Azure Ultra Disk Performance: IOPS and Latency Guide Ultra Disk sits at the top of Azure's block storage lineup, and teams pick it when a database, transaction system, analytics job, or AI pipeline can't afford slow I/O. But a published spec is not an application result. VM limits, queue depth, caching, and workload shape decide what you actually get.

Many teams provision a high IOPS number, see no improvement, and pay for headroom the VM can't use. This guide covers what Ultra Disk performance represents, how its limits work, how to measure it, and where designs go wrong.

Key Takeaways

  • IOPS, latency, and throughput must be evaluated together, not one at a time.
  • Ultra Disk lets you set capacity, IOPS, and throughput independently.
  • The lowest limit across disk, VM, and workload sets your real ceiling.
  • Validate with representative tests and Azure Monitor, not theoretical maximums.

What Ultra Disk Performance Represents in Azure Managed Disks

Microsoft defines three core metrics for disk performance:

  • IOPS: I/O requests issued to the disk per second.
  • Throughput: data read or written per second, equal to IOPS multiplied by I/O size.
  • Latency: total time to complete one I/O request.

IOPS drives transaction rates. Latency drives how long a commit waits. Throughput drives bulk data movement, such as scans and backups.

Ultra Disk is a managed data disk for Azure VMs, aimed at SAP HANA, top-tier databases, and transaction-heavy workloads. It cannot be an OS disk. Its defining trait is that performance is a setting you choose, not a side effect of capacity.

Provisioned, Available, and Consumed

Three numbers get confused constantly:

  • Provisioned: what you configure
  • Available: what Azure lets the disk deliver after VM caps apply
  • Consumed: what your application actually drives

The storage path runs from application to OS, to VM, to Azure host, to disk. The slowest layer becomes the bottleneck. Raising disk IOPS does nothing if the VM is already capped.

What Shapes Real-World Results

VM size and disk limits both apply, and exceeding either one throttles I/O and can raise latency. Supported VM families, regions, and zones also limit where Ultra Disk runs. Regional support doesn't guarantee a specific SKU is available in a given zone.

Workload variables matter just as much:

  • Read/write ratio and sequential versus random access
  • Block size: small I/O favors IOPS, large I/O favors throughput
  • Queue depth and concurrency, where excessive depth can worsen latency
  • Burst patterns and synchronization requirements

Two applications needing the same IOPS can see different latency. A log writer flushing small synchronous writes behaves nothing like a scan-heavy analytics job.

Ultra Disk workload patterns affecting IOPS throughput and latency

Ultra Disk also doesn't support host caching, so guest OS, filesystem, database settings, and background jobs account for gaps between disk metrics and application results.

Range and Key Technical Properties of Azure Ultra Disk Performance

No single universal number describes Ultra Disk performance. Performance comes from configured values, service limits, VM capability, workload conditions, and the margin you leave.

Nominal Operating Range

Per Microsoft's disk types documentation (updated September 15, 2026), current values are:

Property Documented value
Capacity 4 GiB to 64 TiB, in 1 GiB increments
IOPS 100 minimum; up to 1,000 per GiB; capped at 400,000 per disk
Throughput 1 MB/s minimum; up to 0.25 MB/s per provisioned IOPS; capped at 10,000 MB/s per disk
Adjustments Up to 4 performance resizes per 24 hours; each can take up to 1 hour

Resizes can also fail if capacity isn't available. Check the page again before you design, since these values change.

Those are per-disk limits. VM aggregate limits are separate. For example, Standard_M176bds_4_v3 supports up to 650,000 aggregate uncached IOPS and 10,000 MB/s. Striping several disks, which SQL Server on Azure VMs guidance allows, raises aggregate throughput but never removes VM or workload ceilings.

Azure Ultra Disk and VM aggregate performance limits comparison

Boundary Limits

Know the restrictions before committing:

  • Ultra Disk uses LRS only, with no ZRS or availability sets, and isn't supported with Azure Compute Gallery.
  • Shared disks are supported (maxShares 1 to 15) but can't span availability zones.
  • Incremental snapshots are supported. Azure Backup supports Ultra under the Enhanced policy.
  • Azure Site Recovery requires High Churn support and doesn't allow zone-to-zone DR.
  • Azure Disk Encryption and double encryption at rest are unsupported. Server-side encryption and encryption at host are supported.

A disk provisioned below its maximum can still miss your target if the VM, guest OS, or application can't sustain the I/O.

Safe Operating Margin

Don't run at the published limit continuously. Leave room for demand surges, background operations, failover, maintenance, and measurement error.

Microsoft publishes no universal percentage, so set the margin by workload criticality: a payment ledger deserves more headroom than a nightly batch job. Define it from your own measured peaks and service-level objectives.

Stability and Variability

Microsoft describes Ultra Disks as designed for "consistently low submillisecond latency," but that wording carries no P95 or P99 condition. Average latency hides outliers, and outliers hurt transaction logs and user-facing apps.

Latency shifts with queue depth, saturation, I/O size, read/write mix, and concurrent activity. Microsoft's DiskSpd -L option reports P95 and P99 values separately for reads, writes, and totals. Azure Monitor's average latency is not an I/O percentile distribution.

Cost Trade-offs

Billing is hourly on provisioned capacity, IOPS, and throughput, not on usage. Ultra-enabled VMs without an attached Ultra Disk also incur a vCPU reservation charge. Guidance on rounding differs between Microsoft pages, so confirm against the current Azure pricing page. Weigh extra IOPS against a larger VM, and size to measured need.

How Ultra Disk Performance Is Specified, Measured, and Validated

Treat performance as both a design specification and an operational result tested against real service-level objectives.

Specification

Record these for every deployment:

  • Disk size, provisioned IOPS, and provisioned throughput
  • VM SKU, region, and zone
  • Caching mode, redundancy, OS, and filesystem
  • Workload assumptions

Keep four things separate: Microsoft-rated limits, provisioned values, observed benchmark results, and application objectives.

Measurement

Azure Monitor gives you the core signals:

  • Disk metrics: read/write operations and bytes per second, averaged per minute
  • Data Disk IOPS Consumed Percentage and Data Disk Bandwidth Consumed Percentage, where 100% means the disk cap
  • VM Uncached IOPS Consumed Percentage and bandwidth equivalents for VM saturation
  • Data Disk Queue Depth for outstanding I/O
  • Data Disk Latency (confirm availability; docs conflict on preview status and SCSI-only coverage)

For controlled tests, Microsoft's examples use DiskSpd with 4K blocks, four threads, and 128 outstanding I/Os, or fio with direct=1, bs=4k, iodepth=256, and numjobs=4. These are tool examples, not Ultra guarantees.

DiskSpd and fio controlled test configurations for Ultra Disk

Document every run with:

  • Block size, queue depth, and concurrency
  • Read/write mix and duration
  • Caching state
  • Disk-level vs. application-level scope

Run steady-state and peak loads. Report tail latency, not just averages.

Field Versus Lab

Production telemetry diverges from benchmarks because of backups, replication, antivirus scans, database checkpoints, and shifting demand. A practical workflow:

  1. Establish a baseline.
  2. Check whether the disk or the VM is saturated.
  3. Compare provisioned and consumed performance.
  4. Change one variable at a time.
  5. Confirm the application improved, not just the disk graph.

Five-step Ultra Disk performance validation workflow

Operational Visibility

Azure Monitor and Microsoft's limits remain your source of truth. Lucidity can complement them when you need visibility across cloud block storage. Lumen analyzes IOPS, throughput, latency, and cost metrics across AWS, Azure, and Google Cloud so you can spot mis-sized or idle volumes without infrastructure changes.

Sustained saturation starts a chain: I/O queues grow, database transactions slow, response times stretch, timeouts appear, and availability suffers.

In SQL Server, commits wait on log flushes, so WRITELOG waits often point to log-disk latency. Small transactions and CPU scheduling delays can cause them too.

Over-provisioning is the quieter failure. You pay hourly for IOPS the application never consumes when the real bottleneck is the VM, CPU, network, or concurrency.

Watch for these misreadings:

  • Treating provisioned IOPS as a guaranteed application result
  • Comparing latency figures with different I/O sizes, queue depths, regions, or VM SKUs
  • Assuming more IOPS lowers latency when the workload isn't disk-bound
  • Ignoring tail latency, bursts, and failover needs
  • Treating Microsoft's 80% lower P99.9 and 30% lower average latency claim as a benchmark when no baseline, VM, block size, or queue depth is disclosed

Even the BlackRock Aladdin deployment Microsoft cites on Ultra Disk and M-series VMs discloses no measured latency outcome.

Operating near a performance boundary also erodes resilience to spikes. Revalidate the full storage and VM configuration after every change.

Conclusion

Ultra Disk IOPS and latency are governing parameters shaped by provisioning, VM limits, workload behavior, and how you measure. Reliable design needs current Microsoft specifications, realistic testing, operating headroom, and monitoring at both disk and application level. Engineering judgment, not the biggest advertised number, should decide whether Ultra Disk fits your workload.

Frequently Asked Questions

Which Azure VM data disk is best for high-performance data access?

Ultra Disk targets the most demanding, latency-sensitive workloads. Premium SSD v2 or Premium SSD may suffice when measured needs are lower. Confirm VM support and current documentation first.

What are the differences between Azure Files and Azure Disk Storage?

Azure Disk Storage provides block volumes attached to VMs. Azure Files provides managed SMB or NFS file shares. Their performance models, access patterns, and use cases differ.

What are the performance tiers in Azure Files?

Azure Files uses SSD (premium) and HDD (standard) media, with provisioned v1, provisioned v2, and pay-as-you-go billing models. HDD pay-as-you-go adds transaction optimized, hot, and cool access tiers. None of these are managed disk types like Ultra Disk.

What is the maximum IOPS for Azure Ultra Disk?

Microsoft documents up to 400,000 IOPS and 10,000 MB/s per disk. Your VM size and workload set what you can actually reach.

How does Ultra Disk latency compare with Premium SSD v2?

Microsoft describes both as submillisecond, with no matched-load comparison. Compare average and tail latency under your own workload, VM configuration, and service-level objective.

How do I monitor Ultra Disk IOPS and latency?

Use Azure Monitor disk and VM metrics, plus guest OS and application telemetry. Compare consumed IOPS, throughput, queue depth, and latency against provisioned values and VM limits.