transform the problem into a more deterministic form and prove that the objective function is submodular. This enables us to design a greedy algorithm that achieves a (1 − 1/e)-approximation ratio with lower complexity. Extensive experiments on a real-world dataset demonstrate that the proposed algorithm significantly improves profit by 0.23×-3.35× and resource utilization by 0.16×-0.75× while accurately controlling the risk associated with overbooking.

Speculative inference accelerates Large Language Model (LLM) serving by leveraging small speculative models (SSMs) as drafters to generate draft tokens, which are subsequently verified in parallel by the LLM. However, the practical effectiveness of this approach is frequently hindered by the tight coupling of compute-intensive LLMs and memory-intensive SSMs within a single node (e.g., single-GPU device). This coupling induces resource contention, thereby significantly lowering draft acceptance rates and system throughput. To overcome these obstacles, we present CoSine, a novel speculative inference system that decouples sequential speculation from parallel verification. Specifically, CoSine distributes multiple specialized drafters across the heterogeneous nodes and coordinates them via a verification server, enabling cost-effective collaboration to mitigate inference bottlenecks. First, CoSine dynamically routes requests to the proper drafters and incorporates a confidence-based token fusion mechanism to synthesize high-quality drafts. Then, CoSine orchestrates the verification pipeline featuring prioritized draft queueing and adaptive batch management to balance the decoupled workflows and minimize pipeline bubbles. Experimental evaluation demonstrates that CoSine significantly outperforms state-of-the-art approaches. Under equivalent resource constraints, CoSine achieves latency reductions of up to 27.1% and throughput improvements from 1.24× to 1.62×.

Spiking Transformers, which couple Spiking Neural Networks (SNNs) with Transformer architectures, promise low-energy inference and strong accuracy. However, their training is hindered by the intrinsic temporal dynamics of SNNs, which cause severe memory and synchronization bottlenecks. Through an in-depth analysis of the SpikeFormer architecture and training workload, we uncover local dependency in its training process, which enables time parallelism. We present TPipe, the first unified framework for efficient SpikeFormer training that integrates time, data, and pipeline parallelism. We formalize the Automated SpikeFormer Parallelism (ASP) problem to jointly optimize parallelism and placement strategies. We further formalize the SpikeFormer Scheduling (SS) problem, which eliminates pipeline bubbles via fine-grained, sub–micro-batch asynchronous scheduling. We prove that TPipe achieves bounded speed-up over synchronous baselines while preserving training correctness. Experiments on multiple SpikeFormer models and datasets show that TPipe attains 1.1–3.8× speed-up over default pipeline baselines while maintaining accuracy.

High-performance network functions (NFs) are typically designed with modularization, consolidation, and vectorization, forming NF graphs. However, optimizing vectorized NF graphs is challenging due to complex interactions among workload diversity, packet characteristics, and hardware constraints. While intelligent algorithms are promising, tuning them directly on real systems is often impractical due to significant engineering overhead and hardware costs. We argue that a simulator that is interactive, accurate, and efficient is essential for enabling effective NF-graph optimization.We present NGSim, a simulator designed to meet these goals through a hybrid event- and data-driven approach with a decoupled, parallel simulation pipeline. NGSim enables fast and accurate exploration of NF-graph performance under diverse conditions. Evaluation shows that it achieves an average relative error of only 1.38% for throughput and an absolute error of 0.03% for SLO violation rate, while reducing simulation time by 85% compared to its sequential baseline. A case study with reinforcement learning further demonstrates NGSim’s effectiveness, yielding an 11.5% improvement in throughput without violating SLOs.

Training large language models faces frequent interruptions due to various faults, demanding robust fault-tolerance. Existing backup-free methods, such as redundant computation, dynamic parallelism, and data rerouting, each incur performance penalties, whether from ongoing overhead, lengthy reconfigurations, or post-recovery inefficiencies. We propose Chameleon, an adaptive fault-tolerant system that intelligently selects optimal recovery strategies when a failure occurs. Chameleon achieves this through a unified performance model, expedient execution plan search, accurate performance estimation, and efficient communication optimizations. Experiments on a 32-card cluster show that Chameleon maintains a performance gap of within 11.00% between post-recovery and failure-free training, while preserving model convergence and efficient memory usage. Compared to state-of-the-art methods, Chameleon achieves up to 1.229x and 1.355x higher average throughput than Oobleck and Recycle, respectively.

The growing scale of large language models (LLMs) has made communication overhead a critical bottleneck in distributed training, primarily due to imbalanced traffic loads. Existing load balancing methods often lead to severe flow contention when handling low-entropy and high-volume LLM training flows. Motivated by the point-to-point pattern in each collective communication phase and the inherent periodicity of training traffic, we propose PhOrch, a proactive phase-level contention-free flow path orchestration framework tailored for LLM training workloads. We formulate the orchestration as an optimization problem, which is typically NP-hard. To tackle this problem, we develop a segmented edge coloring algorithm for bipartite multigraphs, which efficiently assigns flow paths while avoiding contention. Evaluation results demonstrate that PhOrch reduces the per-cycle communication time by 60% compared to the state-of-the-art methods and achieves contention-free training traffic in non-oversubscribed topologies, indicating that our method substantially mitigates the communication bottleneck during the LLM training process.

Multiview video streaming requires much higher bandwidth than conventional 2D video streaming. Current multi-view video streaming systems mainly leverage view interpolation to reduce redundancy between adjacent 2D views, and do not consider 3D spatial redundancy among all views. In this paper, we propose Argus, a multiview video streaming system that reduces 3D spatial redundancy by transmitting only a sparse subset of views and synthesizing the remaining views through 3D Gaussian reconstruction and splatting. Specifically, we propose: (i) a 3D Gaussian splatting-assisted multiview video coding method featuring real-time 3D Gaussian reconstruction and view synthesis, and (ii) a content-adaptive view selection method that dynamically selects a subset of views to optimize the visual quality of the synthesized views. We develop a prototype system of Argus, supporting up to 50-view autostereoscopic 3D display. Comprehensive experiments show that Argus achieves bitrate reduction by 41.33% and 44.12% compared with two baselines on two datasets, respectively, and supports real-time multiview video streaming at over 30 FPS.

Distributed Quantum Computing (DQC) scales quantum computing capabilities by interconnecting multiple Quantum Processing Units (QPUs) for collaborative computation, but faces the critical challenge of high entanglement costs associated with remote gate operations. Circuit transformation and qubit mapping are critical components of quantum circuit compilation for minimizing entanglement cost, which are interdependent in DQC. Existing approaches that treat circuit transformation and qubit mapping as separate optimization problems fail to account for their fundamental interdependence, resulting in suboptimal entanglement utilization and degraded performance. This paper presents the first optimization framework for circuit transformation and qubit mapping tailored for DQC. To address this joint optimization problem, we propose Qmactr, a two-phase reinforcement learning approach. The first phase trains a qubit mapping agent, and the second phase trains a circuit transformation agent that utilizes the mapping agent as a reward oracle. Extensive evaluation results demonstrate that the proposed approach reduces entanglement consumption by up to 35% compared to state-of-the-art sequential methods while achieving superior circuit fidelity.

Multi-access Edge Computing (MEC) federations allow 5G operators to pool their dispersed micro‑datacenters, ensuring that users are always served by an edge host close to their current location. Yet today’s research stays almost exclusively user‑centric: trimming latency or energy but ignores the operators’ own constraints, i.e. heterogeneous servers, hard resource budgets and, above all, the need for balanced give‑and‑take across domains. This results in a brittle reactive offloading that can overwhelm the most generous partner. We tackle the problem from the operators’ perspective, casting one scheduling epoch of a federation as an exact‑potential game that jointly captures (i) per‑application deadlines, (ii) per‑server CPU/RAM/storage heterogeneity, and (iii) a fairness objective keeping each domain’s exports and imports in line. We then reach the game's equilibrium with a lightweight best‑response approach, \name, where orchestrators exchange only bids, never raw telemetry, and need only one message per epoch. We evaluate \name on a novel open‑source MEC‑federation simulator, configurable down to link latency, bandwidth and instruction counts. Experiments show that our solution reduces response time by 37\%, and boosts overall CPU/RAM utilization by up to 20\% compared to six baselines. All simulator code, scenarios, and experiments are openly available to foster reproducibility and future research.

This paper revisits combinatorial Thompson sampling (CTS) in the semi-bandit setting with sleeping arms, where only a subset of arms is available each round under combinatorial constraints. Such settings frequently arise in networking applications and pose significant challenges due to the need for combinatorial optimization over a dynamically changing action space. A canonical example is wireless mesh routing, where fluctuating link availability and routing constraints lead to a time-varying set of feasible paths between source and destination nodes.The existing works have three key limitations: (1) the lack of worst-case regret guarantees for CTS in semi-bandit settings, even without accounting for sleeping arms; (2) the absence of theoretical guarantees for CTS under adversarially varying arm availability; and (3) the subpar empirical performance of CTS with Gaussian priors (CTS-G).To address these issues, we propose CL-SG, a simple yet effective CTS-G variant that samples a single shared Gaussian seed each round to coordinate exploration across arms. Theoretically, CL-SG achieves a worst-case regret bound of $\tilde O(\sqrt {mNT} )$ and a matching lower bound of $\Omega (\sqrt {mNT} )$. Our analysis can also easily recover a regret bound for CTS-G. With experiments driven by real-world datasets, CL-SG consistently outperforms diverse benchmark algorithms such as CTS-G and CTS-B. We open-source our implementation and experiments to support reproducibility and further research.

Autonomous System (AS) relationship inference is critical for understanding Internet topology and detecting routing anomalies, yet existing methods suffer from accuracy and stability issues. Current algorithms struggle with peer-to-peer (p2p) relationship inference due to insufficient visibility in BGP feeds and lack of robust heuristics, while exhibiting instability under changing vantage point (VP) deployments.We present Apollo, an accurate and stable AS relationship inference method using trusted seeds and semi-supervised learning. We first extract trusted AS relationships by using export policies in Internet Routing Registry (IRR) databases. We then apply the valley-free principle to infer relationships along AS paths using these trusted seeds. For remaining unlabeled links, we employ probabilistic label generation and ensemble learning to achieve balanced training and accurate inference. Our evaluations of BGP data from 2022 to 2024 show that, compared to current methods, Apollo reduces the overall error rates by 2.52×-3.85× and p2p relationship error rates by up to 11.25×. Apollo maintains superior stability with zero relationship transitions in two of the three categories under varying VP deployments. Applied to route leak detection, Apollo uncovers 2.10×-2.64× more historical route leak incidents than prior methods.

RFID has become fundamental to applications ranging from inventory management to indoor localization, where reliably reading each tag multiple times (K-reads) is critical for system reliability and robustness. While existing anti-collision protocols address general tag identification, they fail to guarantee deterministic K-reads completion due to the probabilistic nature of tag responses and physical-layer challenges like multipath fading and interference. This paper presents the first work that formally introduces and addresses the K-reads problem, which aims to ensure each tag is successfully read at least K times with minimal total inventory time. Our key insight stems from the observation that conventional approaches waste much interrogation time re-reading already-counted tags while persistently missing others. To solve this, we propose: 1) a novel power-tuned silence method that exploits channel asymmetry to selectively suppress easily detectable tags without transmission overhead; 2) a hybrid framework that intelligently combines: framed slotted ALOHA for bulk identification, power-tuned silence that selectively suppress easily detectable tags without transmission overhead, and targeted polling for remaining hard-to-read tags. Our approach dynamically optimizes the use of these techniques based on real-time tag response characteristics. Implemented on commercial RFID, the system achieves a 65.7% reduction in inventory time compared to the baseline.

The collective communication libraries (xCCL) are an important networking support for distributed training systems. Optimizing xCCL is crucial for reducing training costs while maintaining accuracy, but existing approaches often require extensive manual intervention, lack user-friendliness, and struggle to adapt to complex environments. Therefore, we introduce xCCLTuner, an out-of-the-box tuning system without needing to categorize or model parameters that automatically optimizes xCCL parameters for distributed training systems. xCCLTuner utilizes Bayesian optimization to navigate the high-dimensional parameter space of xCCL knobs in a black-box manner. It overcomes key challenges through three main designs: (1) Low-dimensional tuning to reduce the complexity of the parameter space, (2) xCCL-specific lookup table to identify and skip fatal parameter combinations, and (3) Bucketization techniques to conquer vast parameter domains. xCCLTuner rapidly delivers optimized training times across various topologies, showcasing its efficiency and ease of use. Our experiments demonstrate that xCCLTuner achieves up to a 5.87X improvement in throughput.

As magnetic resonant coupling (MRC)-based wireless power transfer (WPT) systems continue to evolve, in-band receiver (RX) feedback communication becomes essential for achieving tighter integration, adaptive power control, and reliable system coordination. However, existing on-off keying-based in-band communication inherently interrupts the charging process, leading to reduced charging efficiency and rendering it unsuitable for the frequent feedback required by mainstream standards. In this paper, we propose Chunip, a charging-uninterrupted in-band parallel feedback communication mechanism for MIMO MRC-WPT systems, achieved by integrating resistive load modulation with channel matrix matching. To address the exponential growth of the channel states and dynamic channel variations, we introduce a three-mode RX modulation scheme and a hypothesis-testing-inspired method for efficient channel matrix collection. Moreover, we employ a performance-guided online update strategy that adaptively reuses stable historical information to preserve accurate channel matching despite load fluctuations and RX movements. We develop a Chunip prototype and conduct comprehensive evaluations. Experimental results demonstrate that our system achieves over 95% communication accuracy when serving 12 RXs simultaneously and delivers 27.69% more power than existing state-of-the-art solutions.

Host-side packet processing latency occupies a significant proportion of the total end-to-end latency of TCP flows in DCNs. The monitoring of host-side latency is critical to achieve DCN performance diagnosis, troubleshooting, and service-level objectives. Existing solutions normally incur significant CPU consumption or lose full-coverage due to sampling, and are susceptible to host-side anomalies. In this paper, we propose a robust method termed as Holm to achieve host-side latency monitoring without host CPU consumption by offloading monitoring logic to DPU, and achieve monitoring coverage and resource balancing through differentially treating best-effort and latency-sensitive TCP flows. Holm first identifies best-effort and latency-sensitive TCP flows in DPU, and performs corresponding selective and full-coverage monitoring. A mathematical model is then established for evaluating the overall latency state of latency-sensitive TCP flows, and if abnormal increases are detected, foreboding the threat of packet overflow, a corresponding number of congestion notifications will be delivered to best-effort TCP flow senders for reducing host-side anomaly impacts. Holm is evaluated by our prototype implementation on NVIDIA BlueField-3 DPU. The experimental results indicate that compared to representative methods, it eliminates host CPU consumption with low-overhead and achieves robustness by improving coverage by up to 2.07× even under serious host-side congestion.

Checkpointing is a widely used technique to ensure fault tolerance in distributed machine learning systems. Frequent checkpointing can significantly reduce the wasted training time in the event of a failure. However, due to the large overhead of checkpoint storage and transmission, it is still challenging to achieve frequent checkpointing in practice. In this paper, we propose FlashCheck, a checkpointing framework that enables frequent checkpointing in distributed machine learning systems. FlashCheck first introduces the mergeable delta compression (MDC) technique that provides high compression ratios for checkpoints with low computational overhead. MDC allows for efficient merging of compressed checkpoints to mitigate the continuously increasing storage overhead. To further reduce the transmission overhead, FlashCheck employs a semi-reliable transport protocol (CSTP) that offers low tail transmission latency and fine-grained control over the reliability of checkpoint transmission. Through extensive experiments on real-world distributed machine learning workloads, we demonstrate that FlashCheck can conduct a significant checkpoint size reduction of over 99.17%. It can achieve over 4x speedup in the checkpointing frequency compared to existing solutions.

In this paper, we propose the geometric algebra-informed neural radiance fields (GAI-NeRF), a novel framework for wireless channel prediction that leverages geometric algebra attention mechanisms to capture ray-object interactions in complex propagation environments. Our approach incorporates global token representations, drawing inspiration from transformer architectures in language and vision domains, to aggregate learned spatial-electromagnetic features and enhance scene understanding. We identify limitations in conventional static ray tracing modules that hinder model generalization and address this challenge through a new ray tracing architecture. This design enables effective generalization across diverse wireless scenarios while maintaining computational efficiency. Experimental results demonstrate that GAI-NeRF achieves superior performance in channel prediction tasks by combining geometric algebra principles with neural scene representations, offering a promising direction for next-generation wireless communication systems. Moreover, GAI-NeRF greatly outperforms existing methods across multiple wireless scenarios. To ensure comprehensive assessment, we further evaluate our approach against multiple benchmarks using newly collected real-world indoor datasets tailored for single-scene downstream tasks and generalization testing, confirming its robust performance in unseen environments and establishing its high efficacy for wireless channel prediction.

As large foundation models are routinely trained with hundreds of thousands of GPU compute nodes, the need for simulating a computer network at scale has become more critical and relevant than ever. Without a doubt, packet-level discrete-event simulation (DES) offers the finest granularity, and thus the highest accuracy. Unfortunately, conventional discrete-event simulators were widely known to be slow, and thus unable to accommodate the scale of modern networks. Recent work in the literature attempted to estimate the performance of large-scale networks using deep neural network models, but such estimation inevitably leads to a loss of packet-level accuracy, when compared to the ground truth from discrete-event simulators.But is it really the case that discrete-event simulators are not performant at scale? In this paper, we advocate that a process-based design is a simpler, more scalable, and performant choice than the current event-based design. We challenge the conventional wisdom that discrete-event simulators lack scalability, and progressively introduce the design and implementation of two new DES frameworks, ns.py and Days, built using Python and Rust, and with modern development advances in generators, asynchronous programming, and stackless coroutines. Our new simulation frameworks are designed to be lean and performant, outperforming existing discrete-event simulators and performance estimators by up to three orders of magnitude.

Machine learning-based network traffic classification models are vulnerable to backdoor attacks. Recent work demonstrates that packet-level backdoor attacks are feasible even in black-box settings, but their attack success rates remain low under clean-label conditions (e.g., 36.26% on ISCX VPN-nonVPN). Unlike image-based models, simple trigger patterns are ineffective against traffic classifiers because of the unique characteristics of network data. These include variable-length packet sequences, strict protocol compliance, and semantic constraints that limit allowable modifications. Such challenges make it difficult for fixed-pattern triggers to maintain integrity, saliency, and consistency across diverse network environments. Motivated by this insight, we propose BDpackets, a packet-level clean-label backdoor attack that employs a novel feature fusion technique to nonlinearly integrate salient features and generate semantically rich triggers, thereby significantly improving attack effectiveness. Experiments on multiple datasets and model architectures show that BDpackets achieves an attack success rate of 98.91%, outperforming prior backdoor methods by 62.65% and 45.36% in absolute terms. Moreover, BDpackets maintains stealthiness and successfully evades detection by state-of-the-art defenses.