Dirk Kutscher

Our paper on NetSenseML: Network-Adaptive Compression for
Efficient Distributed Machine Learning has been accepted at the 31st International European on Parallel and Distributed Computing (Euro-Par-2025).

Abstract:
Training large-scale distributed machine learning models imposes considerable demands on network infrastructure, often resulting in sudden traffic spikes that lead to congestion, increased latency, and reduced throughput, which would ultimately affect convergence times and overall training performance. While gradient compression techniques are commonly employed to alleviate network load, they frequently compromise model accuracy due to the loss of gradient information.

This paper introduces NetSenseML, a novel network adaptive distributed deep learning framework that dynamically adjusts quantization, pruning, and compression strategies in response to real-time network conditions. By actively monitoring network conditions, NetSenseML applies gradient compression only when network congestion negatively impacts convergence speed, thus effectively balancing data payload reduction and model accuracy preservation.

Our approach ensures efficient resource usage by adapting reduction techniques based on current network conditions, leading to shorter convergence times and improved training efficiency. We present the design of the NetSenseML adaptive data reduction function and experimental evaluations show that NetSenseML can improve training throughput by a factor of 1.55x to 9.84x compared to state-of-the-art compression-enabled systems for representative DDL training jobs in bandwidth-constrained conditions.

References

Yisu Wang, Xinjiao Li, Ruilong Wu, Huangxun Chen, Dirk Kutscher; NetSenseML: Network-Adaptive Compression for Efficient Distributed Machine Learning; 31st International European on Parallel and Distributed Computing (Euro-Par-2025); August 2025; accepted for publication

Our paper on Trochilus, titled Learning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane has been accepted at USENIX ATC-2025. This is joint work with Qing LI's group at Peng Cheng Lab, and the first author is Guanglin DUAN.

Abstract:
Pattern matching is critical in various network security applications. However, existing pattern matching solutions struggle to maintain high throughput and low cost in the face of growing network traffic and increasingly complex patterns. Besides, managing and updating these systems is labor intensive, requiring expert intervention to adapt to new patterns and threats. In this paper, we propose Trochilus, a novel framework that enables high-throughput and accurate pattern matching directly on programmable data planes, making it highly relevant to modern large-scale network systems. Trochilus innovated by combining the learning ability of model inference with the high-throughput and cost-effective advantages of data plane processing. It leverages a byte-level recurrent neural network (BRNN) to model complex patterns, preserving expert knowledge while enabling automated updates for sustained accuracy. To address the challenge of limited labeled data, Trochilus proposes a semi-supervised knowledge distillation (SSKD) mechanism, converting the BRNN into a lightweight, data-plane-friendly soft multi-view forest (SMF), which can be efficiently deployed as match-action tables. Trochilus minimizes the need for expensive TCAM through a novel entry cluster algorithm, making it scalable to large network environments. Our evaluations show that Trochilus achieves multi-Tbps throughput, supports various pattern sets, and maintains high accuracy through automatic updates.

References

• Guanglin Duan, Yucheng Huang, Zhengxin Zhang, Qing Li, Dan Zhao, Zili Meng, Dirk Kutscher, Ruoyu Li, Yong Jiang, and Mingwei Xu. Learning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane. Usenix ATC 2025. accepted for publication
• Extended Summary by Peng Cheng Lab

Our paper on Rethinking Dynamic Networks and Heterogeneous Computing with Automatic Parallelization has been accepted by the 9th Asia-Pacific Workshop on Networking (APNET'25).

Abstract:
Hybrid parallelism techniques are crucial for the efficient training of large language models (LLMs). However, these techniques often introduce differentiated computational and communication tasks across nodes. Existing automatic parallel planning frameworks typically fail to consider both node heterogeneity and dynamic changes in network topology simultaneously, limiting their practical performance. In this paper, we address this issue by positioning heterogeneous nodes within dynamic network environments and employing a simulator to identify optimal parallel strategies. Our approach achieves fine-grained workload distribution in scenarios featuring node heterogeneity and complex networks, while also matching state-of-the-art performance in regular topologies and stable network conditions. Moreover, to mitigate the excessively long search times caused by large search spaces in existing frameworks, we propose a strategy pruning technique to rapidly eliminate infeasible parallel configurations. We further accelerate the search process by executing search tasks in parallel within the simulator. Preliminary evaluation results demonstrate that our method significantly improves training performance on heterogeneous nodes, and the proposed dynamic network design offers enhanced adaptability for complex scenarios such as cloud computing environments.

References

Ruilong Wu, Xinjiao Li, Yisu Wang, Xinyu Chen, Dirk Kutscher; Rethinking Dynamic Networks and Heterogeneous Computing with Automatic Parallelization; The 9th Asia-Pacific Workshop on Networking (APNET'25); August 2025; accepted for publication

Our paper on ViFusion: In-Network Tensor Fusion for Scalable Video Feature Indexing has been accepted at the ACM International Conference on Multimedia Retrieval 2025 (CCF-B).

Abstract:
Large-scale video feature indexing in datacenters is critically dependent on efficient data transfer. Although in-network computation has emerged as a compelling strategy for accelerating feature extraction and reducing overhead in distributed multimedia systems, harnessing advanced networking resources at both the switch and host levels remains a formidable challenge. These difficulties are compounded by heterogeneous hardware, diverse application requirements, and complex multipath topologies. Existing methods focus primarily on optimizing inference for large neural network models using specialized collective communication libraries, which often face performance degradation in network congestion scenarios.

To overcome these limitations, we present ViFusion, a communication aware tensor fusion framework that streamlines distributed video indexing by merging numerous small feature tensors into consolidated and more manageable units. By integrating an in-network computation module and a dedicated tensor fusion mechanism within datacenter environments, ViFusion substantially improves the efficiency of video feature indexing workflows. The deployment results show that ViFusion improves the throughput of the video retrieval system by 8–22x with the same level of latency as state-of-the-art systems.

Stay tuned for the pre-print.

References

Yisu Wang, Yixiang Zhu, Dirk Kutscher; ViFusion: In-Network Tensor Fusion for Scalable Video Feature Indexing; The 15th ACM International Conference on Multimedia Retrieval; June 2025; accepted for publication.

Our 2024 paper on Networked Metaverse Systems: Foundations, Gaps, Research Directions has been mentioned as one most popular and impactful papers of the IEEE Open Journal of the Communications Society (OJCOMS) 2024–2025.

References

• https://dirk-kutscher.info/publications/networked-metaverse-systems/
• Y. Zhang, D. Kutscher and Y. Cui, "Networked Metaverse Systems: Foundations, Gaps, Research Directions," in IEEE Open Journal of the Communications Society, vol. 5, pp. 5488-5539, 2024, doi: 10.1109/OJCOMS.2024.3426098.

We have published the report of the Dagstuhl Seminar 24402 on Greening Networking: Toward a Net Zero Internet that took place from September 29th to October 2nd 2024. The seminar discussed the most impactful networking improvements for reducing carbon emissions in three different areas: 1) applications, systems, and stakeholders; 2) network technologies; and 3) lifecycle and control loops. As a major result of the seminar, the following problems and topics for future research were identified: 1) characterizing the Internet footprint on carbon emissions accurately; 2) understanding attributional and consequential accounting of carbon emissions in networked systems; and 3) identifying potential solutions to give network systems more flexibility in better supporting energy grids and connecting to renewable energy sources. One of the concrete results of this seminar is a list of technologies and research opportunities for which we estimated the potential impact and time horizons.

References

• https://dirk-kutscher.info/events/dagstuhl-greening-networking/
• Alexander Clemm, Dirk Kutscher, Michael Welzl, Cedric Westphal, Noa Zilberman, and Simone Ferlin-Reiter. Greening Networking: Toward a Net Zero Internet (Dagstuhl Seminar 24402). In Dagstuhl Reports, Volume 14, Issue 9, pp. 167-192, Schloss Dagstuhl – Leibniz-Zentrum für Informatik (2025) https://doi.org/10.4230/DagRep.14.9.167

Our paper on PacTrain: Pruning and Adaptive Sparse Gradient Compression for Efficient Collective Communication in Distributed Deep Learning has been accepted at the Design Automation Conference DAC (2025) (CCF-A).

Abstract:
Large-scale deep neural networks (DNN) exhibit excellent performance for various tasks. As DNNs and datasets grow, distributed training becomes extremely time-consuming and demands larger clusters. A main bottleneck is the resulting gradient aggregation overhead. While gradient compression and sparse collective communication techniques are commonly employed to alleviate network load, many gradient compression schemes do not achieve acceleration of the training process while also preserving accuracy. This paper introduces PacTrain, a novel framework that accelerates distributed training by combining pruning with sparse gradient compression. Active pruning of the neural network makes the model weights and gradients sparse.

By ensuring the global knowledge of the gradient sparsity among all distributed training workers, we can perform lightweight compression communication without harming accuracy. We show that the PacTrain compression scheme achieves a near-optimal compression strategy while remaining compatible with the all- reduce primitive. Experimental evaluations show that PacTrain improves training throughput by 1.25 to 8.72× compared to state-of-the-art compression-enabled systems for representative vision and language models training tasks under bandwidth-constrained conditions.

Stay tuned for the pre-print.

References

Yisu Wang, Ruilong Wu, Xinjiao Li , Dirk Kutscher; PacTrain: Pruning and Adaptive Sparse Gradient Compression for Efficient Collective Communication in Distributed Deep Learning; Design Automation Conference (DAC) 2025

We updated our Internet Draft draft-irtf-icnrg-reflexive-forwarding-00 on Reflexive Forwarding for CCNx and NDN Protocols.

Current Information-Centric Networking protocols such as CCNx and NDN have a wide range of useful applications in content retrieval and other scenarios that depend only on a robust two-way exchange in the form of a request and response (represented by an Interest-Data exchange in the case of the two protocols noted above). A number of important applications however, require placing large amounts of data in the Interest message, and/or more than one two-way handshake. While these can be accomplished using independent Interest-Data exchanges by reversing the roles of consumer and producer, such approaches can be both clumsy for applications and problematic from a state management, congestion control, or security standpoint. This specification proposes a Reflexive Forwarding extension to the CCNx and NDN protocol architectures that eliminates the problems inherent in using independent Interest-Data exchanges for such applications. It updates RFC8569 and RFC8609.

The recent update includes a generalization of the main protocol specification, so that Reflexive Forwarding can be used in both CCNx and NDN.

On October 10th, 2024, I was invited to give a talk at the 2nd Airbus Workshop on Networking Systems. The workshop largely discussed connected aircraft scenarios and technologies and features talks on security and reliability, IoT sensor fusioning, and future space and 6G network architectures.

My talk was on Connected Aircraft – Network Architectures and Technologies, and discussed relevant scenarios from my perspective, such as passenger services and new aircraft management applications. For the technology discussion, I focused on large-scale low-latency multimedia communication over the expected heterogeneous and dynamic aircraft connectivity networks and discussed current and emerging technologies such as Media over QUIC, ICN.

I also introduced the recently established Low-Altitude Systems and Economy Research Institute at HKUST(GZ), a cross-disciplinary research institute for the low-altitude domain (with similar but not identical requirements) and some of our recent projects such as Named Data Microverse.

The term ‘Metaverse’ often denotes a wide range of existing and fictional applications. Nevertheless, there are actual systems today that can be studied and analyzed. However, whereas a considerable body of work has been published on applications and application ideas, there is less work on the technical implementation of such systems, especially from a networked systems perspective.

In a recently published open access journal article, we share some insights into the technical design of Metaverse systems, their key technologies, and their shortcomings, predominantly from a networked systems perspective. For the scope of this study, we define the ‘Metaverse’ as follows. The ‘Metaverse’ encompasses various current and emerging technologies, and the term is used to describe different applications, ranging from Augmented Reality (AR), Virtual Reality (VR),and Extended Reality (XR) to a new form of the Internet or Web. A key feature distinguishing the Metaverse from simple AR/VR is its inherently collaborative and shared nature, enabling interaction and collaboration among users in a virtual environment.

Building on Existing Platforms and Network Stacks

Most current Metaverse systems and designs are built on existing technologies and networks. For example, massively multiplayer online games such as Fortnite use a generalized client-server model. In this model, the server authoritatively manages the game state, while the client maintains a local subset of this state and can predict game flow by executing the same game code as the server on approximately the same data. Servers send information about the game world to clients by replicating relevant actors and their properties. Commercial social VR platforms such as Horizon Worlds and AltspaceVR use HTTPS to report client-side information and synchronize in-game clocks across users.

Mozilla Hubs, built with A-Frame (a web framework for building virtual reality experiences), uses WebRTC communication with a Selective Forwarding Unit (SFU). The SFU receives multiple audio and video data streams from its peers, then determines and forwards relevant data streams to connected peers. Blockchain or Non-Fungible Token (NFT)-based online games, such as Decentraland, run exclusively on the client side but allow for various data flow models, ranging from local effects and traditional client-server architectures to peer-to-peer (P2P) interactions based on state channels; Upland is built on EOSIO, an open-source blockchain protocol for scalable decentralized applications, and transports data through HTTPS. Connections between peers in Upland are established using TLS or VPN tunnels.

Many studies have focused on improving various aspects of Metaverse systems. For example, EdgeXAR is a mobile AR framework using edge offloading to enable lightweight tracking with six degrees of freedom (DOF) while reducing offloading delay from the user’s view; SORAS is an optimal resource allocation scheme for edgeenabled Metaverse, using stochastic integer programming to minimize the total network cost; Ibrahim et al. explores the issue of partial computation offloading for multiple subtasks in an in-network computing environment, aiming to minimize energy consumption and delay. However, these ideas for offloading computation and rendering tasks to edge platforms often conflict with the existing end-to-end transport protocols and overlay deployment models. Recently, a Deep Reinforcement Learning (DRL)-based multipath network orchestration framework designed for remote healthcare services is presented, automating subflow management to handle multipath networks. However, proposals for scalable multi-party communication would require interdomain multicast services, unavailable on today’s Internet.

Disconnect Between High-Level Concepts and Actual Systems

In practice, there is a significant disconnect between high-level Metaverse concepts, ideas for technical improvements, and systems that are actually developed and partially deployed. A 2022 ACM IMC paper titled Are we ready for metaverse?: a measurement study of social virtual reality platforms analyzes the performance of various social VR systems, pinpointing numerous issues related to performance, communication overhead, and scalability. These issues are primarily due to the fact that current systems leverage existing platforms, protocols, and system architectures, which cannot tap into any of the proposed architectural and technical enhancements, such as scalable multi-party communication, offloading computation, rendering tasks, etc.

Rather than merely layering ‘the Metaverse’ on top of legacy and not always ideal foundations, we consider Metaverse as a driver for future network and web applications and actively develop new designs to that end. In our article, we take a comprehensive systems approach and technically describe current Metaverse systems, focusing on their networking aspects. We document the requirements and challenges of Metaverse systems and propose a principled approach to system design for these requirements and challenges based on a thorough understanding of the needs of Metaverse systems, the current constraints and limitations, and the potential solutions of Internet technologies.

Article Overview

1. We present a technical description of the ‘Metaverse’ based on existing and emerging systems, including a discussion of its fundamental properties, applications, and architectural models.
2. We comprehensively study relevant enabling technologies for Metaverse systems, including HCI/XR technologies, networking, communications, media encoding, simulation, real-time rendering and AI. We also discuss current Metaverse system architectures and the integration of these technologies into actual applications.
3. We conduct a detailed requirements analysis for constructing Metaverse systems. We analyze applications specific requirements and identify existing gaps in four key aspects: communication performance, mobility, large-scale operation,and end system architecture. For each area, we propose candidate technologies to address these gaps.
4. We propose a research agenda for future Metaverse systems, based on our gap analysis and candidate technologies discussion. We re-assess the fundamental goals and requirements, without necessarily being constrained by existing system architectures and protocols. Based on a comprehensive understanding of what Metaverse systems need and what end-systems, devices, networks and communication services can theoretically provide, we propose specific design ideas and future research directions to realize Metaverse systems that can meet the expectations often articulated in the literature.

References

• Y. Zhang, D. Kutscher and Y. Cui; Networked Metaverse Systems: Foundations, Gaps, Research Directions; in IEEE Open Journal of the Communications Society, doi: 10.1109/OJCOMS.2024.3426098.
• Tianyuan Yu, Xinyu Ma, Varun Patil, Yekta Kocaogullar, Yulong Zhang, Jeff Burke, Dirk Kutscher, Lixia Zhang; Secure Web Objects: Building Blocks for Metaverse Interoperability and Decentralization; IEEE MetaCom 2024; August 12-14 2024; Hong Kong, China
• Dirk Kutscher, Jeff Burke, Giuseppe Fioccola, Paulo Mendes;
  Statement: The Metaverse as an Information-Centric Network; 10th ACM Conference on Information-Centric Networking (ACM ICN '23); October 9 — 10, 2023, Reykjavik, Iceland
• Giuseppe Fioccola , Paulo Mendes , Jeff Burke , Dirk Kutscher;
  Information-Centric Metaverse; Internet Draft draft-fmbk-icnrg-metaverse-01; Work in Progress; July 2023