arXiv subjects

Yannis Thomas

Publications and source records attributed to Yannis Thomas.

6 recordsLinked to original sources

arXiv2026

Split Federated Learning Architectures for High-Accuracy and Low-Delay Model Training

Can we find a network architecture for ML model training so as to optimize training loss (and thus, accuracy) in Split Federated Learning (SFL)? And can this architecture also reduce training delay and communication overhead? While accuracy is not influenced by how we split the model in ordinary, state-of-the-art SFL, in this work we answer the questions above in the affirmative. Recent Hierarchical SFL (HSFL) architectures adopt a three-tier training structure consisting of clients, (local) aggregators, and a central server. In this architecture, the model is partitioned at two partitioning layers into three sub-models, which are executed across the three tiers. Despite their merits, HSFL architectures overlook the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay, and overhead. This work explicitly captures the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay and overhead by formulating a joint optimization problem. We prove that the problem is NP-hard and propose the first accuracy-aware heuristic algorithm that explicitly accounts for model accuracy, while remaining delay-efficient. Simulation results on public datasets show that our approach can improve accuracy by 3%, while reducing delay by 20% and overhead by 50%, compared to state-of-the-art SFL and HSFL schemes.

cs.LG↗

arXiv2025

Collaborative Split Federated Learning with Parallel Training and Aggregation

Federated learning (FL) operates based on model exchanges between the server and the clients, and it suffers from significant client-side computation and communication burden. Split federated learning (SFL) arises a promising solution by splitting the model into two parts, that are trained sequentially: the clients train the first part of the model (client-side model) and transmit it to the server that trains the second (server-side model). Existing SFL schemes though still exhibit long training delays and significant communication overhead, especially when clients of different computing capability participate. Thus, we propose Collaborative-Split Federated Learning~(C-SFL), a novel scheme that splits the model into three parts, namely the model parts trained at the computationally weak clients, the ones trained at the computationally strong clients, and the ones at the server. Unlike existing works, C-SFL enables parallel training and aggregation of model's parts at the clients and at the server, resulting in reduced training delays and commmunication overhead while improving the model's accuracy. Experiments verify the multiple gains of C-SFL against the existing schemes.

cs.DC↗

arXiv2018

IPTV Over ICN

The efficient provision of IPTV services requires support for IP multicasting and IGMP snooping, limiting such services to single operator networks. Information-Centric Networking (ICN), with its native support for multicast seems ideal for such services, but it requires operators and users to overhaul their networks and applications. The POINT project has proposed a hybrid, IP-over-ICN, architecture, preserving IP devices and applications at the edge, but interconnecting them via an SDN-based ICN core. This allows individual operators to exploit the benefits of ICN, without expecting the rest of the Internet to change. In this paper, we first outline the POINT approach and show how it can handle multicast-based IPTV services in a more efficient and resilient manner than IP. We then describe a successful trial of the POINT prototype in a production network, where real users tested actual IPTV services over both IP and POINT under regular and exceptional conditions. Results from the trial show that the POINT prototype matched or improved upon the services offered via plain IP.

cs.NI↗

arXiv2018

IP Over ICN Goes Live

Information-centric networking (ICN) has long been advocating for radical changes to the IP-based Internet. However, the upgrade challenges that this entails have hindered ICN adoption. To break this loop, the POINT project proposed a hybrid, IP-over-ICN, architecture: IP networks are preserved at the edge, connected to each other over an ICN core. This exploits the key benefits of ICN, enabling individual network operators to improve the performance of their IP-based services, without changing the rest of the Internet. We provide an overview of POINT and outline how it improves upon IP in terms of performance and resilience. Our focus is on the successful trial of the POINT prototype in a production network, where real users operated actual IP-based applications.

cs.NI↗

arXiv2017

Multi-Flow Congestion Control with Network Assistance

A well-known technique for enhancing the performance and stability of content distribution is the use of multiple dissemination flows. Multipath TCP (MPTCP), the most popular multiflow protocol on the Internet, allows receivers to exploit multiple paths towards a single sender. Nevertheless, MPTCP cannot fully exploit the potential gains of multipath connectivity, as it must fairly share resources with (single-flow) TCP, without a clear understanding of whether the available paths do share any bottleneck links. In this paper, we introduce a hybrid congestion control algorithm for multisource and multipath transport that enables higher bandwidth utilization compared to MPTCP, while remaining friendly to TCP-like flows. Our solution employs (i) an in-network module that offers essential topological information and (ii) Normalized Multiflow Congestion Control (NMCC), a novel end-to-end congestion control algorithm. While NMCC is architecture-independent and the in-network module can be adapted for Multi-Protocol Label Switching (MPLS) or Software Defined Networks (SDNs), our prototype was implemented on the Publish-Subscribe Internetworking (PSI) architecture, which offers centralized path formation and source routing. Using an actual protocol implementation deployed on our test-bed, we provide experimental results which validate the effectiveness of our design in terms of performance, adaptation to shifting network conditions and friendliness to other flows.

cs.NI↗

arXiv2016

Object-oriented Packet Caching for ICN

One of the most discussed features offered by Information-centric Networking (ICN) architectures is the ability to support packet-level caching at every node in the network. By individually naming each packet, ICN allows routers to turn their queueing buffers into packet caches, thus exploiting the network's existing storage resources. However, the performance of packet caching at commodity routers is restricted by the small capacity of their SRAM, which holds the index for the packets stored at the, slower, DRAM. We therefore propose Object-oriented Packet Caching (OPC), a novel caching scheme that overcomes the SRAM bottleneck, by combining object-level indexing in the SRAM with packet-level storage in the DRAM. We implemented OPC and experimentally evaluated it over various cache placement policies, showing that it can enhance the impact of ICN packet-level caching, reducing both network and server load.

cs.NI↗