This July 2026 issue contains three technical papers. The first paper, The Carrier Pigeon Internet Protocol: An Algorithmic (and Light-hearted) Perspective by Matthias Bentert, Shay Kutten, Darya Melnyk, Tijana Milentijević, and Stefan Schmid is an amusing read that investigates protocols and complexities of communication networks based on carrier pigeons. While lighthearted in nature, the paper is technically grounded in classic mathematical techniques used to analyze modern computer networks. This creative paper is both entertaining and informative—in fact, the paper proposes follow-up work that I'd love to see.
We present a large-scale measurement of emerging Artificial Intelligence (AI)-oriented web permission and descriptor mechanisms across approximately 4M domains. Beyond traditional robots.txt files, we analyze the adoption, structure, and interactions of newer files, including the permission-oriented ai.txt and the descriptor-oriented llms.txt and llms-full.txt. While robots.txt remains the dominant mechanism for signaling crawler permissions, adoption of AI-specific permission files is emerging, especially among technology-focused domains. Directive usage shows a strong emphasis on restricting AI agents, with prevalent disallow rules and relatively few AI-specific permission controls. Crossfile comparisons reveal that conflicts between robots.txt and ai.txt are infrequent, while some llms.txt link targets point to content blocked by access-control files.
An artifact is a resource essential to the complete understanding of a scientific article, extending beyond the manuscript itself. As more conferences allow authors to submit artifacts alongside their articles, different review processes, badge types are emerging, and badge-acquisition requirements. In this paper, we describe the process for conducting artifact evaluation that has been implemented in the Brazilian community. Our discussion focuses on the two largest Brazilian networking and security venues, where we discuss the challenges we encountered during our interactions and the solutions we ultimately adopted to improve the review process.
On April 28th, 2025, a major power outage caused a disruption of electric service across Spain and Portugal, lasting about ten hours in most regions and even longer in some areas. We analyze the impact of the power outage, showing how it led to a near-instantaneous 33–70% drop in traffic at 4 major Iberian IXPs, while core infrastructure remained largely functional. We show that the Internet was impacted by a second event: a partial outage in a data center that resulted in a second drop of traffic and BGP sessions.
The theoretical model behind the pigeon post as a link layer in a communication network was introduced by Shannon (under the guise of studying One-Time Pads for cryptography). That is, to send a one-hop message to v, a node u needs a mail pigeon bred and raised at v. When sending a message using a pigeon to v, node u loses the pigeon. To send another message to v, node u needs another pigeon of v. It has been demonstrated that the communication bandwidth achievable with pigeon post can exceed that of networks using other media. This has already motivated the introduction of Internet standards that allow the use of pigeons as Internet link-layer media. In this paper, we begin to fill in the missing piece: designing algorithms for breeding and scheduling pigeons to meet a given communication demand efficiently, minimizing the number of pigeons required. We consider singlehop, 2-hop, and multihop pigeon use. While the singlehop variant admits a simple characterization, both the 2-hop and the multihop variants are NP-hard. For the latter variants, we present a polynomial-time algorithm based on demand aggregation that achieves a 2-approximation for the number of pigeons used. We believe that this pigeon-based perspective offers both amusing and instructive insights into network design and hopefully, into ornithology.
Multiple studies have shown the benefits of industry-academia collaborations. In this editorial, we discuss the industry-academic research collaboration from the perspective of SIDN, a Netherlandsbased privately held small tech company with a public-interest mission, which operates the .nl top-level domain. We present and compare five models for industry-academia collaboration and show how each has produced deliverables that benefit not only SIDN and its partners involved but also the broader community, including the SIGCOMM's. We also share lessons learned from building and running an in-house research team. Our goal is to encourage other industry players to pursue open, collaborative research that serves academia, industry, and society alike.
In recent years, learning-based congestion control algorithms (CCAs) show great potential in adaptability and flexibility, particularly in highly dynamic network environments. Offline learning-based CCAs outperform heuristic CCAs by leveraging a large amount of network traces, but may experience performance degradation when deployed in environments significantly different from their training dataset. In contrast, online learning-based CCAs adapt to diverse network environments but require more time to converge. To address these issues, we propose ORC, an advanced CCA that combines the online reinforcement learning-based CCA and the heuristic CCA, which effectively accelerates the convergence process of adaptation to new network environments. Experimental results show that ORC significantly improves throughput and reduces latency compared to the state-of-the-art learning-based CCAs.
eBPF allows developers to write safe operating system extensions, but writing these extensions remains challenging because it requires detailed knowledge of both the extension's domain and eBPF's programming interface. Most importantly, the extension must pass the eBPF verifier. This paper argues that DSLs for extensions should guarantee verifier-safety: valid DSL programs should result in eBPF code that always passes the verifier. This avoids complex debugging and the need for extension developers to be eBPF experts. We show that three existing DSLs for different domains are compatible with verifier-safety. Beyond verifier-safety, practical extension DSLs must also achieve good performance. Inspired by database query optimization, we sketch an approach to creating DSL-specific optimizers capable of maintaining verifier-safety. A preliminary evaluation shows that optimizing verifier-safe extension performance is feasible.
The ongoing Russia-Ukraine conflict has significantly impacted global Internet infrastructure. In March 2022, Cogent Communications, a major Tier-1 ISP, voluntarily disconnected from Russian networks, raising critical questions about the resilience and adaptive capacity of Internet routing under geopolitical pressures. In this study, we present a comprehensive, data-driven analysis of Cogent's disconnection by leveraging BGP routing data from RIPE RIS, RouteViews, and PCH over a three-year period (one year before and two years after the event). Our multi-dimensional evaluation quantifies the impact on network reachability, examines AS path modifications, and delineates the timeline of connectivity recovery. The findings reveal heterogeneous recovery dynamics among Russian ASes: while some experienced substantial disruptions-manifested as significant declines in Cogent-dependent reachability and increased AS path lengths-others maintained stable connectivity by rapidly rerouting via alternative transit providers. Our results suggest that Cogent's disconnection was primarily symbolic rather than a permanent technical blockade. These findings contribute to a deeper empirical understanding of ISP de-peering dynamics and offer preliminary implications for improving network resilience in the face of geopolitical disruptions.
The scalability of distributed Machine Learning (ML) systems heavily relies on efficient data exchange between multiple devices, which is typically achieved through collective operations such as allgather, reduce-scatter and allreduce. Many ML models, including transformer-based models, exhibit significant sparsity in activations and gradients that are exchanged through these collectives. In this work, we introduce lightweight sparse collectives designed to exploit this data sparsity, aiming to minimize the communication volume while keeping the overhead of sparsity-aware compression and decompression low. These sparse collectives deliver substantial improvements in collective performance, significantly reducing their completion time. Experiments on AMD Instinct(TM) MI210 and MI300X GPU nodes demonstrate up to 2.96x allreduce, 2.6x allgather, and 2.85x reduce-scatter speedup. These results highlight the potential of sparse collectives to accelerate large-scale distributed training and inference systems.
Remote Direct Memory Access (RDMA) has become a cornerstone technology in modern datacenter networks due to its high throughput and extremely low latency. However, recent works have revealed that congestion arises in the "last mile" of the RDMA I/O path--between DRAM and CPU registers--due to inefficiencies in the memory hierarchy, where severe cache misses and memory bandwidth contention degrade performance. We identify the root cause of this I/O congestion as the speed mismatch between network ingress and CPU processing, which leads to data accumulation and, eventually, last-level cache overflow. To address this, we propose CARC, a credit-based rate control mechanism that dynamically aligns network ingress speed with CPU processing speed. Our preliminary evaluation on eRPC over RDMA, a widely used RPC framework, demonstrates that CARC effectively mitigates I/O congestion, reducing flow completion time by up to 1.40x and improving throughput by up to 1.35x compared to prior work.
eBPF offers a lightweight method to extend the Linux kernel without modifying the source code in existing modules. However, writing correct and efficient eBPF programs is hard due to its unique verifier constraints and cumbersome debugging processes specific to the kernel execution environment. To tackle such an obstacle, we present a system, SimpleBPF, aiming at offloading the tedious eBPF development task. Developers only need to express their intent in a high-level domain-specific language, while the underlying eBPF code generation is handled automatically. SimpleBPF integrates four key components: a concise DSL, an LLM-based generator, a semantic checker, and an LLM-based optimizer. We use few-shot prompting to build both the code generator and optimizer in SimpleBPF, and evaluate the system on programs written in a representative DSL. The preliminary evaluation result shows that SimpleBPF can generate valid eBPF programs that pass the kernel verifier and exhibit competitive runtime performance. We also outline future directions based on current findings.
The disaggregation enabled by Open Radio Access Network (O-RAN) technology offers unprecedented flexibility for multi-tenant deployments on shared infrastructure, making it a promising solution for neutral hosts managing RAN resources across multiple operators. However, this flexibility introduces new challenges in resource allocation and cost optimization. This paper presents a programmable platform designed to help neutral hosts dynamically allocate resources to minimize energy and infrastructure costs while meeting tenant performance needs. We propose a lightweight heuristic to validate the platform's ability to adapt CPU core allocation and DU-level power consumption in real time, using Intel SST-CP and CPU power limit adjustments. The architecture is implemented on a bare-metal OKD cluster, offering a cost-effective and reconfigurable foundation for practical O-RAN experimentation.
The surge in private deployment of large language models (LLMs) driven by open-source advancements has intensified challenges in computational scalability, infrastructure costs, and data privacy. While cloud-edge collaborative inference frameworks alleviate local resource constraints through elastic cloud offloading, their efficacy in wide-area networks (WANs) is hindered by communication inefficiencies and privacy risks. This paper proposes CROSS-SEC, a novel cloud-edge collaborative inference framework integrating cross-WANs PD disaggregation with split learning (SL) for data security preservation. To mitigate transmission bottlenecks, CROSS-SEC introduces a layerwise KVCache computation-communication overlapping mechanism, coupled with asychromous concurrent transmission to eliminate ACK-induced latency. For congestion control, a dual-grained scheduling strategy is proposed: (1) KVCache-level priority scheduling across multi-user/multi-prefill requests ensures first-come-first-serve processing, and (2) latency-sensitive prioritization of latent variables over KVCache transfers guarantees TPOT compliance with SLAs. Experimental validation demonstrates that CROSS-SEC reduces TTFT by 19.83% and improves throughput by 3.63% compared to state-of-the-art frameworks, while maintaining data privacy through SL-based input/output confinement.
Recent advances in programmable optics have shown great promise in providing runtime control over a network's topological behavior to achieve spatial adaptability (e.g., dynamically provision new wavelengths). At the same time, the emergence of programmable data plane technologies has revolutionized how a network's forwarding behavior can be controlled at runtime to accomplish temporal flexibility (e.g., "on-the-fly" traffic aggregation). Unfortunately, a lingering chasm between optical systems and digital packet systems researchers prevents modern-day network applications to simultaneously benefit from both of these exciting developments. To overcome this divide, we propose in this paper ShapeShifter, a novel and principled approach towards integrating programmability in both packet and optical layers and jointly realizing spatial adaptability and temporal flexibility in practice. To provide the necessary technological foundation for this integration, ShapeShifter relies on recent progress in runtime programmability in both communities.
The increasing adoption of Artificial Intelligence (AI), particularly large language (LLMs) and vision-language models (VLMs) has led to a sharp rise in energy demand. While most of the studies predominantly assess energy consumption during training and inference, they often neglect the energy required to transport contextual data---such as text, images, or video---from far-edge devices to AI models, especially over mobile networks. We measure and analyze energy consumption for AI inference both on model-level and network-level. Our approach leverages a combined cross-layer and in-band network telemetry approach to estimate application-level energy usage. Our experiments show that the energy used by the network can be on par with that used by energy efficient AI models for certain tasks. Furthermore, we also estimate the total CO2 emissions of these inference workflows. These results highlight the critical need to incorporate network consumed energy into sustainable AI system design.
Modern large language model (LLM) serving systems increasingly adopt a prefill-decode disaggregation architecture to enhance inference efficiency. While this design improves resource utilization, it introduces latency due to the transfer of key-value (KV) cache. The community has generally assumed that this latency is bandwidth-bound and can be effectively mitigated by high-speed interconnects. In this paper, however, we reveal a contrasting observation: in geo-distributed deployments, KV cache transfer is predominantly round-trip time (RTT)-bound, resulting in significantly reduced effective throughput even when physical bandwidth is abundant. This performance bottleneck stems from the sequential transmission of non-contiguous memory blocks allocated by Paged Attention. Through analysis, we show how current block-wise transmission mechanisms collapse under high RTT conditions, mirroring the classic TCP small send window problem. We conclude by outlining promising directions for overcoming these limitations and enabling efficient LLM inference in geo-distributed environments.
eBPF enables high-performance kernel-level execution by eliminating networking stack traversal and context switching. Despite the advantages, eBPF applications face strict memory management constraints due to the eBPF verifier requirements that mandate static memory allocation. This limitation imposes a fundamental tradeoff between application performance and memory efficiency, ultimately restricting the potential of eBPF. We present Kerby, a dynamic memory pool allocator for eBPF that enables eBPF applications to dynamically manage pre-allocated memory by representing variable-length data as collections of fixed-size blocks. This allows applications to increase the amount of kernel-resident data while minimizing internal fragmentation. Our preliminary evaluation with key-value store implementations demonstrates that Kerby achieves significant improvements in both memory utilization and throughput.
Modern network function (NF) deployments face a fundamental trade-off: kernel-based extended Berkeley Packet Filter (eBPF) NFs provide safety, portability, and an extensive tooling ecosystem, but are limited in performance, while kernel-bypass frameworks deliver high throughput but lack integrated verification and ease of deployment. We present uXDP, a new runtime that unifies these worlds by running unmodified, verified XDP programs in userspace. uXDP ensures compatibility and preserves the verification-driven safety, portability, and familiar workflows of eBPF while moving execution into the userspace, enabling more aggressive optimizations and flexibility. Without recompiling eBPF code, uXDP achieves throughput gains of up to 3.3x over in-kernel execution and improves Meta's Katran load balancer performance by 40%, all while retaining the trusted eBPF development model and deployment simplicity.
