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 allreduce primitive. Experimental evaluations show that PacTrain improves training throughput by 1.25 to $8.72 \times$ compared to state-of-the-art compression-enabled systems for representative vision and language models training tasks under bandwidth-constrained conditions.

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.

Real-time streaming of point cloud video, characterized by massive data volumes and high sensitivity to packet loss, remains a key challenge for immersive applications under dynamic network conditions. While connection-oriented protocols such as TCP and more modern alternatives like QUIC alleviate some transport-layer inefficiencies, including head-of-line blocking, they still retain a coarse-grained, segment-based delivery model and a centralized control loop that limit fine-grained adaptation and effective caching. We introduce INDS (Incremental Named Data Streaming), an adaptive streaming framework based on Information-Centric Networking (ICN) that rethinks delivery for hierarchical, layered media. INDS leverages the Octree structure of point cloud video and expressive content naming to support progressive, partial retrieval of enhancement layers based on consumer bandwidth and decoding capability. By combining time-windows with Group-of-Frames (GoF), INDS's naming scheme supports fine-grained in-network caching and facilitates efficient multi-user data reuse. INDS can be deployed as an overlay, remaining compatible with QUIC-based transport infrastructure as well as future Media-over-QUIC (MoQ) architectures, without requiring changes to underlying IP networks. Our prototype implementation shows up to 80% lower delay, 15-50% higher throughput, and 20-30% increased cache hit rates compared to state-of-the-art DASH-style systems. Together, these results establish INDS as a scalable, cache-friendly solution for real-time point cloud streaming under variable and lossy conditions, while its compatibility with MoQ overlays further positions it as a practical, forward-compatible architecture for emerging immersive media systems.

There is a growing trend in deploying DNNs on tiny microcontroller (MCUs) to provide inference capabilities in the IoT. While prior research has explored many lightweight techniques to compress DNN models, achieving overall efficiency in model inference requires not only model optimization but also careful system resource utilization for execution. Existing studies primarily leverage arithmetic logic units (ALUs) for integer-only computations on a single CPU core. Floating-point units (FPU) and multi-core capabilities available in many existing MCUs remain underutilized. To fill this gap, we propose AdaptQNet, a novel MCU neural network system that can determine the optimal precision assignment for different layers of a DNN model. AdaptQNet models the latency of various operators in DNN models across different precisions on heterogeneous processing units. This facilitates the discovery of models that utilize FPU and multi-core capabilities to enhance capacity while adhering to stringent memory constraints. Our implementation and experiments demonstrate that AdaptQNet enables the deployment of models with better accuracy-efficiency trade-off on MCUs.

No abstract available.

No abstract available.

Internet of Things (IoT) is a growing topic of interest. Billions of IoT devices are expected to connect to the Internet in the near future. These devices differ from the traditional devices operated in the Internet. In this paper, we argue that an information-centric networking (ICN), a new networking paradigm, is a more suitable architecture for the IoT compared to the currently prevailing IP-based network. We observe that recent works that propose to use ICN for IoT, either do not cover the need to integrate sensor networks with the Internet to realize IoT or do so inefficiently. There is a need to understand effective ways to integrate the various heterogeneous sensor networks with the Internet without affecting their current mode of operation. In this paper, we study the essential requirements for integrating sensor networks to the Internet. We provide an architecture with gateways for paving a way for the sensor networks to become a part of the IoT family. We further provide a naming schema for efficient operation of the resource constrained sensor networks, discuss mobility, security, communication patterns, and propose the most suitable choices for IoT networks.

The vast majority of mobile data transfers today follow the traditional client-server model. Although in the fixed network P2P approaches have been exploited and shown to be very efficient, in the mobile domain there has been limited attempt to leverage on P2P (D2D) for large-scale content distribution (i.e., not DTN-like, point-to-point message transfers). In this paper, we explore the potential of a user-operated, smartphone-centric content distribution model for smartphone applications. In particular, we assume source nodes that are updated directly from the content provider (e.g., BBC, CNN), whenever updates are available; destination nodes are then directly updated by source nodes in a D2D manner. We leverage on sophisticated information-aware and application-centric connectivity techniques to distribute content between mobile devices in densely-populated urban environments. Our target is to investigate the feasibility of an opportunistic content distribution network in an attempt to achieve widespread distribution of heavy content (e.g., video files) to the majority of the destination nodes. We propose ubiCDN as a ubiquitous, user-operated and distributed CDN for mobile applications.

This paper introduces the concept of dynamic Service Boost and proposes deployment solutions in mobile networks focusing on the 3GPP Long Term Evolution (LTE) architecture. The main idea is to introduce a time bound preferential service to subscribers based on predefined service contracts. By applying a light-weight, dynamic Quality of Service (QoS) control, operators achieve both, efficient network utilization and adequate QoS for users and content/application providers. This helps operators to use resources more efficiently, for example to enable a more efficient capacity sharing in the presence of increasing mobile traffic. We initially investigate the impact of Service Boost on Over-The-Top (OTT) traffic transmitted without any specific QoS guarantees over the so-called default bearer in LTE. Then we consider the design and analysis of a Service Boost architecture and framework for managing and prioritizing service requirements for certain applications within LTE. Finally, we elaborate the realization of Service Boost through congestion accountability.

In this work, we outline how to enable Information-Centric Networking (ICN) on existing IP networks, such as ISP or data center networks, using Software-Defined Networking (SDN) functions and control. We describe a mechanism that requires neither new or extended network/L3 and transport/L4 protocols nor changes of ICN host network stacks, and supports aggregation of routes inside the SDN controlled network. The proposed solution is agnostic of the specific ICN protocol in use, and does not require all network elements to be SDN-enabled. It supports advanced ICN routing features like request aggregation and forking, as well as load-balancing, traffic engineering, and explicit path steering (e. g., through ICN caches). We present the design as well as our first implementation of the proposed scheme-based on the Trema OpenFlow controller-framework and CCNx.

Capacity sharing - the fair usage of the available capacity amongst a set of users - is relying solely on the end user/end hosts in today's Internet. This in principle allows malicious end users to consume disproportional amounts of capacity without obvious ways of stopping or charging them. Re-ECN is a resource sharing framework that enables such resource usage accountability by, amongst other things, employing enforcement points in the network. These are often described as packet droppers at or near the egress of an end-to-end path. This paper presents a packet dropper design that allows implementing the required congestion declaration enforcement efficiently. Our dropper design is based on a hybrid approach providing the effectiveness close to a complete stateful flow monitoring with the efficiency of a more stateless approach. We evaluate the effectiveness and the resource requirements through simulation.

This paper presents an approach to apply congestion exposure concepts to mobile wireless communication, with a particular focus on LTE's E-UTRAN. Assuming Re-ECN as the congestion exposure mechanism, we show how the different functional Re-ECN components can be mapped to logical elements in the E-UTRAN architecture. Moreover we present a novel concept of Wireless ECN, which allows for generating congestion (cost) indication taking wireless resource consumption into account. We show how such a mechanism can be used to enhance throughput and reduce buffering delays in wireless networks with heterogeneous Modulation and Coding Scheme configurations for different users.

Base stations (BSs) are the main energy expenditure elements of cellular networks, considering the high coverage requirements and the fact that the total provisioned capacity is intended to match peak hour traffic demand. In this paper, we introduce energy saving algorithms based on coordination between network elements. We introduce the notion of energy partitions - associations of powered-on and powered-off BSs - to deliver energy saving with the objective of matching offered capacity with the traffic demand in a flexible manner. Our energy saving algorithms are based on shared knowledge of load and coverage information and enable an appropriate cell reconfiguration for achieving a network-level energy saving. Through a simulation-based evaluation, we analyze the performance of centralized and distributed algorithms under different network topologies and traffic conditions, highlight the benefits and drawbacks, and provide recommendations for deployment scenarios.

Internet access for highly mobile users in high-speed trains or cars, underground walkways and tunnels is often challenged by intermittent connectivity with short and unpredictable connectivity opportunities. In such opportunistic communication scenarios, utilizing these connectivity opportunities to the best potential for the mobile user can be crucial to achieve an acceptable overall perceived service quality. When wireless access networks are shared between mobile and non-mobile TCP users, this objective can be difficult to achieve especially considering the long-term throughput balance. In this paper, we analyze the performance issues with current TCP implementations and propose a new rate control mechanism called RACING (Rate Control for Intermittent Networking) for increasing long-term TCP throughput for Intermittent Connectivity Nodes (ICNs) in such challenged networks. In the proposed algorithm, the sending rate is adjusted based on the expected traffic during the outage to, in a sense, compensate the ICN for the disruption. The algorithm also considers the fairness with other flows by gradually changing its algorithm closer to the original TCP according to the compensated amount. Computer simulation results show that RACING reasonably improves the performance of the ICN.

The quality of Internet access for mobile users may suffer from highly variable communication characteristics (packet loss, delay, throughput) and from temporary disconnections. The former occur as a result of changing radio properties, handovers, or variable system load, the latter whenever wireless coverage is insufficient. While quite some research has tackled improving radio coverage to keep users always best connected and numerous approaches pursue improving wireless performance and robustness, these often assume a greenfield deployment or a tight integration with operators. We present the mobility support system architecture developed in the CHIANTI project. While building on related work in various technical respects, the CHIANTI architecture is specifically designed to be instantly and incrementally deployable in today's Internet landscape, considering real-world legal and deployment constraints, and supports roles of different (independent) players. We also report on our proof-of-concept implementation.

Future heterogeneous wireless networks will encompass different link layer technologies and allow selecting the most appropriate network depending on different criteria. To support mobile nodes in the selection process, network information services are developed that provide the mobile node with sufficient information about its network neighborhood, typically focusing on the optimization of handover processes. In this paper, we take a more general approach towards network information services, which is needed to support mobile communications in the existing environments of WLAN hotspots and wide area mobile communications networks. We introduce the notion of "service maps", a mobile data management approach allowing a mobile user to obtain a detailed view of available networks and the services they offer depending on the user context such as geographic position, mobility paths, and application requirements. We review our experiences with networking in highly mobile scenarios from which we derive a set of requirements for a general information service and present our system design and data model that also accounts for business and (incremental) deployment considerations

Today's mobile, wireless, and ad-hoc communications often exhibit extreme characteristics challenging assumptions underlying the traditional way of end-to-end communication protocol design in the Internet. One specific scenario is Internet access from moving vehicles on the road as we are researching in the drive-thru Internet project. Using wireless LAN as a broadly available access technology leads to intermittent - largely unpredictable and usually short-lived - connectivity, yet providing high performance while available. To allow Internet applications to deal reasonably well with such intermittent connectivity patterns, we have introduced a supportive drive-thru architecture. A key component is a "session" protocol offering persistent end-to-end communications even in the presence of interruptions. In this paper, we present the design of the persistent connectivity management protocol (PCMP) and report on findings from our implementation.

This paper reports on measurement results for the use of IEEE 802.11 networks in drive-thru scenarios: we have measured transmission characteristics for sending and receiving high data volumes using UDP and TCP in vehicles moving at different speeds that pass one or more IEEE 802.11 access points at the roadside. We discuss possibilities and limitations for the use of scattered WLAN cells by devices in fast moving vehicles and provide an analysis of the performance that can he expected for the communication in such scenarios. Based on these observations, we discuss implications for higher-layer protocols and applications.