Understanding IOPS, Throughput and Storage Latency
Storage performance is characterized by three key metrics: IOPS, throughput, and latency. Understanding these metrics helps you select the right storage for your workloads and diagnose performance issues.
IOPS (Input/Output Operations Per Second)
IOPS measures how many individual read or write operations a storage device can perform per second. This metric matters most for workloads that perform many small, random I/O operations.
Databases are the primary consumer of IOPS. A PostgreSQL database performing random index lookups generates many small reads. Each read is an I/O operation.
Typical IOPS values: - HDD: 100 to 200 IOPS - SATA SSD: 90,000 to 100,000 IOPS - NVMe SSD: 500,000 to 1,000,000+ IOPS
Throughput
Throughput measures the rate of data transfer, typically in MB/s. This metric matters for workloads that read or write large contiguous blocks of data.
Sequential operations that benefit from high throughput: - Database backups and restores - Large file transfers - Log file processing - Media streaming
Typical throughput values: - HDD: 100 to 200 MB/s - SATA SSD: 500 to 550 MB/s - NVMe SSD: 3,000 to 7,000 MB/s
Latency
Latency is the time between issuing an I/O request and receiving the response, measured in microseconds or milliseconds.
Typical latency values: - HDD: 5,000 to 10,000 microseconds (5 to 10 ms) - SATA SSD: 50 to 100 microseconds - NVMe SSD: 10 to 20 microseconds
Random vs Sequential
Random I/O (database queries, virtual machine disk access) is characterized by many small operations at random locations. IOPS and latency are the relevant metrics.
Sequential I/O (file transfers, backups, log writes) involves reading or writing large contiguous blocks. Throughput is the relevant metric.
Choosing Storage for Your Workload
Database server: prioritize IOPS and latency. NVMe storage is strongly recommended.
Web server: SATA SSD is usually sufficient. Random I/O is moderate.
Backup target: throughput matters more than IOPS. Standard SSD or even HDD may be adequate.
File server: depends on access patterns. Many concurrent small reads benefit from IOPS; large file transfers benefit from throughput.
Measuring Storage Performance
Use fio (Flexible I/O Tester) to benchmark storage: fio --name=test --ioengine=libaio --direct=1 --bs=4k --rw=randread --size=1G --numjobs=4 --runtime=30
This measures random read IOPS with 4 KB block size.
Cloud Storage Considerations
Cloud VPS storage performance varies by provider and storage type. When evaluating cloud infrastructure, ask about IOPS limits, throughput caps, and latency guarantees for the storage options available.
Key Takeaways
- IOPS measures random I/O capacity and matters most for databases; throughput measures sequential transfer rates and matters for backups and large file operations
- Latency is the delay per I/O operation — NVMe SSDs deliver 10–20 microseconds versus 5–10 milliseconds for HDDs, a difference that compounds under sustained load
- Match storage to workload: prioritize IOPS and latency for databases, throughput for backup targets, and SATA SSDs for general-purpose web servers
- Cloud storage performance varies significantly by provider and tier — always verify IOPS limits, throughput caps, and latency guarantees before deploying
- Use tools like `fio` to benchmark real-world storage performance rather than relying solely on vendor specifications