CAREER: Enabling Secure and Efficient AI Infrastructure with CXL-Enabled Disaggregated Shared Memory

Large machine learning models are advancing healthcare, scientific discovery, economic innovation, and national security. As these models increase in scale and capability, they require unprecedented memory capacity, creating a critical bottleneck for future Artificial Intelligence (AI) infrastructure. Emerging technologies such as Compute Express Link (CXL) enable large-scale memory resources to be shared across multiple servers, expanding capacity beyond the limits of a single server. However, sharing memory among multiple users introduces significant challenges to data privacy and system security. Without effective safeguards, cloud providers and data centers may be unable to safely deploy next-generation AI infrastructures based on CXL-enabled shared memory. This project addresses this challenge by developing secure and scalable abstractions and architectural designs for AI infrastructure using CXL-enabled shared memory, enabling continued technological advancement while strengthening economic competitiveness and national security. Research findings and simulation artifacts will be disseminated through publications, professional meetings, and a public website, and will be integrated into undergraduate and graduate curricula as well as K–12 outreach activities to advance education and workforce development in emerging computing technologies. This project establishes a secure and efficient AI infrastructure based on CXL-enabled disaggregated shared memory by addressing two research questions: how to secure disaggregated shared memory in multi-host and multi-tenant environments, and how to reduce security overhead to fully realize system performance. The project develops a new enclave abstraction tailored to disaggregated AI infrastructures that defines protected memory regions spanning multiple hosts and enables secure multi-tenant data allocation and management with strong confidentiality and integrity guarantees. The research investigates challenges unique to disaggregated settings, including cross-host isolation, secure resource coordination, and protection mechanisms against both software and hardware attacks, extending enclave-based protection beyond traditional single-server deployments. To reduce security overhead, the project introduces performance–security co-design techniques that incorporate workload data access characteristics, particularly those of machine learning models, into memory protection strategies. These techniques include workload-aware page prefetching and secure page migration mechanisms that improve efficiency while preserving strong security guarantees. The project conducts rigorous architectural and system-level evaluation of the proposed hardware and software designs using large-scale machine learning and memory-intensive workloads to assess scalability, security strength, and performance trade-offs. The resulting framework provides broadly applicable foundations for secure, scalable, and high-performance AI infrastructure built on CXL-enabled disaggregated shared memory. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2543427 | Program: 01003031DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Yuanchao Xu | Institution: University of California-Santa Cruz, SANTA CRUZ, CA | Award Amount: $464,384 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543427 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543427.html