Dirk Kutscher

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.

The IETF has recently published an interview with me on the IETF Blog.

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 are organizing the 2025 HKUST Internet Research Workshop (HKIRW) in the week before the IETF-122 meeting in Bangkok. This workshop aims to bring together researchers in computer networking and systems around the globe to a live forum discussing innovative ideas at their early stages. The mission of the workshop is that promising but not-yet-mature ideas can receive timely feedback from the community and experienced researchers, leading them into future IRTF work, Internet Drafts, or IETF working groups.

The workshop will operate like a “one day Dagstuhl seminar” and will focus on discussion and ideas exchange and less on conference-style presentations. The objective is to identify topics and connect like-minded people for potential future collaboration.

Please see https://hkirw.github.io/2025/ for details.

References

• Workshop homepage
• Report from HKIRW-2024

On Monday, December 9th, 2024, we held our Decentralization of the Internet (DIN) workshop at ACM CoNEXT-2024. It brought together network researchers, law and policy experts, and digital right activists to discuss the observed consolidation and centralization of the existing Internet applications, services, and the infrastructure in recent years. This trend has economic as well as technical implications for attributes commonly associated with the Internet, such as user-centricity and permissionless innovations.

The Decentralization of the Internet Research Group (DINRG) of the Internet Research Task Force (IRTF) has been working on identifying the root causes and consequences of Internet centralization at IETF meetings and focused workshops in the past, which has led to significant insights, especially with regard to the centralization of infrastructure and control power. This recent DIN workshop at ACM CoNEXT-2024, organized by my DINRG co-chairs Lixia Zhang and myself, provided a forum for academic researchers to present and discuss on-going efforts on this topic, and to create a greater awareness of this important issue in the broader network research community. The workshop attracted a diverse set of researchers who are working on Internet decentralization in fields such as Internet technologies, economics and law-making. The workshop featured two keynotes, two technical paper presentation sessions, and an interactive panel discussion.

Keynotes

The keynotes were presented by two renowned experts:

Keynote: Cory Doctorow: DISENSHITTIFY OR DIE!

Cory Doctorow, member of the Electronic Frontier Foundation (EFF), gave a talk titled DISENSHITTIFY OR DIE! How computer scientists can halt enshittification to make a new, good internet and condemn today's enshitternet to the scrapheap of history. Cory’s talk vividly explained the historic development of a process that he called enshittification, a process in which the providers of online products and services changed their policies subtly and gradually over time, grabbing the control of user data for profitability. Doctorow also discussed potential remedies and countermeasures, including removing the barriers for users to exit platforms and reinstalling the end-to-end principle in future application developments.

Keynote: Michael Karanicolas: The Fediverse Papers: Constitutional, Governance, and Policy Questions for a New Paradigm of Networking

Michael Karanicolas, the executive director of the UCLA Institute for Technology, Law and Policy, talked about the Fediverse Papers: Constitutional, Governance, and Policy Questions for a New Paradigm of Networking. Michael provided an overview of the history of digital speech and content governance. He highlighted the challenges in supporting effective content moderation in today’s Internet contexts, including issues around monetization, legislation, privacy, and the need for governance mechanisms to meet users, content owners, and governments’ expectations. He emphasized the importance of intentionality and a structured process to identify the essential policy questions and to evaluate various design choices for the future of decentralized platforms.

Decentralized Systems

Bluesky and the AT Protocol: Usable Decentralized Social Media

Authors: Martin Kleppmann, Paul Frazee, Jake Gold, Jay Graber, Daniel Holmgren, Devin Ivy, Jeromy Johnson, Bryan Newbold, Jaz Volpert

Abstract: Bluesky is a new social network built upon the AT Protocol, a decentralized foundation for public social media. It was launched in private beta in February 2023, and has grown to over 10 million registered users by October 2024. In this paper we introduce the architecture of Bluesky and the AT Protocol, and explain how the technical design of Bluesky is informed by our goals: to enable decentralization by having multiple interoperable providers for every part of the system; to make it easy for users to switch providers; to give users agency over the content they see; and to provide a simple user experience that does not burden users with complexity arising from the system’s decentralized nature. The system’s openness allows anybody to contribute to content moderation and community management, and we invite the research community to use Bluesky as a dataset and testing ground for new approaches in social media moderation.

ReP2P Matrix: Decentralized Relays to Improve Reliability and Performance of Peer-to-Peer Matrix

Authors: Benjamin Schichtholz, Roland Bless, Florian Jacob, Hannes Hartenstein, Martina Zitterbart

Abstract: Matrix is a decentralized middleware for low-latency group communication, most renowned for its use in the Element instant messenger. Proposals for peer-to-peer (P2P) Matrix architectures aim to decentralize the current architecture further, which is based on federated servers. These proposals require that the receiver and the originator, or another peer that already successfully received the message, are simultaneously online. We introduce relay-enhanced P2P Matrix (ReP2P Matrix) in order to improve message delivery between peers that are online at different times. The design maintains the advantages of P2P Matrix and integrates well into it, e.g., it reuses existing mechanisms for authentication and authorization. Using an extended real-world group messaging traffic dataset, we evaluate P2P Matrix by comparing it to P2P Matrix without relays. The results show that relays do not only improve reliability in message delivery, but also increase the share of low delivery latencies by 50% points in groups with up to 30 members.

On Empowering End Users in Future Networking

Authors: Tianyuan Yu, Xinyu Ma, Lixia Zhang

Abstract: In today's Internet, end users communicate largely via cloud-based apps, and user data are stored in cloud servers and controlled by cloud providers. Recent years have witnessed multiple efforts in developing decentralized social apps with various design approaches, although the community at large is yet to fully understand the effectiveness, viability, and limitations of these different designs. In this paper, we make a proposition that a necessary condition of moving towards Internet decentralization is enabling direct user-to-user (U2U) communications, and discuss the design choices in several decentralization efforts and identify their limitations. We then articulate why a DNS-derived namespace is the best choice in U2U app developments in general, and use a recently developed decentralized app, NDN Workspace (NWS), as an example to show how NWS' use of DNS-derived namespace enables secure U2U communications.

Technologies for Decentralization

Atomicity and Abstraction for Multi-Blockchain Interactions

Authors: Huaixi Lu, Akshay Jajoo, Kedar S. Namjoshi

Abstract: A blockchain enables secure, atomic transactions among untrusted parties. Atomicity is not guaranteed, however, for transactions whose operations span several blockchains; multi-chain atomicity must be enforced by a protocol. Such protocols are known only for special cases, such as cryptocurrency swaps, which are limited only to two chains. We propose a novel two-phase protocol that facilitates atomic executions of general multi-chain (>= 2) transactions. We formally analyze the protocol correctness and show that the proposed abstraction considerably simplifies the development of multi-chain applications. Our experiments with a prototype implementation show that the performance of the general atomicity protocol is comparable to that of custom-built implementations.

Communication Cost for Permissionless Distributed Consensus at Internet Scale

Authors: David Guzman, Dirk Trossen, Jörg Ott

Abstract: The diffusion of information that evolves a distributed computing state is a fundamental operation of a permissionless distributed consensus system (DCS). This permissionless participation decentralized the consensus over the distributed computing state, e.g., in cryptocurrencies and voting systems. For this, a permissionless DCS implements protocols to establish relationships among peers, which is then used to diffuse information. The relation establishment constitutes the control plane of the DCS, while the state diffusion is the data plane. The prevalent mechanism to realize both is a randomized peer-centric iterative diffusion. In this paper, we contrast this approach against a multicast-based design, focusing our comparison on the costs (bytes transmitted) for maintaining the relations, the control plane. We develop suitable models to account for those costs, parameterized through Internet-scale experimental insights we derived from existing DCS deployments. Our results show that the communication costs can be reduced by 30 times.

Towards a Decentralized Internet Namespace

Authors: Yekta Kocaogullar, Eric Osterweil, Lixia Zhang

Abstract: The Domain Name System (DNS) has been providing a decentralized global namespace to support all Internet applications and usages over the last few decades. In the recent years, a number of blockchain-based name systems have emerged with the claim of providing better namespace decentralization than DNS. The community at large seems uncertain with regard to which of these systems is the best in providing decentralized Internet namespace control. In this paper, we first deconstruct the design of DNS, identify its three essential components and explain who controls each of them. We then examine the Ethereum Name Service (ENS) as a representative example of blockchain-based naming systems, gauge the degree of its decentralization. Finally, we conduct a comparative analysis between DNS and ENS to assess the validity and affordability of each design and the (de)centralization in their namespace control and name system operations.

Panel Discussion: Decentralization of the Internet – Quo Vadis?

An interactive panel discussion with (from left to right) Michael Karanicolas (UCLA), Paul Mockapetris (ThreatSTOP), Dan Massey (USC ISI, NSF), and Cory Doctorow (EFF), articulated various next steps for countering Internet centralization. Among many things discussed, the panel and audience identified the notion of enabling direct user-to-user communication without reliance on third parties, and the required functionality to support that, such as how to provide user owned identities, tools for user mutual authentications and secure communications.

These and additional related topics will be further discussed at the IRTF DIN research group, which is a forum with open participation to serve the purpose of continuous international collaborative research on Internet decentralization.

References

• Workshop homepage at ACM CoNEXT-2024
• Workshop papers
• Workshop videos
• IETF blog post on the workshop
• IRTF DINRG

I am delighted that I have been appointed as the next Chair of the Internet Research Task Force (IRTF) by the Internet Architecture Board (IAB).

I have been involved in the IRTF for many years. It is a unique organization that conducts research of importance to the evolution of the Internet protocols, applications, architecture and technology. It has initiated and supported many important technology developments for the Internet in the past, in fields such as network architecture, security and privacy, congestion control, and many more.

The IRTF focuses on longer term research issues, and its various research groups are enabling international collaboration for continuous research on critical topics for the Internet by working with academic and industry research communities.

My term starts in March 2025. I am sincerely grateful for all the support I have received, I am looking forward to working with this community to help making the Internet work better through good research work.

References

• https://www.iab.org/announcements/dirk-kutscher-to-be-next-internet-research-task-force-chair/

Our ACM CoNEXT-2024 workshop on the decentralization of the Internet on Monday, December 9th 2024 in LA has an exciting agenda – don't miss it! Check out the workshop homepage for up-to-date information.

09:00 Session 1: Keynotes

1. Keynote by Cory Doctorow: DISENSHITTIFY OR DIE! How computer scientists can halt enshittification to make a new, good internet and condemn today's enshitternet to the scrapheap of history.
2. Keynote by Michael Karanicolas: The Fediverse Papers: Constitutional, Governance, and Policy Questions for a New Paradigm of Networking

11:00 Session 2: Decentralized Systems

1. Martin Kleppmann, et al.; Bluesky and the AT Protocol: Usable Decentralized Social Media
2. Benjamin Schichtholz et al.; ReP2P Matrix: Decentralized Relays to Improve Reliability and Performance of Peer-to-Peer Matrix
3. Tianyuan Yu et al.; On Empowering End Users in Future Networking

14:00 Session 3: Technologies for Decentralization

1. Huaixi Lu et al.; Atomicity and Abstraction for Multi-Blockchain Interactions
2. David Guzman et. el; Communication Cost for Permissionless Distributed Consensus at Internet Scale
3. Yekta Kocaogullar et al.; Towards a Decentralized Internet Namespace

15:00 Session 4: Decentralization of the Internet – Quo Vadis?

• Organizers: Lixia Zhang & Dirk Kutscher
• Interactive panel discussion with Cory Doctorow, Michael Karanicola, and paper authors