Azure Ultra Disk Latency and IOPS Explained Your database team provisions an Ultra Disk at 160,000 IOPS, runs the workload, and sees half that. The disk isn't broken. Something else is capping the result, and finding it is the real work.

Latency and IOPS are the two numbers that decide whether Ultra Disk earns its price on databases, transaction processing, analytics, and AI workloads. But provisioned performance is a ceiling, not an outcome. VM limits, I/O size, queue depth, and read/write mix all change what your application sees.

This guide covers what the metrics mean, the documented operating range, how to measure real behavior, and the mistakes that cause the most confusion.

Key Takeaways

  • IOPS, throughput, and latency measure different things and must be evaluated together.
  • The lower of the disk limit and the VM limit governs real performance.
  • Average latency, tail latency, I/O size, and queue depth shape application behavior.
  • Validate limits against current Microsoft documentation and Azure Monitor data before resizing.

What Azure Ultra Disk Latency and IOPS Represent

Microsoft defines the three core terms in its premium storage guidance:

  • IOPS: I/O requests issued to a disk per second
  • Throughput: data read or written per second
  • Latency: total time to complete one I/O request

They aren't interchangeable. A workload of 4 KiB random reads can hit an IOPS cap while moving very little data. A backup streaming large blocks can hit the throughput cap with modest IOPS.

Ultra Disk is Azure managed block storage. Your application issues I/O, the VM and storage path process it, and the disk's provisioned IOPS and throughput act as ceilings. They are not guaranteed application results.

Ultra Disk also lets you configure capacity, IOPS, and throughput independently. You can raise IOPS without buying more capacity, which most other disk tiers don't allow. Confirm current ranges and adjustment rules in Microsoft's documentation before you design around them.

Factors That Influence Latency and IOPS in Practice

Several variables decide whether a workload ever reaches the published ceiling:

  • I/O size: Throughput = IOPS × I/O size. Larger requests consume more bandwidth per operation.
  • Queue depth: Queue depth = IOPS × latency. More concurrency can raise IOPS, but excessive depth worsens latency.
  • Read/write mix and bursts: Write-heavy or bursty patterns can raise observed latency.
  • VM size: Aggregate uncached IOPS and bandwidth limits on the VM can become the real bottleneck.
  • Caching: Ultra Disk doesn't support host caching, so you can't rely on it to absorb reads.

Which Workloads Fit Ultra Disk

Microsoft positions Ultra Disk for I/O-intensive workloads such as SAP HANA, top-tier databases, and transaction-heavy systems. Other tiers may fit better when latency demands are looser.

Tier Max IOPS Max throughput
Ultra Disk 400,000 10,000 MB/s
Premium SSD v2 80,000 2,000 MB/s
Premium SSD 20,000 900 MB/s
Standard SSD 6,000 750 MB/s

Azure managed disk tiers compared by maximum IOPS and throughput

These are per-disk maximums from Microsoft's disk types page, not application guarantees. A web server or dev/test environment rarely justifies Ultra pricing.

Range and Key Technical Properties of Azure Ultra Disk Performance

Usable performance depends on disk configuration, the attached VM, workload profile, region, and supported features. Treat every published figure as conditional.

Nominal Operating Range

Current documented provisioning rules:

Property Documented rule
Capacity 4 GiB to 64 TiB, in 1 GiB increments
IOPS 100 minimum; up to 1,000 per GiB; 400,000 per disk
Throughput 1 MB/s minimum; up to 0.25 MB/s per provisioned IOPS; 10,000 MB/s per disk

Microsoft describes Ultra Disk latency as consistently submillisecond. That is a general claim, not a published p99 or p99.9 guarantee.

In its November 2025 announcement, Microsoft reported an 80% reduction in P99.9 and outlier latency and a 30% improvement in average latency. Those are relative gains, with no absolute values or test methodology supplied.

Boundary Limits

The disk is only half of the equation. Microsoft's performance documentation states that both disk and VM caps apply, and the lower one limits you. A disk provisioned at 160,000 IOPS on a VM capped at 80,000 uncached IOPS delivers 80,000 at best.

Azure Ultra Disk and VM limits showing effective IOPS bottleneck

Other boundaries to check before deploying:

  • Data disk only; no OS disk use
  • Locally redundant storage only; no ZRS or availability sets
  • Support varies by VM series, size, region, and zone
  • Shared disks supported, with maxShares up to 15
  • Azure Disk Backup supports Ultra; VM Backup requires the Enhanced policy

Snapshot background copying can also raise latency on affected disks, so schedule restores with care.

Safe Operating Margin

Microsoft identifies 100% consumption as the cap. It gives no fixed Ultra-specific spare-capacity percentage, so set your own based on data.

  • Keep headroom for traffic spikes, maintenance, and background jobs.
  • Use Azure Monitor history and workload forecasts to decide what to raise.
  • Increase only the constrained resource: IOPS, throughput, capacity, or VM size.

Upgrading everything at once is expensive and rarely necessary.

Stability and Variability

Average latency hides outliers. A disk averaging 0.5 ms can still produce slow requests that stall a transaction.

Tail metrics such as P99 and P99.9 matter for latency-sensitive systems, but Azure Monitor's Data Disk Latency metric reports an average, not a percentile. Capture tail latency in your benchmark tool or application telemetry.

Latency also moves with queue depth, request size, read/write activity, contention, and short bursts, even when provisioned settings stay constant.

Coupling With Adjacent Variables

Use this formula to see which limit you'll hit first:

Throughput (MB/s) = IOPS × I/O size (MB)

  • 100,000 IOPS at 8 KiB is roughly 800 MB/s. IOPS binds first.
  • 20,000 IOPS at 512 KiB is roughly 10,000 MB/s. Throughput binds first.

Azure IOPS and throughput calculation examples by I/O size

Trade-offs follow from there:

  • More disk performance: simple, but useless if the VM caps you.
  • Bigger VM: raises ceilings, but also compute cost.
  • More disks: spreads load, adds operational complexity.
  • Different tier: cheaper, but may not hold latency targets.

Azure Disk Storage reservation discounts don't apply to Ultra Disks, and an Ultra-enabled VM without an attached Ultra Disk incurs a per-vCPU charge.

How Azure Ultra Disk Latency and IOPS Are Specified, Measured, and Validated

Treat an Ultra Disk configuration as a hypothesis: "this setup meets our SLO." Then test it against measured behavior.

Specification and Documentation

Use current Microsoft documentation for limits, VM scalability targets, region support, and live-change rules. Skip old Ultra SSD preview material.

Then translate requirements into numbers:

  1. Set a target transaction or log-write latency.
  2. Estimate peak concurrent I/O and typical block size.
  3. Calculate required IOPS and throughput from those.
  4. Add batch windows and recovery needs.
  5. Check the result against the VM's uncached limits.

Five-step Azure Ultra Disk performance specification workflow

Measurement and Verification

Azure Monitor gives you most of what you need, according to Microsoft's disk metrics guide:

  • Data Disk IOPS Consumed Percentage
  • Data Disk Bandwidth Consumed Percentage
  • Data Disk Queue Depth
  • Data Disk Latency (average)
  • VM Uncached IOPS and Bandwidth Consumed Percentage

Compare the disk percentages with the VM percentages. If the VM hits 100% first, resizing the disk won't help.

Metrics show whether you are throttled; benchmarks show whether the disk meets your latency and IOPS targets. Microsoft recommends DiskSpd on Windows and fio on Linux. Microsoft's guide has no Ultra-specific profile, so build your own:

  • Vary block size, queue depth, threads, and read/write mix
  • Run long enough to see steady behavior
  • Record average and tail latency, not just peak IOPS

Azure Ultra Disk benchmark variables and latency validation workflow

In production, validate during representative peaks. Separate disk throttling from VM saturation, application queuing, network delays, and database configuration issues.

For enterprise teams running many disks, Lucidity's Lumen tool offers continuous visibility into IOPS, throughput, latency, and cost across Azure, AWS, and Google Cloud. Its recommendations rest on historical data rather than point-in-time snapshots, and it can flag disks that belong on a different tier.

The chain is simple. Too little IOPS or throughput causes queuing, and queuing raises latency. Too much provisioning raises cost with no application benefit.

Lucidity's research pegs that second problem at over 50% storage underutilization and an estimated $21.5 billion in wasted block storage spend by 2028.

Performance Loss Mechanisms

Per Microsoft, requests beyond disk or VM allocations are capped, and applications can see increased latency. Sustained saturation, undersized VMs, unsuitable I/O patterns, and no headroom lead to:

  • Slow transactions and timeouts
  • Longer batch windows
  • Degraded user experience

Reliability Considerations

Azure abstracts the physical media, so media wear isn't your concern. Throttling, queue buildup, and service limits are the risks that matter in practice.

Microsoft's reliability guidance lists at least 99.999999999% durability with three replicas, but Ultra supports LRS only.

It's a single-region service with no native automatic cross-region failover, so a regional outage can make disks unavailable. Plan disaster recovery with tools such as Azure Site Recovery.

Common Misinterpretations

  • A published maximum is a guarantee. It isn't; VM caps and workload shape decide.
  • More IOPS always means lower latency. Not when queue depth is the cause, or when the workload is throughput-bound.
  • Disk limits are the only constraint. VM limits often bind first.
  • Lab benchmarks predict production. They rarely match real mixes and bursts.
  • Over-provisioning is safe insurance. It costs money every hour without measured demand.

Conclusion

Ultra Disk latency and IOPS set hard bounds on what your workload can achieve. Reliable performance depends on knowing the supported range, the VM dependencies, the workload's I/O behavior, and the conditions behind each published number.

Validate specifications against Azure Monitor and application telemetry. Then keep right-sizing. Match tier and limits to measured demand instead of defaulting to the fastest option.

Frequently Asked Questions

What is IOPS and throughput in Azure?

IOPS is the number of I/O operations per second; throughput is the data transferred per second. Both are capped by disk provisioning and by the VM's limits.

What are the limitations of Premium SSD v2 disks in Azure Virtual Machines?

Premium SSD v2 can't be an OS disk, doesn't support host caching, and has region and zone dependencies. It tops out at 80,000 IOPS and 2,000 MB/s, versus 400,000 and 10,000 MB/s for Ultra.

What type of storage offers the lowest latency in Azure?

It depends on the path. Local temporary NVMe avoids the network but loses data after stop, deallocation, or maintenance. Ultra Disk is persistent and durable, and Microsoft publishes no universal latency ranking.

What is the latency of Azure Ultra Disk?

Microsoft describes it as consistently submillisecond, without a published absolute p99 or p99.9 figure. Results depend on I/O size, queue depth, VM limits, and workload mix.

How do Azure VM limits affect Ultra Disk IOPS?

Effective IOPS can't exceed the lower of the disk's provisioned IOPS and the VM's uncached limit. Add up all attached disks when evaluating the VM.

Can you change Azure Ultra Disk IOPS without downtime?

Yes. You can adjust IOPS and throughput without detaching the disk or restarting the VM. Microsoft allows four changes per 24 hours, taking up to one hour to apply. Capacity expands online but can't shrink, and is blocked during background copy.