
This guide is for Azure architects, DevOps, ITOps, database, and FinOps teams. It covers how to measure real demand, tune IOPS and throughput independently, check the VM limit, validate every change, and stop paying for performance nobody uses.
Key Takeaways
- IOPS, throughput, and latency are separate targets; tuning one can leave another as the constraint.
- Premium SSD v2 separates capacity from performance, so size each to its own requirement.
- The lowest limit wins: disk, VM, guest OS, or application.
- Baseline with Azure Monitor and OS tools before raising any setting.
What IOPS and Throughput Mean in Azure Premium SSD v2
IOPS, throughput, and latency are different tuning targets
Microsoft defines IOPS as storage requests per second, throughput as data transferred in an interval, and latency as the total time to complete a request. The three connect through one equation: IOPS x I/O request size = throughput.
That is why block size matters. Small random I/O hits the IOPS ceiling first. Large sequential I/O hits the throughput ceiling first.
Three hypothetical workloads show the difference:
- OLTP database: many 8 KB random operations, so IOPS and latency dominate
- Analytics job: large sequential scans, so throughput dominates
- File server: mixed sizes with moderate IOPS, so a balanced profile fits
Premium SSD v2 versus Premium SSD
| Attribute | Premium SSD v2 | Premium SSD (v1) |
|---|---|---|
| Capacity | 1 GiB to 64 TiB, 1 GiB increments | 4 to 32,767 GiB, rounded to a tier |
| Baseline | 3,000 IOPS, 125 MB/s included | Tier-dependent (smallest: 120 IOPS, 25 MB/s) |
| Max per disk | 80,000 IOPS, 2,000 MB/s | 20,000 IOPS, 900 MB/s (burst 30,000 / 1,000 MB/s) |
| Billing | Capacity, IOPS, and throughput separately | Tier, redundancy, optional bursting |
| OS disk | No | Yes |
| Host caching | No | Disks under 4 TiB |
| Redundancy | LRS only | LRS or ZRS |
Source: Microsoft Learn disk types and the deployment guide. Confirm values against current documentation and the Azure pricing page before you commit.
If you need an OS disk, host caching, or ZRS, Premium SSD stays on the table.
Translate application demand into storage requirements
Collect these before choosing settings:
- Read and write IOPS
- Average and peak throughput
- Block size and read/write mix
- Queue depth
- Latency objective
Averages alone mislead. Size for sustained peaks, and treat checkpoints, backups, scheduled jobs, and traffic surges separately. Then write the goal in application terms, such as transaction latency or batch completion time. A disk spec is not the goal.
Identify the active performance boundary
Effective performance is the lower of the disk limit and the VM's uncached limit. Those VM limits are aggregate budgets shared across all attached disks, not per-disk allowances. Record the VM SKU and disk layout in every assessment.

If disk and VM metrics both look healthy, look at guest-OS queueing, filesystem behavior, database settings, CPU, memory, or the network.
How to Tune Premium SSD v2 IOPS and Throughput
Establish a measured target before changing the disk
Build the target from observed peak IOPS, peak throughput, latency, workload criticality, and a written headroom policy. Separate normal baseline from short bursts and sustained peaks. Provision for the sustained peak, and decide deliberately how to handle shorter spikes.
Set different targets for read-heavy, write-heavy, random, and sequential volumes. One generic number for every disk wastes money.
Set capacity, IOPS, and throughput independently
Size capacity on usable data, growth, snapshots, and backups. Don't buy capacity just to unlock performance. Then set IOPS from measured operations and throughput from block size and transfer volume.
Microsoft's example creates a 100 GiB disk with 5,000 IOPS and 150 MB/s:
az disk create -g "$rg" -n "$disk" --location "$region" --zone "$zone" \
--sku PremiumV2_LRS --size-gb 100 --logical-sector-size 4096 \
--disk-iops-read-write 5000 --disk-mbps-read-write 150
az disk update -g "$rg" -n "$disk" \
--disk-iops-read-write 5000 --disk-mbps-read-write 200
The default sector format is 4Kn. Use --logical-sector-size 512 for 512E.
Validate the relationship between IOPS and throughput
Provisioned IOPS caps throughput at 0.25 MB/s per IOPS, up to 2,000 MB/s. Microsoft's examples: 3,000 IOPS allows 750 MB/s, 4,000 allows 1,000 MB/s, and 8,000 or more allows 2,000 MB/s. Above 6 GiB, maximum IOPS scales at 500 per GiB, so 80,000 IOPS needs at least 160 GiB.
A hypothetical: 10,000 IOPS at 8 KB blocks moves about 80 MB/s. Raising IOPS won't help if you're already throughput-bound, and raising throughput won't help if the app issues too few operations to use it. This is an illustration, not a guarantee.

Apply changes safely
Each disk allows four performance adjustments per 24 hours, and creation counts as one. Adjustments need no downtime. Microsoft's deployment guidance doesn't specify propagation time, so don't count on reactive changes during an incident.
- Schedule planned changes
- Record the previous configuration
- Validate afterward
- Keep a rollback or workload-protection plan
Account for workload and platform behavior
Several factors cut realized performance:
- I/O alignment and sector size
- Queue depth and application concurrency
- Filesystem settings and database page size
For SQL Server, Microsoft recommends 64 KB NTFS allocation units on data-file disks. Also note that v2 has no host caching, and Microsoft's bursting documentation does not cover Premium SSD v2. Revalidate any Premium SSD bursting assumptions during migration.
Before production, test with representative data, read/write patterns, block sizes, concurrency, and peak load.
How to Measure and Validate Premium SSD v2 Performance
Start with Azure Monitor
On the VM resource, watch these metrics:
- Data Disk IOPS Consumed Percentage
- Data Disk Bandwidth Consumed Percentage
- VM Uncached IOPS Consumed Percentage
- VM Uncached Bandwidth Consumed Percentage
- Data Disk Read/Write Operations/Sec and Bytes/Sec
Split disk metrics by LUN. A high disk percentage means the disk is near its limit. A high VM-uncached percentage means the VM is near its limit. Data Disk Latency is listed for SCSI; check current Azure Monitor docs for NVMe latency coverage.
In the portal, open the VM, go to Metrics, and chart the percentages. From the CLI:
az monitor metrics list --resource "$vmId" --interval 1m \
--metrics "Data Disk IOPS Consumed Percentage" \
"VM Uncached IOPS Consumed Percentage" --aggregation Average
Cover normal traffic, peaks, scheduled jobs, and past incidents. For guest counters, Azure Monitor Agent feeds the Perf table:
Perf
| where ObjectName == "LogicalDisk"
| summarize avg(CounterValue) by CounterName, InstanceName, bin(TimeGenerated, 5m)
Use OS-level tools for root-cause analysis
On Linux, iostat shows reads and writes per second, kB_read/s and kB_wrtn/s for throughput, await for latency, and queue depth. On Windows, Performance Monitor tracks disk reads and writes per second, bytes per second, average disk seconds per read or write, and queue length.
Correlate both with Azure Monitor. Guest counters can't reveal VM-level or platform limits.
Benchmark carefully and interpret results correctly
Run FIO or DiskSpd only in a test environment or maintenance window, on a test file that can't overwrite production data. Microsoft's examples:
- FIO random write:
direct=1,ioengine=libaio,bs=4k,iodepth=256,numjobs=4 - DiskSpd:
-w100 -b8K -F4 -r -o128 -W30 -d30 -Sh - Throughput test: 64 KB requests, high queue depth, multiple threads
Document block size, access pattern, read/write mix, queue depth, concurrency, duration, warm-up, caching, and VM SKU. Microsoft's benchmark page publishes no v2-specific results, so treat any synthetic number as a ceiling check, not a prediction of application behavior.

Validate after every tuning change
Compare before and after:
- IOPS, throughput, average and tail latency
- Queueing and application response time
- Cost
Confirm the new setting is actually consumed and that the bottleneck didn't just move to the VM, OS, or database. Set an observation period, acceptance criteria, and alerts for renewed saturation or unexpected cost growth.
Monitor large estates without manual disk-by-disk analysis
Across many subscriptions, regions, and VM SKUs, checking disks one at a time doesn't scale. Lucidity Lumen is one optional visibility layer. It analyzes IOPS, throughput, latency, and cost metrics for every disk and scores each disk's tier against actual usage.
Its recommendations rest on historical trends, not point-in-time snapshots. That helps surface chronically over-provisioned, idle, or repeatedly saturated volumes. The tuning method above works with or without it.
Troubleshoot Bottlenecks and Platform Constraints
Diagnose throttling and low realized performance
Hitting a disk or VM limit rarely produces a clear storage error. Look for rising latency, growing queues, flat effective IOPS, and application slowdown. Microsoft notes that exceeding either IOPS or throughput allocation causes capping and can raise latency.
Use this decision path:
- Disk consumption high: review IOPS or throughput provisioning.
- VM consumption high: evaluate a larger VM SKU.
- Both low: investigate the guest OS, application, database, network, CPU, or memory.

More Premium SSD v2 IOPS cannot overcome a lower VM ceiling.
Check deployment and compatibility limitations
Verify these before design or migration:
- Region and zone: East US, East US 2, and West US 2 are zonal; North Central US, West Central US, and West US are nonzonal
- Live SKU check: run
az vm list-skus --resource-type disksto confirm current support - VMs: a Premium Storage-capable size is required, and nonzonal VMs have restrictions in most availability zone regions
- Unsupported: OS disks, host caching, Azure Compute Gallery, and ZRS
- Protection: Azure Backup needs the Enhanced policy; Site Recovery supports v2 as a source disk only, with High Churn enabled
- Snapshots: incremental only, and attaching a disk during background copy can affect performance
Avoid common tuning mistakes
- Assuming the highest IOPS is best
- Assuming capacity alone raises v2 performance; IOPS and capacity are set separately
- Trusting one benchmark run
- Raising disk performance before checking the VM limit
- Ignoring throughput
- Sizing from averages only
- Treating cost optimization as one-time work
Document the reason for every change and recheck demand as data volumes and application versions evolve.
Control Cost and Choose the Right Azure Disk
Compare total cost, not just the disk label
Premium SSD v2 bills hourly on provisioned GiB, IOPS above 3,000, and throughput above 125 MB/s. Prices vary by region and configuration, so pull current US-region figures from the Azure pricing page or Retail Prices API.
Compare like-for-like setups: the same capacity, IOPS, throughput, VM size, and redundancy. To find waste, calculate:
- Provisioned IOPS minus observed peak IOPS
- Provisioned MB/s minus observed peak throughput
- Provisioned GiB minus used GiB
- VM uncached headroom beyond what any disk can use
The FinOps Foundation suggests comparing provisioned versus utilized IOPS over three months, evaluating peaks separately.
Build a right-sizing operating loop
- Review utilization after peak periods.
- Reduce unused performance, respecting the four-per-day change limit.
- Alert on persistently low utilization, repeated saturation, unexpected configuration changes, and workload growth.
- Validate with the same monitoring used for tuning.
Cost cuts must preserve latency and availability objectives. Deleting guest files never shrinks a managed disk. Microsoft doesn't support in-place shrinking.
Choose among Premium SSD v2, Premium SSD, and Ultra Disk
- Premium SSD v2: independently tunable data-disk IOPS and throughput that fit the workload and deployment constraints
- Premium SSD: OS-disk support, caching, ZRS, or compatibility needs take priority
- Ultra Disk: workloads that exceed v2 per-disk limits (Ultra reaches 400,000 IOPS and 10,000 MB/s), such as SAP HANA or transaction-heavy systems

Verify limits, regions, VM compatibility, pricing, and features before finalizing. Then run this checklist:
- Measure workload peaks and latency goals
- Set disk capacity, IOPS, and throughput separately
- Check VM aggregate limits
- Confirm region, zone, backup, and DR constraints
- Enable alerts and monitoring
- Calculate cost like-for-like
- Document the validation plan
Frequently Asked Questions
What are the differences between Azure Premium SSD v2 and Premium SSD?
Premium SSD v2 provisions capacity, IOPS, and throughput independently, with up to 80,000 IOPS and 2,000 MB/s per disk. Premium SSD uses fixed tiers but supports OS disks, host caching, and ZRS.
What are the limitations of Azure Premium SSD v2 disks in Virtual Machines?
Premium SSD v2 is data-disk only, with no host caching, ZRS, or Azure Compute Gallery. It needs a Premium Storage-capable VM in a supported region or zone, and allows four performance changes per 24 hours.
How can I achieve 99.99% availability in Azure?
Disk performance doesn't determine availability. Microsoft's 99.99% VM SLA requires at least two instances across two or more availability zones in one region. Add backup, monitoring, and application-level resilience.
Which Azure disk types cannot be used as an OS disk?
Premium SSD v2 and Ultra Disk are data disks only. Confirm support for your VM and architecture before designing around either.
What are the key differences between Azure Premium storage and standard storage?
Microsoft positions Premium SSD for production and performance-sensitive workloads, and Standard SSD or HDD for less demanding use. Premium costs more but delivers more predictable performance.
Can I shrink my Azure disk size?
No. Azure doesn't support in-place shrinking, and deleting files inside the guest doesn't reduce provisioned size. Create a smaller data disk, copy the data with Robocopy or rsync, verify it, then detach and delete the original. OS-disk downsizing isn't supported.