Understanding NVMe IO Determinism in the NVMe 1.4 Specification

The NVMe Association presented the NVMe 1.4 roadmap at FMS2018, highlighting three major updates: IO Determinism, Persistent Memory Region, and Multipathing, with IO Determinism being the primary focus.

IO Determinism aims to provide relatively independent access spaces, increase IOPS, reduce latency, and deliver superior QoS.

It consists of three parts: NVM Sets, Predictable Latency Mode (PLM), and Read Recovery Level (RRL).

NVM Sets partition SSD physical resources into independent sets, each containing multiple flash channels and dies, allowing applications to operate without blocking each other, thereby improving overall system QoS.

PLM divides operation time into Deterministic Windows (DTWIN) and Non‑Deterministic Windows (NDWIN). During DTWIN the SSD guarantees deterministic latency for read/write commands, while NDWIN is used for background tasks such as garbage collection. Hosts can obtain typical and reliable‑estimate values via the Get Log Page command and can force the SSD into a specific window using Set Features.

The relationship between typical and estimate values is illustrated by a read‑DTWIN graph; when the estimate value falls below a threshold, the SSD automatically switches to NDWIN.

Key timing parameters include NDWIN Time Minimum Low, NDWIN Time Minimum High, and DTWIN Time Maximum, which together define the proportion of time the SSD spends in deterministic versus non‑deterministic modes.

RRL addresses varying read error recovery needs. Different recovery levels (0–15) trade off between data recovery effort and read latency, allowing hosts to select the appropriate level for their workload.

As SSD performance peaks, customers now prioritize stable performance and low latency; NVMe IO Determinism is positioned to meet these requirements, and its adoption is expected to grow with the final release of NVMe 1.4.