Storage-class memory (SCM)
Storage-class memory (SCM) is non-volatile solid-state memory between DRAM and block storage in latency, often byte-addressable and used as a tier close to the CPU.
A 2008 review framed SCM as combining solid-state performance with the archival capability and low cost of disks, targeting 50 to 1,000 ns access, 100 MB/s and 10-year retention against 3 to 5 ms for disk. Among flash extensions, ferroelectric, magnetic, phase-change and resistive memories it named phase-change and solid-electrolyte memory as front-runners, all with open scaling, endurance or retention questions. [2]
A 2019 persistent-memory DIMM study reports 305 ns random loads against 81 ns for local DRAM with cold caches. Its 94 ns store result, versus 86 ns for DRAM, times a store followed by a cache flush and fence reaching the protected ADR domain, not completion at physical media. The DRAM comparison is a timing baseline, not evidence that ordinary DRAM retains data without power. [3]
One module reached 6.6 GB/s read and 2.3 GB/s write, six modules 39.4 and 13.9 GB/s. Memory mode used DRAM as a cache without persistence, while App Direct mode exposed load-store persistent memory. [3]
The 256-byte media granularity is not a fixed small-write penalty. XPBuffer can combine adjacent 64-byte requests. Uncombined small stores can require read-modify-write. Effective write ratio (EWR) is bytes issued by the integrated memory controller (iMC) divided by media bytes written, measured using hardware counters. It can exceed one through write combining. [5]
Linux DAX bypasses the page cache for suitable memory-like storage. It needs continuous CPU byte addressability and does not work on architectures with virtually mapped caches. A fault at a mapped location with a known or latent media error can deliver SIGBUS; this describes mapped access, not every storage error path. [4]
Relations
- related
- 3D XPoint, Persistent memory, NVDIMM-N, Byte-addressable storage, Solid-state media, Phase-change memory (storage), Memristor storage, Direct Access, Write amplification, Read-modify-write
Limitations
- DAX does not imply O_SYNC guarantees. It removes the kernel page-cache copy, not the need for a persistence protocol. For a supported MAP_SYNC mapping, appropriate CPU instructions can persist mapped updates without explicit fsync. [6]
- PMDK documents pmem_persist() as a durability operation, not an atomic transaction. A failure during the call can leave only part of a range persistent, and stores may reach persistence before the call. Applications that require several stores to recover as one update need a separate consistency protocol. [7]
- In the studied platform, XPBuffer is already inside the protected persistence domain. It can merge updates despite ordering requirements imposed by memory fences. A durability boundary therefore need not coincide with an individual physical-media transfer. [5]
Citations
- Overview of candidate device technologies for storage-class memory
- Izraelevitz et al. Basic Performance Measurements of the Intel Optane DC PMM, arXiv v3: executive summary and sections 2.1-2.3, 3.1-3.2.
- Direct Access for files, Linux kernel documentation
- Yang et al. An Empirical Guide to the Behavior and Use of Scalable Persistent Memory. FAST 2020, sections 2.1-2.3 and 5.1, pp. 170-171, 174-175.
- Linux man-pages: statx(2), STATX_ATTR_DAX and MAP_SYNC semantics
- PMDK libpmem(3), version 1: pmem_persist durability and non-transactional semantics
Further reading
Sources & caveats
- c2 is an IBM Research review (2008); its figures are targets, not results.
- c3 measures 256 GB Intel Optane DC PMMs on a dual-socket Cascade Lake engineering-sample platform (Table 1), not general SCM properties.
- No prices, market size or product status are given.
- Arithmetic: 305 / 81 = 3.8; 3 ms / 1,000 ns = 3,000 and 5 ms / 50 ns = 100,000.
- c3 v3's abstract gives 346 ns, while its executive summary and section 3.1.1 give 305 ns. The note follows the latter locations; it does not resolve the discrepancy or claim a new measurement.
- c3's store latency stops at the ADR persistence domain. Sustainable write bandwidth is a separate experiment; the reciprocal of this latency is not a measured device throughput.
- c5 section 5.1: single-thread random non-temporal stores had EWR 0.25 for 64-byte accesses and 0.98 for 256-byte accesses. These are reported Optane measurements, not rerun or universal SCM constants.