Computational storage drive
A computational storage drive combines persistent data storage with compute engines that execute supported storage-associated functions.
Computational storage is the broader architecture, not a synonym for a drive. SNIA distinguishes a computational storage drive, which includes persistent data storage, from a computational storage processor, which supplies computation for associated storage without providing persistent data storage. A computational storage array is a storage array containing one or more compute engines. [2]
Compute engines, such as CPUs or FPGAs, run functions including compression, erasure coding and encryption. A drive can still serve ordinary storage requests through its existing host interface. [2]
For I/O controllers supporting the NVMe Computational Programs Command Set revision 1.3, Execute Program is mandatory but Load Program is optional (section 2.7.1). Device-defined programs are supplied by the namespace and cannot be changed by the host; downloadable programs require a supported type. Inspect the Program List and Downloadable Program Types List for that compute namespace, not merely the NVMe label. [3]
Loading, activation and execution are distinct operations. Execution requires the program to be activated on the target compute namespace. The Program List reports activation status when its query was processed; that status may have changed by the time the host uses it (sections 2.1.4.2 and 4.1.4.1). Downloading does not necessarily reserve compute resources; activation may reserve them. Host-loaded programs and activation state are cleared on power-on resets (sections 4.2.1 and 4.2.3). Flush has no effect on compute namespaces (section 2.2), so it cannot make this program state survive that reset. This program-state rule does not describe the persistence of every output dataset. [3]
Relations
- related
- NVMe, Enterprise SSD, IceClave, Willow, Flash-Cosmos, ParaBit
Advantages
- Implementations supporting the NVMe Computational Programs Command Set can expose program operations through that standard interface. [3]
Limitations
- Engine environments differ, from operating systems and containers to eBPF and FPGA bitstreams, and how a processor associates with its storage is implementation specific, which limits portability. [2]
- Conformance means tracking several separately revised documents. [2, 3]
- A common command interface does not establish that an arbitrary host binary is accepted or portable between devices. [2, 3]
- The HotStorage 2019 virtual-object prototype uses a vendor-specific EXEC and temporary extent/length descriptions; it is not the standardized NVMe Execute Program interface. [4]
- FIEMAP-based block-device access while the filesystem is mounted is undefined in Linux 6.12; FIEMAP_FLAG_SYNC does not waive that restriction, and encoded extents need separate handling. [5]
- Programmability does not define a threat model: Willow assumes an uncompromised host kernel, while IceClave considers hostile platform operators. [6, 7]
Citations
- SNIA Computational Storage Architecture and Programming Model, Version 1.2, 21 April 2026.
- NVM Express Computational Programs Command Set 1.3 (31 July 2026); copyright 2008-2026 NVM Express, Inc., all rights reserved.
- Adams, Keys and Mesnier, Respecting the block interface - computational storage using virtual objects, HotStorage 2019, sections 3.1-3.2.
- Linux 6.12, Fiemap Ioctl: Request Basics and Extent Mapping, including FIEMAP_EXTENT_ENCODED.
- Seshadri et al. Willow: A User-Programmable SSD. OSDI 2014, pp. 67-80, sections 2.6, 3, 4.2 and 4.6.
- Xue et al. Building A Trusted Execution Environment for In-Storage Computing. arXiv:2205.06361v1 (2022), section 2.
Further reading
Sources & caveats
- Architecture and command semantics reviewed against SNIA 1.2 and NVMe Computational Programs 1.3; these specifications are not performance measurements. No device experiment, speedup or output-durability result is claimed.
- The established computational-storage slug and broad alias are retained for compatibility; the definition follows the record's existing drive-specific title.
- The original HotStorage 2019 paper, sections 3.1-3.2, is a research design, not a generic Linux filesystem integration recipe. The Linux 6.12 FIEMAP contract is checked separately. No performance number or raw-block experiment is adopted. [4] [5]