How to Process SSD Benchmark Logs and Analyze Performance Data

Introduction

The series on SSD benchmarking covers four topics; this final part focuses on data processing. After collecting raw fio logs, the goal is to transform them into a usable format for analysis and visualization.

Log Formats

Fio generates separate logs for bandwidth (bw), IOPS, and latency (clat, slat, lat). Each line follows a simple CSV‑like structure: a timestamp, metric values, and a block size.

<code># fio bandwidth log
0,21005,0,4096
0,20378,0,4096
0,21222,0,4096
0,22755,0,4096
# fio iops log
0,1,0,4096
0,1,0,4096
# fio latency log
0,453,0,4096
0,435,0,4096
</code>

The prefix (e.g., log_read_fio_4k_qd16_bw ) identifies the test, while the suffix ( bw , iops , clat , etc.) indicates the metric.

Metric Explanation

bw stands for bandwidth (KB/s). clat is the completion latency, the time from the device receiving an I/O request to finishing it. slat is the submission latency, the time the OS takes to submit the request; it is usually negligible and can be ignored. lat equals clat + slat , but because slat is tiny, lat and clat are almost identical. For SSD benchmarking the tests often disable slat and clat to focus on device performance.

Data Processing

The raw logs can be stripped of headers, converted to CSV, and aggregated (e.g., min, max, average) as needed. Bandwidth values are derived from block size / clat , so the bw log adds little value when clat and iops are already available. After processing, the data is ready for plotting.

Results and Discussion

Two enterprise SSDs (SanDisk and Micron) were tested with 4 KB random reads. The latency charts show sub‑millisecond reads for both drives, but SanDisk consistently has lower latency. Micron’s latency varies more, though still within acceptable limits. IOPS charts reveal similar trends: SanDisk’s IOPS remain stable, while Micron shows slight fluctuations.

Key observations:

Maximum latency stays below 2 ms for both drives, indicating acceptable service quality.

Latency spikes can become a service‑quality killer if they exceed tolerance (e.g., >200 ms).

Random write tests (3 reads / 7 writes) show similar patterns; write amplification is a concern for SSD lifespan.

Recommendations

During testing, avoid zero‑IOPS or zero‑throughput points.

Keep read/write latency within predictable bounds (generally <5 ms for high‑load scenarios).

Do not rely on TRIM alone to maintain SSD performance.

Consider PCIe/NVMe SSDs for production workloads, but verify firmware and RAID‑card interactions.

Overall, thorough log processing and metric analysis are essential for selecting SSDs that meet business performance and reliability requirements.

Illustrative Charts