Large-Scale Field Study of SSD Operational Characteristics in Enterprise Storage Systems

The work, published at FAST 2022, reports the first large‑scale field study of SSD operational characteristics based on data from nearly 2 million SSDs deployed in enterprise storage systems of a major vendor (NetApp).

Background and Motivation – Understanding SSD behavior in production is crucial because their operational patterns differ from HDDs, affecting wear, write amplification (WAF), and overall durability.

Research Questions

How much data do SSDs write in production and how close are they to wear limits?

What is the observed write amplification and how does it compare to prior academic reports?

What are the wear‑leveling characteristics in the field?

How do factors such as FTL version, workload type, DWPD, RAID role, capacity, and interface impact WAF?

System Description

~2 million devices from three manufacturers, 20 series, covering various capacities and NAND types.

Data collected via NetApp Active IQ (SMART‑like telemetry) includes host read/write, total physical writes, wear level, and write amplification per drive.

Two system classes: WBC (SSD cache for HDD) and AFF (All‑Flash).

Operational Features

1. Write Rate (DWPD) – Average DWPD is 0.36, satisfying most drives; however 7 % exceed 3 and 2 % exceed 10. WBC drives show significantly higher write rates (average 3.6×, 99th percentile 10.6×) than AFF drives.

2. Write Amplification (WAF) – Overall WAF distribution differs markedly from prior studies; 98.8 % of SSDs have WAF > 1.3. Certain families (I‑C, I‑D, I‑E) exhibit WAF an order of magnitude higher, largely due to background retention‑mitigation work.

3. Wear Leveling – Median erase ratio is 1.55, indicating some blocks wear 55 % more than average; 5 % of SSDs have an erase ratio > 6, showing uneven wear.

4. Fullness – Average system fullness is ~45 % (median 45 %). Fullness grows faster in the first two years, then slows; larger‑capacity systems tend to be fuller.

Factors Impacting WAF

FTL version: drives with firmware FV3 show different WAF than FV2 for the same model.

Workload proxies: DWPD, RAID role, capacity, interface, and read/write ratio.

DWPD: Higher DWPD correlates with lower WAF, suggesting SSDs operate more efficiently under heavier write loads.

RAID role: Partitioned SSDs have higher WAF, likely due to mixed workload handling.

Capacity: Larger SSDs (8 TB, 15 TB) exhibit lower WAF than smaller ones.

Interface: NVMe drives show significantly lower WAF, possibly due to different workload placement.

Read/Write Ratios – Approximately 94 % of drives are read‑dominant with a median R/W ratio of 3.6 : 1 (95th percentile 61 : 1), contrasting with prior HDD‑centric studies where writes dominate.

Most Important Findings (Summary)

Most SSDs consume program‑erase (PE) cycles very slowly.

Cache SSDs experience higher host write intensity but lower WAF, indicating higher‑endurance drives are not always required for caching.

WAF varies by drive family and manufacturer by orders of magnitude.

Background work for retention mitigation is a major contributor to WAF.

Wear leveling is imperfect; 5 % of SSDs have erase ratios > 6.

AFF systems are on average 43 % full; larger systems are fuller.

Read‑dominant workloads (94 % of SSD‑related jobs) highlight a stark contrast with HDD‑centric environments.

Source: SCS Storage Committee; original paper: https://www.usenix.org/system/files/fast22-maneas.pdf