NET4 – Device-Edge-Cloud continuum in B5G/6G era
Wednesday, 5 June 2024, 16:00-17:30, room Gorilla 3
Session Chair: Giada Landi (Nextworks, IT)
Resource Variability Telemetry in Edge Computing
Panagiotis Giannakopoulos (Eindhoven University of Technology, The Netherlands); Bart van Knippenberg (Thermo Fisher Scientific, The Netherlands); Kishor Chandra Joshi (TU Eindhoven, The Netherlands); Nicola Calabretta (COBRA Research Institute, The Netherlands); George Exarchakos (Eindhoven University of Technology, The Netherlands)
Edge computing is employed as a common execution platform for a variety of network functions and user applications. Consequently, the platform needs to satisfy a variety of execution time requirements. Towards a reliable edge computing, this study proposes a method to detect the factors that influence the variability of task execution time at the edge. A set of Single Particle Analysis (SPA) algorithms for an electro-microscopy use case are executed in a Kubernetes cluster, monitored by Prometheus. Our experiments show a big increase on the variability of the round trip time of a task running over shared resources. Our methodology selects the most related performance metrics among 554 captured metrics, where the correlation reaches 87%.
Uncore Frequency Tuning for Energy Efficiency in Mixed Priority Cloud Native Edge
Paul Veitch (BT & Bt, United Kingdom (Great Britain)); Chris MacNamara and John J Browne (Intel Corporation, Ireland)
A key challenge for telco cloud deployments is achieving a suitable balance of performance for the mix of workloads, and the power consumption of the compute infrastructure hosting those workloads. Recent generations of Intel x86 servers have a broad suite of technologies aimed at granular tuning of the multi-core processor resources: Cache Allocation Technology (CAT), Memory Bandwidth Allocation (MBA) and Speed Select Technology (SST) are key examples. Further fine-tuning is possible by exploiting frequency scaling properties of the “uncore” component of the CPU which include the last level cache and memory bandwidth control units- i.e., these are distinct from the part of the processor socket silicon footprint where “cores” reside. Such tuning capabilities can realise deterministic performance in the presence of mixed priority workloads which is a major consideration for more resource-constrained environments in the cloud compute continuum, e.g., the network edge where a mix of Containerised Network Functions (CNFs) can co-exist with other workload types. This paper explores a method of combining CAT, MBA, SST and uncore frequency tuning in tandem to achieve improved performance and power efficiency where a high priority workload co-exists with a “Noisy Neighbour” low priority workload. In one specific case, a performance increase of 29.6% is achieved in conjunction with a power reduction of 18.2% when compared with the baseline “untuned” set-up.
TF2AIF: Facilitating Development and Deployment of Accelerated AI Models on the Cloud-Edge Continuum
Aimilios Leftheriotis (University of Patras & National Technical University of Athens, Greece); Achilleas Tzenetopoulos, George Lentaris and Dimitrios Soudris (National Technical University of Athens, Greece); Georgios Theodoridis (University of Patras, Greece)
The B5G/6G evolution relies on connect-compute technologies and highly heterogeneous clusters with HW accelerators, which require specialized coding to be efficiently utilized. The current paper proposes a custom tool for generating multiple SW versions of a certain AI function input in high-level language, e.g., Python TensorFlow, while targeting multiple diverse HW+SW platforms. TF2AIF builds upon disparate toolflows to create a plethora of relative containers and enable the system orchestrator to deploy the requested function on any peculiar node in the cloud-edge continuum, i.e., to leverage the performance/energy benefits of the underlying HW upon any circumstances. TF2AIF fills an identified gap in today’s ecosystem and facilitates research on resource management or automated operations, by demanding minimal time or expertise from users.
When and How to Update AI/ML Models in 6G Resource-Constrained Network Domains?
Flavio Brito (Ericsson AB, Sweden); Josue Cisneros (Ericsson, Sweden); Zere Ghebretensae (Ericcson, Sweden); Neiva Linder (Ericsson Research, Sweden); Per Ödling (Lund University, Sweden)
Artificial intelligence and Machine Learning (AI/ML) are steadily becoming widespread across all layers in the current mobile network generation. The next generation of networks, namely 6G, will consider AI/ML as a foundational block to deliver all expected results of such networks. Thus, the management of AI/ML applications across different layers of 6G is paramount for the success of the next generation of mobile networks. However, current approaches to the management of AI/ML, namely Machine Learning Operations (MLOps), focus mostly on independent AI/ML operations for a specific problem with the goal of improving the performance of the overall models of the AI/ML application. While this approach is useful for scenarios where resources are plentiful, such as the cloud layer, in resource-constrained network domains, focusing only on performance is not the best approach. For example, in the network edge domain, resources such as energy and computation are limited; thus, when and how to update an AI/ML model is a critical question to answer. In this paper, we focus on MLOps in scenarios with limited resources. To this end, we propose a network component for 6G, namely the Model Manager. This component automatically decides when and how to update a given AI/ML model based on both performance and resource consumption points of view. For this, we introduce a Model Manager model score to decide the approach to updating an AI/ML model. Our experiments show that by using this score, the model manager could find suitable situations on how to update a model without manual configuration.
Implementation of a Traffic Flow Path Verification System in a Data Network
Javier Velazquez Martinez (Telefonica Innovacion Digital, Spain); Mattin Antartiko Elorza Forcada (Telefonica Innovación Digital, Spain); Antonio Pastor (Telefonica I+D & Universidad Politécnica de Madrid, Spain); Diego Lopez (Telefonica I+D, Spain); Jesús A. Alonso-López (Universidad Complutense of Madrid, Spain)
This paper focuses on one of the recent concerns that has arisen regarding the network softwarization, specifically, traffic attestation in service chaining. The central focus of the paper is the design, development, and evaluation of an implementation of Ordered Proof of Transit (OPoT) as a solution to validate flow paths in the network. This solution uses Shamir’s Secret Sharing (SSS) system to add metadata to each packet, updating them at each node or service it traverses until reaching the final destination. This method ensures the validation of services traversed by the packet at the last crossing point, providing an additional layer of security and preventing unauthorized modifications to the flow of data traffic. We report here how a programmable data plane, based on the P4 language, can be used to provide OPoT features dynamically, according to user and network policy requirements. Additionally, a controller will be developed to configure the network nodes, execute OPoT, and monitor the system state.
