2. Performance Fundamentals

Outcome

Calculate basic workload relationships and identify whether additional concurrency exposes service capacity or only adds waiting time.

Level: Foundation · Time: 40 minutes · Prerequisite: Storage Building Blocks

Learn

throughput ≈ IOPS × average transfer size
concurrency ≈ IOPS × average latency     (Little's Law, consistent units)
response time = service time + queueing time

IOPS without transfer size is incomplete. Throughput without latency distribution is incomplete. Average latency can hide a damaging tail, so operational reviews should include percentiles and workload shape.

Sequential access favors large transfers and locality. Random access increases seek, mapping, metadata or flash-translation pressure depending on the medium. Reads and writes may have different cache and protection paths.

Practice

Open the NVMe Queue Lab. Hold the device ceiling constant, increase queue depth gradually and note the point where throughput stops improving while latency continues to rise.

Check your understanding

1. What throughput does 25,000 IOPS × 16 KiB approximately produce?
2. Why can higher queue depth increase benchmark throughput but harm an application SLO?
3. Which evidence would distinguish device saturation from fabric congestion?

Production boundary

The lab is a queueing baseline. Real latency includes application, operating system, multipath, protocol, fabric, controller, cache and media behavior.

Next step

Continue to Data Protection and Recovery.