NetLLM: Adapting Large Language Models for Networking Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, Fangxin Wang. 2024. NetLLM: Adapting Large Language Models for Networking. In ACM SIGCOMM 2024 Conference (ACM SIGCOMM '24), August 4–8, 2024, Sydney, NSW, Australia. ACM, New York, NY, USA, 18 pages. https://doi.org/10.1145/3651890.3672268 This paper presents NetLLM, a framework that adapts large language models (LLMs) for networking tasks. Current deep learning (DL) approaches for networking require extensive engineering efforts and suffer from poor generalization. Motivated by the success of LLMs, this work explores using LLMs as foundation models for networking, which can achieve better performance and stronger generalization. NetLLM provides a coherent design to harness the powerful capabilities of LLMs with low effort to solve networking problems. Specifically, NetLLM enables the LLM to process multimodal data in networking and efficiently generate task-specific answers. It also drastically reduces the costs of fine-tuning the LLM to acquire domain knowledge for networking. Across three networking-related use cases - viewport prediction, adaptive bitrate streaming and cluster job scheduling, we showcase that the NetLLM-adapted LLM significantly outperforms state-of-the-art algorithms. The main challenges of adapting LLMs for networking include large input modality gap, inefficiency of answer generation, and high adaptation costs. NetLLM addresses these challenges by designing a multimodal encoder to process multimodal data, a networking head to generate answers efficiently, and a data-driven low-rank networking adaptation (DD-LRNA) scheme to reduce adaptation costs. The multimodal encoder automatically projects task inputs to the same feature space as language tokens, enabling the LLM to understand and utilize these inputs for task solving. The networking head enables direct answer generation for specific tasks, ensuring the validity of answers. The DD-LRNA scheme reduces the costs of fine-tuning LLMs by introducing low-rank matrices to learn domain knowledge. NetLLM has been implemented for three networking tasks: viewport prediction (VP), adaptive bitrate streaming (ABR), and cluster job scheduling (CJS). The results show that the NetLLM-adapted LLM significantly outperforms state-of-the-art algorithms in terms of performance metrics. The framework is efficient, reliable, and transferable, with the potential to be applied to other fields beyond networking.NetLLM: Adapting Large Language Models for Networking Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, Fangxin Wang. 2024. NetLLM: Adapting Large Language Models for Networking. In ACM SIGCOMM 2024 Conference (ACM SIGCOMM '24), August 4–8, 2024, Sydney, NSW, Australia. ACM, New York, NY, USA, 18 pages. https://doi.org/10.1145/3651890.3672268 This paper presents NetLLM, a framework that adapts large language models (LLMs) for networking tasks. Current deep learning (DL) approaches for networking require extensive engineering efforts and suffer from poor generalization. Motivated by the success of LLMs, this work explores using LLMs as foundation models for networking, which can achieve better performance and stronger generalization. NetLLM provides a coherent design to harness the powerful capabilities of LLMs with low effort to solve networking problems. Specifically, NetLLM enables the LLM to process multimodal data in networking and efficiently generate task-specific answers. It also drastically reduces the costs of fine-tuning the LLM to acquire domain knowledge for networking. Across three networking-related use cases - viewport prediction, adaptive bitrate streaming and cluster job scheduling, we showcase that the NetLLM-adapted LLM significantly outperforms state-of-the-art algorithms. The main challenges of adapting LLMs for networking include large input modality gap, inefficiency of answer generation, and high adaptation costs. NetLLM addresses these challenges by designing a multimodal encoder to process multimodal data, a networking head to generate answers efficiently, and a data-driven low-rank networking adaptation (DD-LRNA) scheme to reduce adaptation costs. The multimodal encoder automatically projects task inputs to the same feature space as language tokens, enabling the LLM to understand and utilize these inputs for task solving. The networking head enables direct answer generation for specific tasks, ensuring the validity of answers. The DD-LRNA scheme reduces the costs of fine-tuning LLMs by introducing low-rank matrices to learn domain knowledge. NetLLM has been implemented for three networking tasks: viewport prediction (VP), adaptive bitrate streaming (ABR), and cluster job scheduling (CJS). The results show that the NetLLM-adapted LLM significantly outperforms state-of-the-art algorithms in terms of performance metrics. The framework is efficient, reliable, and transferable, with the potential to be applied to other fields beyond networking.