IOPS
IOPS (Input/Output Operations Per Second) is the primary metric for measuring the throughput capacity of a storage system for random-access workloads. NVMe SSDs deliver 500K–7M+ random IOPS; NVMe-oF over TCP achieves similar figures at 25–40µs latency across the network.
IOPS and NVMe Storage
IOPS is most meaningful for random I/O workloads — databases, VMs, Kubernetes volumes — where data is accessed at non-sequential locations. Legacy HDD delivers ~150 IOPS; SATA SSD ~80K; NVMe SSD 500K–7M. A single NVMe-oF storage node can serve millions of IOPS to multiple hosts simultaneously.
IOPS, Queue Depth, and Latency
IOPS = Queue Depth / Latency. To maximize IOPS from an NVMe device, the host must maintain a deep enough queue to keep the device busy. NVMe's 64,000 queue pairs enable far higher parallelism than SCSI's 1 queue × 128 commands, which is why NVMe achieves dramatically higher IOPS than iSCSI on the same physical media.
Measuring IOPS
The standard tool is fio (Flexible I/O Tester). A typical 4K random read IOPS benchmark:
--rw=randread --bs=4k --direct=1 --size=4G \
--filename=/dev/nvme0n1 --runtime=60