persistent-memory · limited

Persistent memory

Persistent memory is nonvolatile storage exposed through a memory-addressable interface, with durability governed by the platform's persistence contract.

Related Observatory article: Read on hesela.com

A CPU store can remain in an unprotected cache even when its destination is persistent memory, and another thread may observe it before it is durable. [1] ADR and eADR protect different parts of the data path. [2] [3]

DAX avoids a page-cache copy, not every CPU cache. A block driver supports it by translating sector numbers in 512-byte units to page frames, and file access is supported only when the filesystem block size equals the kernel page size. [4]

A library persist call flushes the range and then drains, so earlier flushes are not reordered after the drain. It may widen the range to meet platform alignment. Stores can also become persistent earlier through normal cache eviction, so recovery code must not assume that unflushed data is still absent. Byte addressability does not imply transactional atomicity, and a recovery protocol is still needed. [5]

Remote access adds another boundary: an RDMA write completion does not by itself prove remote persistence. The required protocol depends on whether protection covers memory-controller queues, CPU caches and NIC buffers, and whether DDIO directs writes into caches. A remote-visibility guarantee must not be substituted for that persistence contract. [7]

Relations

related
NVDIMM-N, Storage-class memory (SCM), Intel Optane, Byte-addressable storage, Persistence domain, Asynchronous DRAM Refresh, Extended Asynchronous DRAM Refresh, Direct Access, MAP_SYNC, Persistence ordering, Magnetoresistive RAM, Strict persistency, FPTree, Cacheline-Conscious Extendible Hashing, NOVA file system, BPFS, SplitFS, Strata, Assise, Aerie, Kuco, PACTree, Pangolin persistent object library, NOVA-Fortis, MadFS, NoveLSM, Viper, Clover, Hotpot, CXL memory sharing, Twizzler

Advantages

  • Mapping the storage directly into user space removes the extra page-cache copy for reads and writes. [4]

Limitations

  • A persist call is neither atomic nor transactional, so after a crash part of the range may hold new stores and part old ones. [5]

Citations

  1. PMDK project. Concurrency considerations in libpmemobj-cpp, 17 September 2021. (accessed 2026-10-13)
  2. Intel. Second Generation Xeon Processor Scalable Family Technical Overview: The Persistence Domain. (accessed 2026-10-13)
  3. Intel. Third Generation Xeon Processor Scalable Family on Two Socket Platforms: eADR. (accessed 2026-10-13)
  4. Linux kernel documentation. Direct Access for files: motivation and S_DAX state. (accessed 2026-10-23)
  5. PMDK 1.13.1. pmem2_get_persist_fn(3): early persistence and non-transactional semantics. (accessed 2026-10-23)
  6. Sanidhya Kashyap, Dai Qin, Steve Byan, Virendra J. Marathe and Sanketh Nalli. Correct, Fast Remote Persistence. arXiv:1909.02092v1, sections 2-3. (accessed 2026-11-15)

Further reading

  1. Hesela. Persistent memory: DAX is not a durable commit. (accessed 2026-10-13)
Sources & caveats
  • dax-docs re-read 2026-10-23 for the page-cache motivation, block-size rule and the 512-byte sector translation.
  • pmdk-persist re-read at tag 1.13.1 for the flush-then-drain sequence, range widening and early persistence.
  • A companion analysis on persistence domains is listed but not used in prose.
  • The execute-in-place entry links here because direct access to memory-addressable storage without a page-cache copy is shared with in-place execution.
  • 2026-11-15: added source-scoped remote visibility versus persistence qualification from Kashyap et al., sections 2-3; no hardware experiment.

JSON: /data/terms/persistent-memory.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. Persistent memory. Hesela. 2026-11-16. https://hesela.dev/terms/persistent-memory/