로고 로고

Ultra-Low-Latency Storage-Driven I/O Subsystem for Large Language Models

Secure global leadership in AI data center system software through storage-driven I/O technologies,
realize next-generation infrastructure capable of efficiently managing ultra-large-scale AI data on storage, and
cultivate world-class system software researchers

KEYWORDKeyword

System Co-design LLM Full-Stack AI Advanced Computing Reasoning

OBJECTIVE Objective

Centering on the three core technologies required for ultra-low-latency storage utilization in large language models—storage-driven I/O, a CPU-GPU integrated file system, and a CPU-GPU integrated I/O operating system— this project aims to achieve a phased reduction of host software overhead and establish the capability to manage petabyte-scale AI data. Through this, the project seeks to move beyond the conventional CPU-centric I/O paradigm and establish novel system software technologies in which storage actively participates in the data path, thereby building a foundation for technological self-reliance in AI data center system software. Furthermore, by reducing dependence on costly GPU memory and lowering AI infrastructure deployment costs, strengthening the competitiveness of the domestic semiconductor and storage industry through technology transfer and open-source contributions, and improving AI infrastructure accessibility in the public sector, the project aims to promote the broad adoption of AI technologies across industry and society.

INTRODUCTION Director's Message





At a time when AI is rapidly transforming industry and society, the importance of system software that underpins AI infrastructure has never been greater.
While much attention is directed toward algorithms and semiconductors, it is the operating systems and storage systems that make them work in practice that truly form the foundation of AI technology.

Over the past two decades of researching system software, I have witnessed firsthand how the role of the systems layer grows ever more critical at each paradigm shift.
This project is a challenge to secure, with our own hands, the core system software technologies essential to AI infrastructure at this very moment of transition

Above all, people are at the heart of this project.

The most important value we pursue is enabling talented researchers to develop independent research capabilities on KAIST's infrastructure and KISTI's state-of-the-art AI supercomputing infrastructure, and to grow into next- generation leaders in the field of system software.

Through close collaboration with our co-investigators and the KISTI research team, we are committed to producing meaningful outcomes.

Thank you.

IMPACT Impact

The research outcomes of this project can contribute to strengthening competitiveness in the AI data center software domain through technology transfer to the domestic semiconductor and storage industry.

By validating the practicality of the developed technologies through demonstrations on government research institute infrastructure, the project aims to improve the efficiency of public AI infrastructure.

Furthermore, open-source releases and patent filings will lay the groundwork for broader technology dissemination, while system-level techniques that reduce dependence on GPU memory are expected to lower the barriers to AI infrastructure deployment and promote wider adoption of AI across both industry and the public sector.

TALENT Ideal Talent

The postdoctoral researchers to be cultivated through this project are expected to possess a deep understanding of systems software, including operating systems and storage, and to be capable of independently carrying out kernel-level design, implementation, and verification.

We place great emphasis on the ability to analyze the limitations of existing software layers in complex system environments involving heterogeneous hardware and to propose new abstractions that overcome those limitations. We also seek individuals with the capacity to extend their domain expertise into AI infrastructure by incorporating workload-specific characteristics into system-level optimization.

In terms of values, we aim to nurture researchers who embody a pioneering spirit that challenges established paradigms rather than accepting them, an open and collaborative attitude toward joint research with government research institutes, and a long-term vision for growing into next-generation research leaders by building an independent track record through publications, patents, and open-source contributions.

FACULTY Faculty