Understanding Disk I/O Performance, IOPS, and RAID Calculations

Disk I/O refers to read and write operations where the controller sends commands to the disk, which then returns data to the operating system; a single I/O operation completes either a read or a write.

Random access occurs when successive I/O requests target sectors far apart, requiring large head movements, while sequential access targets nearby sectors, allowing faster consecutive operations.

IO size (IO Chunk Size) varies; operating system and controller caches may combine multiple requests, so the size is not fixed and can range from small (1‑8 KB) to large (32‑64 KB) depending on workload.

IOPS (IO operations per second) measures how many I/O operations a system can perform. For a single disk, IOPS can be estimated using the formula IOPS = 1000 / (Tseek + Trotation) , where Tseek is average seek time and Trotation is average rotational delay.

RAID configurations affect IOPS. Read IOPS for a RAID array can be approximated as read IOPS = disk_IOPS / (1 - disk_buffer_read_hit_ratio) * disk_num . Write IOPS includes a penalty factor depending on RAID level, e.g., RAID 0 has no penalty, RAID 1/10 has a factor of 2, RAID 5 a factor of 4, and RAID 6 a factor of 6, using the formula write IOPS = disk_IOPS / (1 - disk_buffer_write_hit_ratio) * disk_num / penalty .

Real‑world performance often exceeds theoretical calculations due to caches, advanced features (TCQ/NCQ), and SSD characteristics such as SLC, MLC, and eSLC hierarchies.

When evaluating storage for desktop virtualization, consider both total system IOPS demand (e.g., number of VMs × per‑VM IOPS) and per‑LUN IOPS supply, adjusting for RAID level and cache‑induced performance factors.

Transfer rate (throughput) can be derived from IOPS and I/O size ( Transfer Rate = IOPS * IO Size ), but for large sequential workloads, raw transfer speed of the disk becomes the limiting factor rather than IOPS.

IO response time (latency) includes queue waiting time in addition to the disk's internal processing time; as IOPS utilization approaches 70 % of the maximum, latency grows non‑linearly, indicating the need for capacity planning.