Transparent Memory Offloading (TMO)
Transparent Memory Offloading is a Linux-based system that uses pressure feedback to reclaim cold application memory into compressed memory or storage-backed swap.
Control loop: Senpai selects how much memory to reclaim using PSI; kernel reclaim selects the pages. The ASPLOS 2022 design considers anonymous memory, file cache and supporting services. It monitors I/O pressure as well as memory pressure, and regulates SSD writes. [1]
Evidence: The authors report 20-32% total-memory savings across Meta's fleet after more than a year in production. Their Web load tests used Skylake 64GB hosts and throughput at a target tail latency. These are deployment-specific results, not a transferable savings guarantee. [1]
Scope: The paper's backend selection between zswap and SSD swap was manual. CXL was future work. Its 1 MB/s write-regulation threshold reflects that fleet's SSD analysis, not a universal endurance budget. [1]
Request versus effect: memory.reclaim supplies an amount to attempt, without installing a persistent memory limit. Linux 6.12 explicitly allows over- or under-reclaim; EAGAIN can follow partial progress. Its write return value is not a reclaimed-byte counter. Use the linked interface record for the versioned contract. [2] [3]
Interpretation: A lower swap-in count alone is not evidence of a better result. Read it with application throughput, tail latency and pressure in the same workload interval. PSI reports time in stall states, not SSD wear or a latency percentile. [4]
Relations
Citations
- Weiner et al. TMO: Transparent Memory Offloading in Datacenters. ASPLOS 2022, pp. 609-621. DOI 10.1145/3503222.3507731; sections 3-5.
- Linux 6.12 cgroup v2: memory.reclaim and memory.high
- Linux v6.12 memory_reclaim: cumulative progress, error paths and write return
- Linux 6.12 Pressure Stall Information: some, full and observation scope
Sources & caveats
- Primary-source synthesis of the 2022 paper and Linux 6.12 documentation/code; no TMO deployment or workload benchmark was run by Hesela.
- Shipping describes the authors' documented production deployment, not a supported turnkey product or every current Linux distribution. Historical results and future-work statements retain their 2022 scope.