Analyzing Performance Factors of NVMe SSDs and Optimizing Storage Systems

NVMe SSD performance can be unpredictable, so it is necessary to open the "SSD black box" and analyze influencing factors from multiple perspectives, then consider how storage software can best utilize NVMe SSDs to accelerate flash adoption in data centers.

1. Evolution of storage media – Semiconductor storage (NVMe SSD) has replaced magnetic disks, offering superior reliability, performance, and power efficiency. NVMe SSDs use PCIe interfaces and NAND Flash, with a Flash Translation Layer (FTL) that abstracts NAND quirks and presents a block‑device interface.

2. NVMe SSD becomes mainstream

2.1 NAND Flash development – 3D stacking and higher‑bit cells (TLC, QLC) increase density; capacities now reach 128 TB per 3.5‑inch drive.

2.2 Multi‑queue technology – NVMe introduces multiple submission/completion queues, allowing each CPU core to communicate with the SSD via independent hardware queue pairs, eliminating the single‑queue bottleneck of AHCI.

2.3 SSD hardware details – NAND Flash cells, multi‑plane architecture, ECC, LDPC, and controller designs (SMP vs MPP) all impact performance and reliability.

3. Factors influencing NVMe SSD performance

Hardware factors – NAND type, channel count, controller processing power, controller architecture, DRAM capacity for mapping tables, PCIe bandwidth, temperature, and wear level.

Software factors – Data layout, garbage‑collection (GC) and wear‑leveling scheduling, over‑provisioning, error‑correction handling, FTL algorithms, I/O scheduling, driver design (kernel vs user‑space), and I/O patterns.

Objective factors – Age‑related wear, temperature‑induced throttling, and long‑term power‑off effects.

3.1 Impact of GC – GC generates background traffic that reduces foreground I/O performance; SSDs show large performance gaps between empty‑drive and steady‑state conditions.

3.2 Impact of I/O pattern – Sequential large‑block writes minimize write amplification and background traffic, while random small writes increase GC overhead; read‑write interference further degrades performance, mitigated by sophisticated I/O schedulers and techniques such as program/erase suspension.

4. SSD write‑performance model – Derives the relationship between user write bandwidth (U), total PCIe bandwidth (B), and write‑amplification factor (WA): U = B / (2·WA – 1). The model matches measured spec values for devices like Intel P4500.

5. Conclusion – Flash storage technology is rapidly advancing; NVMe SSD performance is affected by many hardware and software factors, and optimizing I/O patterns at the software level can significantly improve overall storage‑system performance.