---

Summary

CXL and related memory-pooling systems make it possible to access remote memory at cache-line granularity, but remote access latency and bandwidth costs make the right transfer granularity unclear. This paper compares conventional page-granular remote memory access with cache-line-granular access combined with aggressive prefetching.

The results show that cache-line-granular prefetching is a natural direction for remote memory devices, but matching or beating page-granular access is difficult for the workloads studied.

Bibtex

@inproceedings{mcmahon:prefetching,
  author = {James McMahon and Vinita Pawar and Ryan Stutsman},
  title = {{Remote Memory Prefetching: Is Coarse-grained Fine?}},
  booktitle = {Companion of the 16th ACM/SPEC International Conference on Performance Engineering},
  series = {ICPE '25},
  year = {2025},
  pages = {174--179},
  publisher = {ACM},
  url = {https://doi.org/10.1145/3680256.3721318},
  doi = {10.1145/3680256.3721318},
}

Generously Sponsored By

---