SCONE Meeting 29-10-2025

Date and time

From midday on Wednesday 29th October 2025

Venue

Abertay cyberQuarter
1-3 Bell Street
Dundee
DD1 1LH

How to get here

abertay cyberQuarter is around 10 minutes walk from Dundee Bus Station, and around 15 minutes from the train station. Both have regular services from Aberdeen, Glasgow, and Edinburgh.

You can access the building through the door at the Euclid Crescent corner, or through Annie’s Café. When you reach the model SCADA Factory, then you are in the right place. If in doubt, ask at the reception desk in the Kydd Building on the opposite side of Bell Street.

Registration

Please register here by Friday 17th October, indicating whether you would like to give a talk, and any dietary requirements or preferences that you have.

Organisers

Gregor Haywood, Stephen McQuistin

Schedule

  • 12.00-13.00: Arrival, lunch, and welcome
  • 13.00-14.15: Session 1
    • “5G Research, Projects and Use Cases for Ultra-Low Latency and Cybersecurity Applications” — Laith Al-Jobouri, Abertay University
      • This presentation will explore advanced research methodologies, experimental deployments, and practical implementations of 5G private network architectures and 5G in a box, with a particular emphasis on their capacity to deliver ultra-low latency performance across a range of latency-sensitive domains. Key application areas include competitive E-gaming environments, real-time virtual production workflows, mission-critical search and rescue operations, and live immersive experiences such fan zone for live events. Additionally, the talk will focus on how the 5G technologies intersect with cybersecurity frameworks, addressing some of the challenges related to secure data, and threat detections in high-performance edge computing scenarios.
    • “On Correctness in Distributed Systems with Session Types” — Leonid Nosovitskiy, University of St Andrews
      • Communication protocols are ubiquitous in distributed systems. Errors in protocol implementations impose serious consequences such as financial losses and availability of correctly working services. Such errors can arise from both ambiguous specifications and the presence of common concurrency issues. Session Types help prevent such errors by verifying protocol specifications and guiding the implementations thereof. However, Session Types are often limited by unrealistic assumptions. In this talk, we discuss the challenges imposed by these limitations and present tooling to facilitate the use of Session Types in a real-world distributed setting. 
    • “Through Misfortune or Carelessness: Reassuring Reliability Through Repetition” — Adam Barwell, University of St Andrews
    • “xMem: A CPU-Based Approach for Accurate  GPU Memory Prediction in Deep Learning Training” — Jiabo Shi, University of Glasgow
      • The widespread adoption of Deep Learning (DL) in diverse application areas has significantly increased the demand for GPUs. Consequently, GPU resources are scarce and are managed in clusters to maximize resource utilization. However, this shift introduces new debugging challenges when training DL models on shared clusters particularly Out-Of-Memory (OOM) errors, an issue commonly reported in industry and academic literature. Existing solutions for avoiding OOM primarily rely on static analysis of the DL model’s computational graph, or leverage GPU resources directly or indirectly to estimate the peak memory required for training the given task on the target GPU. Unfortunately, relying on GPUs for these predictions exacerbates resource contention and increases scheduling challenges. Furthermore, the dynamic nature of model development limits the accuracy of static analysis to estimate peak memory usage. To address these limitations, we propose xMem, a novel tool that uses CPU-based analysis to accurately predict the memory required for model training on a GPU. By eliminating the reliance on GPUs for memory estimation, xMem promotes efficient GPU utilization while mitigating OOM errors. Our empirical evaluation of 16 DL models (a total of 5,040 runs) demonstrates that, compared to state-of-the-art GPU memory estimators, xMem decreases the median relative error by 84.32%, reduces the average probability of estimation failure by 73.44%, accelerates the runtime by 50.16%, and improves memory conservation by 125.36%.
  • 14.15-14.30: SICSA Update — Tristan Henderson, University of St Andrews
  • 14.30-15.00: Tea & coffee break
  • 15.00-16.15: Session 2
    • “Genetic Algorithms for Service Function Chain Embedding” — Theviyanthan Krishnamohan, University of Glasgow
      • Service Function Chains (SFCs) virtualise network functions, and embed them on a physical network and route traffic through them programmatically. Optimally embedding SFCs on a physical traffic is an NP-hard problem. We propose the use of Genetic Algorithms, which are metaheuristic algorithms to tackle this optimisation problem. 
    • “OptoFlood: Controllable Flooding for NDN Producer Mobility” — Yuting Wan, University of Glasgow
      • Producer mobility in systems that use Named Data Networking (NDN) protocols can lead to increased latency and packet loss due to problems with slow routing convergence. This affects applications, such as live video streaming and video conferencing, where even a brief interruption in packet delivery can degrade user experience. To address this, we propose OptoFlood, a controllable flooding mechanism that supports Data and Interest packet delivery immediately following producer movement events, to rapidly establish temporary bidirectional communication paths. OptoFlood also accelerates global routing convergence through integration with the Named-data Link State Routing (NLSR) protocol.
    • “Opening Up Kernel-Bypass TCP Stacks” — Michio Honda, University of Edinburgh
      • We have seen a surge of kernel-bypass network stacks with different design decisions for higher throughput and lower latency than the kernel stack, but how do they perform in comparison to each others in a variety of workload, given that modern stacks have to handle both bulk data transfers over multi-hundred gigabit Ethernet and small request-response messages that require low latency? We found that even representative kernel-bypass stacks have never been compared for a set of basic workloads, likely because of difficulty to run their implementation. This paper takes the first step towards answering that question by comparing six in-kernel or kernelbypass stacks. We show that existing stacks cannot handle those workloads at the same time or lack generality. We then use those observations to discuss possible pathways towards practical kernel-bypass stacks. The paper has been presented in USENIX ATC’25: https://micchie.net/files/bypass-atc25.pdf
    • “Remote Connection Offload with XO” — Steven Chien, University of Oxford
      • Layer 7 load balancers (L7LBs) play an important role in per-request server selection. However, L7LBs introduce substantial CPU and network overhead. XO enables an L7LB to offload TCP connections and application request processing to a backend server in a request granularity, outperforming conventional L7LBs in throughput by efficiently utilizing the available server CPU and network resources. We demonstrate XO in two real-world systems, Ceph and Nginx.
  • 16.15-17.00: PhD Rapid Research Introductions
  • 17.00-17.05: Closing
  • 1730: Dinner (tbc)

Cookie preferences