IOPS Explained
What Are IOPS?
IOPS stands for Input/Output Operations Per Second. It measures how many individual read or write operations a storage device or storage system can complete during one second under a particular workload.
Higher IOPS can indicate greater ability to handle large numbers of storage requests, but an IOPS number has little meaning without context. Block size, read/write mix, random versus sequential access, queue depth, latency and workload conditions all affect the result.
What Does IOPS Actually Measure?
An I/O operation is a request to read data from storage or write data to storage. IOPS measures the rate at which those operations are completed.
Read Operations
Measures how many read requests the storage system can complete per second under the specified test conditions.
Write Operations
Measures how many write requests the storage system can complete per second under the specified workload.
Mixed Read/Write Workloads
Represents workloads containing both reads and writes, such as 70% reads and 30% writes. Mixed workloads often better represent real applications.
IOPS vs Throughput vs Latency
IOPS, throughput and latency describe different aspects of storage performance. A complete storage evaluation should consider all three rather than treating any single measurement as the entire performance story.
IOPS & Throughput
For a workload using a consistent block size, IOPS and throughput are mathematically related.
For example, 10,000 IOPS using 4 KB operations represents approximately 40,000 KB of data transferred per second before unit conversions and other real-world considerations.
4K IOPS Is Not the Same as 128K IOPS
An IOPS figure must be interpreted together with the I/O size. Thousands of small 4 KB operations represent a very different workload from thousands of 128 KB operations.
Small-block random I/O is commonly important for databases and virtual machines, while larger sequential operations are often more relevant to streaming, backup and large-file workloads.
Random IOPS vs Sequential I/O
Data From Many Locations
Random I/O accesses data from many different locations. This pattern is common in databases, virtualization, transactional applications and multi-user server environments.
Data in Continuous Regions
Sequential workloads access larger contiguous regions of data. Backup, media streaming and large-file transfers commonly place greater emphasis on sustained throughput.
HDD IOPS vs SSD IOPS
Mechanical Access Limits Random I/O
HDDs must physically move their read/write heads and wait for rotating media to reach the requested location. Mechanical seek time and rotational latency therefore limit random IOPS.
Faster rotational speeds and optimized workloads can improve performance, but HDD random I/O remains fundamentally constrained by mechanical movement.
Flash Enables Much Higher Random IOPS
SSDs have no mechanical seek operation. This allows solid-state storage to process random requests much more quickly and enables far greater IOPS potential than mechanical drives.
Actual SSD performance depends on the controller, NAND, firmware, interface, queue depth, workload and whether the operations are reads or writes.
SATA, SAS & NVMe IOPS
The storage interface can influence the maximum performance a device or array can deliver, but interface type alone does not determine IOPS. Media type, drive architecture and workload characteristics remain critical.
Why Queue Depth Changes IOPS Results
Queue depth describes how many I/O requests are outstanding at a given time. A queue depth of 1 represents a very different workload from a queue depth of 32, 64 or greater.
Many storage devices can achieve higher aggregate IOPS when more requests are available to process concurrently. However, increasing queue depth can also increase request latency. For this reason, comparing two drives using IOPS figures measured at different queue depths can be misleading.
How Much Work Gets Done?
IOPS measures the number of storage operations completed over time. It is particularly useful when evaluating workloads that generate large numbers of relatively small I/O requests.
How Long Does Each Request Take?
Latency measures the time required to service an I/O request. A storage system can report impressive aggregate IOPS while still delivering latency that is unsuitable for a latency-sensitive application.
How RAID Affects IOPS
Combining drives into a RAID array can increase aggregate storage performance because multiple devices can participate in I/O. However, RAID level, controller architecture, workload and the proportion of reads to writes all affect the result.
Mirrored and parity-based RAID levels handle writes differently. RAID 5 and RAID 6, for example, require parity-related operations that can increase the work associated with small random writes. This is commonly described as the RAID write penalty.
Can You Add Drive IOPS Together?
As a simplified planning concept, multiple drives can provide greater aggregate I/O capability than a single drive because operations can be distributed across devices. However, simply multiplying one drive's IOPS by the number of drives does not guarantee the application's real-world IOPS.
RAID overhead, write penalties, caching, controller limits, PCIe or SAS/SATA bandwidth, filesystem behavior, operating system I/O, queue depth and application behavior can all change the result.
When Do IOPS Matter Most?
| Workload | Common I/O Characteristic | Important Metrics |
|---|---|---|
| Databases | Frequent small random reads and writes | IOPS + latency |
| Virtualization | Mixed random I/O from multiple virtual machines | IOPS + latency + queue behavior |
| Transaction Processing | Small, frequent storage operations | IOPS + latency |
| Backup | Often large sequential transfers | Throughput + capacity |
| Media / Large Files | Large sequential reads or writes | Throughput |
| AI / Analytics | Workload-dependent parallel access to datasets | IOPS + throughput + latency |
Never Compare IOPS Numbers Without the Test Conditions
A statement that one drive delivers more IOPS than another is not enough to establish that it will be faster for your application. Benchmark conditions must be comparable.
How Many IOPS Do You Need?
There is no universal number of IOPS that makes a storage system "fast." The correct target depends on the application's actual I/O demand and acceptable latency.
Storage sizing should ideally begin with measurements from the existing workload. Determine peak and sustained IOPS, read/write ratio, block sizes, latency requirements, throughput, capacity growth and workload concurrency before selecting drives and RAID architecture.
What to Evaluate Before Choosing Storage
Related Storage Guides
IOPS FAQ
What does IOPS stand for?
IOPS stands for Input/Output Operations Per Second. It measures how many storage read or write operations can be completed per second under specified conditions.
Is higher IOPS always better?
Not necessarily. Higher IOPS can be useful, but latency, throughput and workload characteristics must also meet the application's requirements.
What are 4K IOPS?
4K IOPS describes I/O operations using a 4 KB block size. The result should also identify whether the workload is random or sequential, its read/write mix and queue depth.
Why do SSDs have more IOPS than HDDs?
SSDs do not need to mechanically seek between data locations. This allows them to process random storage requests much more quickly than rotating hard drives.
What is the difference between IOPS and MB/s?
IOPS measures the number of storage operations completed per second. MB/s measures the amount of data transferred per second.
What is a good IOPS number?
There is no universal good IOPS number. The appropriate target depends on the workload, block size, read/write mix, concurrency, latency requirements and storage architecture.
Does RAID increase IOPS?
RAID can increase aggregate I/O capability by distributing operations across multiple drives, but RAID level, controller limits and write overhead affect real-world performance.
Does NVMe have higher IOPS than SATA?
Modern NVMe SSDs generally have substantially greater IOPS potential than SATA SSDs because NVMe uses PCIe and was designed for highly parallel non-volatile storage.