TPP: Transparent page placement for CXL-enabled tiered-memory

#CXL #memory_management #tiered_memory #Linux_kernel

Meta Info

Presented in ASPLOS 2023.

Authors: Hasan Al Maruf (UMich), Hao Wang (NVIDIA), Abhishek Dhanotia, Johannes Weiner (Meta), Niket Agarwal, Pallab Bhattacharya (NVIDIA), Chris Petersen (Meta), Mosharaf Chowdhury (UMich), Shobhit Kanaujia, Prakash Chauhan (Meta).

Code: https://lwn.net/Articles/876993/

Understanding the paper

  • TPP: Transparent Page Placement

  • It characterizes memory usage patterns of applications in Meta.

    • Show the opportunities to offload colder pages to slower memory tiers.

  • It proposes an OS-level application-transparent pace placement (TPP) for CXL-enabled memory.

    • It identifies and places hot/cold pages to appropriate memory tiers (i.e., local memory or CXL memory).

Implementation

A major portion of it has been merged to Linux kernel v5.18. The rest of it is under an upstream discussion.

Evaluation

  • Evaluated in the production server fleet with early samples of new x86 CPUs with CXL 1.1.

  • Improve application’s performance on default Linux by 18%.

  • Outperforms NUMA Balancing and AutoTiering, two state-of-the-art tiered-memory management mechanisms, by 5–17%.

Last updated