Skip to content

Independent marketplace for verified cloud accounts

Reading Cloud Storage Performance Specifications

IOPS, throughput and latency are three promises, not one. How to read a storage spec sheet so the performance you get matches the performance you expected.

Storage specifications are where cloud providers do their most careful writing. Nothing on the page is false, and yet the number you remember is almost never the number you get. The gap is not dishonesty, it is that disk performance is several variables pretending to be one, and the marketing surface only has room for the flattering one.

If you can read a storage spec properly you can predict your own performance before you provision, which is a better position than discovering it under load.

IOPS, throughput and latency are three different promises

IOPS counts operations per second. Throughput counts bytes per second. Latency measures how long one operation takes. They are related but you cannot derive any one of them from the other two, and a volume can be excellent at one while being poor at another.

Which one you care about depends entirely on your access pattern. A transactional database doing many small scattered reads lives and dies on IOPS and latency, and barely registers on throughput. A video pipeline or an analytics scan reading large sequential blocks cares about throughput and is almost indifferent to IOPS. A build server doing both looks fine on the spec sheet and feels slow in practice, because its pattern is mixed and the volume is tuned for neither.

Work out which of the three your workload actually consumes before comparing products. Most people compare the IOPS number because it is the largest.

Block size turns one number into the other

Here is the relationship that explains most confusion: throughput equals IOPS multiplied by block size. A volume rated for a given IOPS figure at four kilobyte blocks delivers sixteen times the bytes per second at sixty four kilobyte blocks, assuming nothing else limits it. Something else almost always limits it.

Cloud block volumes typically carry both an IOPS ceiling and a separate throughput ceiling, and you hit whichever you reach first. With small blocks you hit the IOPS limit while throughput sits idle. With large blocks you saturate throughput long before the IOPS number means anything. Providers quote the IOPS figure at the smallest block size, because that is where it is highest, and quote throughput at a large block size, because that is where it is highest. Both numbers are true and they cannot be achieved simultaneously.

Check your own block size before believing any figure. Database engines usually use a fixed page size you can look up. Filesystems have a default you can query. If your real block size sits between the two measurement points, interpolate and expect to land below the line.

Burst credits and the cliff at the end of them

Many entry level volume types are not sold at a sustained rate at all. They are sold at a baseline rate plus an accumulating allowance that lets you exceed it for a while. Idle time accrues credit, busy time spends it, and when the balance reaches zero performance drops to the baseline without warning.

This is the single most common cause of the complaint “it was fast for the first twenty minutes”. Benchmarks finish inside the burst window and report the burst figure. Production runs for hours and lives at the baseline. The baseline is often a small fraction of the burst rate, and on some volume types it scales with provisioned capacity, meaning a small volume is slow regardless of what you are willing to pay for IOPS.

Burst behaviour applies to more than disks: network bandwidth and CPU on smaller instance families frequently work the same way. When a workload runs continuously rather than in spikes, burstable resources are a trap, and the fix is to buy the sustained tier even though its headline number looks worse.

Queue depth, and why a benchmark hits numbers your application never will

The published IOPS ceiling assumes many operations are in flight at once. Storage reaches peak throughput through parallelism: dozens of outstanding requests being serviced concurrently. Benchmarking tools default to high queue depths because that is how you demonstrate the maximum.

Plenty of real applications are single threaded and synchronous. They issue one read, wait for it, then issue the next. At queue depth one, your achievable IOPS is simply one divided by the latency per operation, and no amount of provisioned IOPS changes that. A volume advertised at a large IOPS number will deliver a few thousand at best to a strictly serial workload, because physics and the round trip to a network attached device set the ceiling.

The practical consequence: if your application is latency bound rather than parallelism bound, buying more provisioned IOPS is wasted money. Buy lower latency instead, which usually means local storage attached to the host rather than network attached volumes, and accept that local storage does not survive the instance.

Read, write, and the cost of durability

Writes are generally more expensive than reads, and the gap widens as durability increases. A durable network volume acknowledges a write only once it has been committed to multiple independent copies, and that replication adds latency on the write path that reads never pay.

Spec sheets sometimes quote a single IOPS figure covering both, sometimes separate figures, and sometimes a mixed ratio such as seventy percent read. If your workload is write heavy, a figure quoted at a read heavy mix is not describing your situation. Equally, filesystem and database behaviour can amplify writes considerably: a small logical update may become a journal write, a page write and an index write, so the operations reaching the volume outnumber the operations your application issued.

Measure at the volume, not at the application, when you are sizing. The difference between the two is write amplification, and it is frequently a factor of several.

What the spec sheet does not tell you

Three things are routinely absent. First, tail latency. The average is usually fine and the ninety ninth percentile is what your users feel, particularly in a system where one slow request blocks a connection pool. Second, behaviour under contention from other tenants, which on shared storage is real and unquantified. Third, how performance changes as the volume fills, which for some underlying technologies is significant.

Filesystem and operating system layers add their own distance between the spec and your experience. The scheduler, the page cache, the choice of filesystem and whether barriers and flushes are honoured all sit between your application call and the device. A workload that fits comfortably in page cache will report numbers far above anything the volume can do, which is flattering and useless. A workload issuing synchronous flushes on every commit will report numbers far below, which is accurate and alarming. Both are measuring the stack rather than the disk.

There is also a quieter one: whether the quoted figure is a guarantee or a maximum. Provisioned performance tiers promise a rate you have paid for. General purpose tiers describe a rate you may reach. These read almost identically on the page and behave entirely differently at month end when everyone else’s batch jobs run too.

Providers differ here in philosophy as much as in hardware. Operators selling simple fixed specification instances tend to be more predictable and less configurable, while hyperscaler storage is highly tunable and highly conditional. Accounts such as Buy UpCloud Account and Buy Hetzner Account sit toward the predictable end, while Buy AWS Account gives you more dials and more ways to misread them. If you want that contrast laid out directly, the AWS and Hetzner comparison covers it.

Running your own measurement

Half an hour of testing beats any amount of specification reading. Run the test for longer than the burst window, at least thirty minutes and ideally an hour, so you see the sustained rate rather than the opening sprint. Use your application’s real block size, not the tool’s default. Run at the queue depth your application actually produces, which for most single threaded services is one. Record percentiles, not just the mean. And test a volume of the same size as the one you will run, because on many products performance scales with provisioned capacity.

Then write the result down next to the vendor figure. Over a few projects you build a personal correction factor for each provider and volume type, and that correction factor is worth more than any published benchmark. Pair this with our guide to the different cloud storage types so you are at least comparing the right category before you start measuring.

Related comparisons

AWS vs Hetzner
vs

AWS vs Hetzner

For steady always-on compute the price gap between Hetzner and AWS is wide. This comparison works out what that gap…

AWS vs Google Cloud
vs

AWS vs Google Cloud

A comparison built around data gravity: how AWS and Google Cloud expect you to store, move and query data, and…

AWS vs IBM Cloud
vs

AWS vs IBM Cloud

IBM Cloud is built around OpenShift, regulated industries and hardware you cannot rent elsewhere. This comparison sets that platform thesis…

AWS vs Digital Ocean
vs

AWS vs Digital Ocean

How far does a small team get on DigitalOcean before platform complexity starts costing more than it saves? This comparison…

Accounts discussed in this guide

Buy AWS Account
Buy AWS Account
Amazon Web Services

Buy AWS Account

Amazon Web Services has become synonymous with cloud computing itself, and for millions of developers and businesses around the world, an AWS account is…

From$25 14 options
Buy Hetzner Account
Buy Hetzner Account
Hetzner

Buy Hetzner Account

Hetzner has earned a reputation among developers as one of the most powerful and affordable cloud platforms in the world.

From$50 3 options

Keep reading

How to Verify Cloud Credits Before and After a Transfer
Cloud Accounts 7 min read

How to Verify Cloud Credits Before and After a Transfer

Credit balances are provider-controlled, time-limited and service-restricted. This guide shows where each billing console displays them and how to verify a balance…

Root Email, MFA and API Keys: Securing an Inherited Account
Security and IAM 7 min read

Root Email, MFA and API Keys: Securing an Inherited Account

Credential hygiene for an account someone else created. Covers the root mailbox, replacing authentication factors, closing recovery paths and rotating keys without…

Service Limits and Quotas Explained
Cloud Accounts 7 min read

Service Limits and Quotas Explained

Quotas decide what your account can create before you ever hit a technical ceiling. This guide covers soft and hard limits, where…

Need the account, not just the theory?

Browse documented configurations across every provider in the catalog, or ask us which one fits your workload.

Telegram