Using compression to enhance bandwidth on low-performance networks

• Autores: Cristian Peñaranda Cebrián
• Directores de la Tesis: Carlos Reaño González (dir. tes.), Federico Silla Jiménez (dir. tes.)
• Lectura: En la Universitat Politècnica de València ( España ) en 2025
• Idioma: inglés
• Tribunal Calificador de la Tesis: Margarita Amor López (presid.), José Salvador Oliver Gil (secret.), Horacio Gonzalez Velez (voc.)
• Programa de doctorado: Programa de Doctorado en Informática por la Universitat Politècnica de València
• Materias:
  • Ciencias tecnológicas
    • Tecnología de los ordenadores
    • Tecnología de las telecomunicaciones
• Enlaces
  • Tesis en acceso abierto en: RiuNet
• Resumen

  • The Internet of Things (IoT) has experienced significant growth in recent years, driving the need to process data on devices with limited resources efficiently. However, these devices typically have reduced computing capabilities, making it challenging to run demanding applications such as those based on artificial intelligence. A common strategy is to offload specific tasks to the edge, but the computational power of these devices remains insufficient for more complex workloads.

    Remote GPU virtualization emerges as a viable solution to address these computational limitations, enabling IoT devices to leverage GPUs located on remote servers to accelerate application processing. In this way, devices can efficiently execute GPU-intensive tasks by offloading the computation to a more powerful machine. This approach significantly extends the computing capabilities of IoT systems.

    However, remote GPU virtualization introduces a new challenge: network dependency. For the system to function efficiently, data must be quickly transferred between the device and the remote GPU. Data transmission can drastically affect application performance in networks with limited bandwidth, such as those commonly found in IoT environments. Therefore, reducing the amount of transmitted data and optimizing communication between the client and the server becomes a key factor.

    In this thesis, several solutions have been developed to mitigate this problem through data compression within the communication layer of remote GPU virtualization. Different compression techniques have been explored, including a pipeline-based compression system to improve data transmission, and abstraction libraries that group many compression libraries: 41 CPU-based and 8 GPU-based. In addition, a specific dataset was created to evaluate the compression performance in this context. Finally, tests were carried out using AI applications on low-performance networks to analyze the impact of these solutions and propose future improvements.