The Connected Intelligence unit at RISE Research Institutes of Sweden accesses EuroHPC JU supercomputer MeluXina to work with AI-Powered Network Optimization. They applied using the Development Access Call which is continuously open all year round.

Organization Involved 

RISE Research Institutes of Sweden is a state-owned research institute with over 3000 employees and offices located throughout Sweden. The organization is dedicated to fostering innovation and advancing knowledge across various scientific and technological fields, to foster research and innovation to serve Swedish industrial and societal sectors towards strengthening Sweden’s global competitiveness, while aligning with sustainability objectives.

Specializing in areas such as networking, artificial intelligence, machine learning, cybersecurity, distributed systems, software engineering, and data analytics, the Computer Science department at RISE collaborates with industry partners, academic institutions, and government agencies to address complex challenges and drive technological advancements.  

The project grant was awarded to members of the Connected Intelligence unit at RISE, in the Computer Science Department. For information about the project, please contact Daniel Perez.

Technical/Scientific Challenge 

As future communication and computer networks grow in scale, optimizing resource allocation becomes increasingly vital to ensure energy efficiency and sustainability. Modern networks face NP-hard challenges, such as balancing energy, spectrum utilization, and latency while meeting application demands.

A prominent use case is scheduling in IoT networks augmented with backscatter communication. These networks rely on battery-free sensor tags, which require carefully planned schedules for carrier provisioning from neighboring devices. Solving this scheduling problem at scale, while maintaining efficiency and adaptability, is a critical challenge for sustainable digital infrastructure. 

Proposed Solution 

We developed an AI-driven scheduler to address resource allocation challenges in IoT networks. Using Deep Learning, Graph Neural Networks, and Transformers, the scheduler generates efficient carrier schedules by learning from optimal solutions in small IoT networks and generalizing to networks up to 100× larger without retraining. With the Meluxina HPC cluster, we scaled training pipelines for both simulated and real-world network topologies.

The HPC resources allowed us to train over 50 ML models and identify the hyperparameters and embedding configurations that allowed our model to robustly generalize to such large networks. Optimizing workflows for GPU acceleration on NVIDIA A100 GPUs, the scheduler achieves polynomial runtime complexity, enabling it to dynamically adapt to changing network conditions. This solution not only scales effectively but also significantly reduces energy and spectrum utilization. 

Business Impact 

Access to MeluXina supercomputer has been instrumental in accelerating research and innovation in energy-efficient networking. The scheduler achieves up to 2× reductions in energy and spectrum usage, offering a scalable solution for large-scale IoT networks. These improvements support sustainable digitalization while positioning our organization as a leader in AI-powered optimization for network resource allocation.

Moreover, the planned open-source release of the developed code, models, and datasets ensures that our work benefits the broader research community and industrial stakeholders, fostering real-world applications in diverse sectors like smart cities, healthcare, and industrial IoT. 

Benefits 

• Time savings: Substantial reduction in training and experimentation cycles using HPC-enabled parallel training, allowed us to widen the hyperparameter exploration space. 
• Cost savings: our developed solution achieves up to 2× better resource utilization in the IoT networks studied. 
• Product optimization: Scalable scheduling solutions adaptable to dynamic network conditions. 
• Sustainability: Enhanced energy efficiency for future digital infrastructures in the context of 6G networks

Do you need more computational power for your projects?

Learn more about how you can access supercomputers for your projects. ENCCS can help! Visit https://enccs.se/get-supercomputer-access/