Research
Making emerging memory useful across the systems stack.
My work connects memory hardware with the software that uses it: operating systems, communication libraries, and LLM inference runtimes. I am interested in how these layers can work together to improve performance on heterogeneous memory systems.
01 / MEMORY & OPERATING SYSTEMS
Managing tiered memory
Adding a memory tier creates new decisions about where data should live and when it should move. My work studies these decisions through real-device characterization, adaptive tuning, and programmable memory management.
- TierBPF uses eBPF to control admission to page migration with policies informed by the memory architecture.
- Hybrid Adaptive Tuning combines offline profiling and online learning to tune memory-tiering parameters.
- CXL characterization measures real CXL devices and studies their use in HPC, LLM workloads, and memory placement.
TierBPF · preprint / Adaptive tuning · preprint / Characterization · IPDPS 2025
02 / LLM SYSTEMS
Inference across CPUs, GPUs, and memory tiers
Long-context generation places pressure on both compute and memory. In HybridGen, we study CPU–GPU collaborative attention for LLM inference with expanded tiered memory. The work explores how to distribute computation across processors as the memory demands of inference grow.
03 / PARALLEL & DISTRIBUTED SYSTEMS
Communication and computation with shared memory
CXL memory sharing offers another way for machines to exchange data. cMPI explores MPI one-sided and two-sided communication over shared CXL memory. I carried out this research during an internship at SK hynix.
I also collaborate on OCEAN, a CXL emulation framework for studying multi-host memory sharing and pooling, and mLR, which uses memoization and memory offloading for scalable laminography reconstruction.
The full publication list includes papers, preprints, and available code. For research conversations, reach me at swang166@ucmerced.edu.