Name of participant: Zahra Ebrahimi
Project’s name: X-DNet: Energy-Efficient Distributed and In-Network Computing via Approximation of Applications and Accelerators
Project description:
Developing sustainable, energy-efficient, and real-time solutions is the main challenge in 5G/6G era, especially considering the ever-increasing complexity of computing algorithms in stream processing and AI applications. Currently, executing medium- to large-scale applications is only possible on the cloud due to the resource constraints on edge devices. However, transmitting large data and offloading the computations to the cloud comes with huge penalties. First, only transferring data takes 50-90% of the total energy in, e.g., AI inference, and this has been a major barrier to Germany’s Energiewende (energy transition) and Nachhaltigkeit (sustainability) goals. Second, such transmission delays a real-time response and increases the risk of network congestion. Third, transmitting raw data to the cloud jeopardizes data privacy for users. Hence, relying solely on cloud-based services can no longer be viable. To address these concerns in future 6G networks, the computing is envisioned to become distributed and/or performed on-the-fly, while data is transmitted through the network elements (switches, base stations, SmartNICs, etc), called In-Network Computing (INC). However, the concept of INC is still not easily applicable to medium-to-large complex applications because they are resource-hungry and, therefore, cannot be deployed in resource-constrained network switches with limited processing capability. To address this challenge, the X-DNet project aims to enable a real-time and energy-efficient in-network acceleration for large and multi-kernel applications through cross-layer approximation, optimization, and distribution techniques. To this end, we first modify and optimize the structure of complex applications and reduce their complexity to make them deployable on resource-constraint 5G-6G network nodes. This is achieved through using various optimization and approximation techniques. Afterward, we partition and distribute the application kernels on a chain of network nodes, not only w.r.t the available resources but also in a way to minimize the transmission costs.
Software Campus Partner: TU Dresden und Huawei Technologies Düsseldorf GmbH
Implementation period: 15.05.2023 – 31.01.2025
