
Many teams struggle with the same problems: they buy capacity to get throughput they don't need, or they size right at a ceiling and watch latency climb. Both mistakes come from skipping the details behind the ratio: service level, scope, units, and workload.
This guide covers the definitions, formulas, limits, validation steps, and common pitfalls.
Key Takeaways
- Throughput per TiB is throughput divided by provisioned capacity, not total volume or pool throughput.
- Standard, Premium, and Ultra tie throughput to capacity; Flexible separates them.
- IOPS, latency, I/O size, concurrency, and client limits shape real results.
- Leave headroom, and validate with production-like tests before sizing.
What Throughput per TiB Represents in Azure NetApp Files
Throughput per TiB is the storage throughput tied to each tebibyte of provisioned capacity, expressed in MiB/s per TiB. Microsoft documents binary units (GiB, TiB, MiB/s), so keep them consistent rather than mixing in decimal GB or MB/s.
The basic calculation
Divide throughput by the relevant provisioned capacity:
Throughput per TiB = throughput (MiB/s) ÷ provisioned capacity (TiB)
Before you divide, decide what the numerator covers: a volume, a capacity pool, one workload, or a whole deployment. Also label it as documented, provisioned, benchmarked, or observed.
For example, a 2-TiB Premium volume at 64 MiB/s per TiB gives 128 MiB/s. The same service level on a 100-GiB volume yields only 6.25 MiB/s.
What the ratio is not
The ratio normalizes differently sized configurations for comparison. It is not a separate service limit or an end-user guarantee. It also differs from related metrics:
- Total throughput is the absolute rate, with reads and writes combined.
- IOPS counts operations per second; throughput is IOPS times I/O size.
- Latency is the time each operation takes.
- Capacity utilization is how much of the provisioned space holds data.
Two configurations with the same ratio can feel very different. Small random requests from many clients stress IOPS and latency, while large sequential reads stress throughput.
Microsoft's FAQ relates the two with IOPS = (MBps Throughput / KB per IO) * 1024.
Factors That Influence Throughput per TiB in Real-World Operation
A calculated ratio often differs from production results because the ratio only describes the storage allowance. Everything between your application and the volume also counts.
Configuration factors
- Service level, volume size, and QoS type (auto or manual) set the volume limit.
- Regular volumes, large volumes, and breakthrough-mode volumes have different ceilings.
- Quotas and regional availability vary, so check the latest Microsoft documentation.
Workload factors
- Sequential versus random access changes how efficiently you use the allowance
- Read/write mix and I/O size shape realized throughput under the same limit
- Concurrency, burstiness, and client count affect whether you sustain the volume cap
Environmental factors
- VM storage-I/O and network limits cap what a client can push. Adding NICs doesn't raise total bandwidth.
- NFS
nconnectsupports up to 16 connections, and Microsoft identifies 4 as optimal in most environments. Combining it withsec=krb5*can degrade performance. - SMB Multichannel is on by default, while signing and encryption cost client CPU.
- Caching and single-flow tests can distort results.
- Monitoring intervals hide short spikes.
Range and Key Technical Properties of Throughput per TiB
Throughput per TiB is not one fixed number. It depends on ANF design limits and operating conditions, so confirm every figure against current Microsoft sources.
Nominal Operating Range
| Service level | Documented model |
|---|---|
| Standard | Up to 16 MiB/s per TiB |
| Premium | Up to 64 MiB/s per TiB |
| Ultra | Up to 128 MiB/s per TiB |
| Flexible | 128 MiB/s included per pool, up to 640 MiB/s per TiB of pool capacity |
Flexible works differently. The included 128 MiB/s applies to the pool regardless of size, and you can add extra throughput in 1 MiB/s increments.
A 1-TiB Flexible pool runs 128 to 640 MiB/s. A 10-TiB pool runs 128 to 6,400 MiB/s. Per TiB is a derived comparison here, not the provisioning rule.

Boundary Limits
According to Microsoft's resource limits, pools run 1 to 2,048 TiB, with a default regional capacity quota of 100 TiB that support can raise. Volume ceilings by class:
- Regular volumes: 50 GiB to 100 TiB, with 1,600 MiB/s for Standard and 4,500 MiB/s for Flexible, Premium, and Ultra.
- Conventional large volumes: 50 TiB to 1 PiB, up to 12,800 MiB/s.
- Breakthrough-mode large volumes: up to 80 GiB/s, depending on workload and placement, with an access request required.
Microsoft's benchmarks are peak results, not sustained guarantees. A 101-TiB Ultra large volume with twelve clients reached roughly 12,761 MiB/s reads and 8,518 MiB/s writes with 256-KiB sequential I/O. Don't size production against that number.

Change timing rules matter as well. Flexible throughput can rise anytime, but can only drop 24 hours after any throughput change. Fixed-tier volumes can move up a tier with no wait, then need 24 hours before moving down.
Safe Operating Margin
Microsoft states that at the throughput ceiling, I/O queueing and latency can rise rapidly. There is no universal headroom percentage, so set your own from observed peaks.
Leave room for:
- Traffic spikes
- Metadata operations
- Concurrency growth
- Measurement uncertainty
Sizing right at a boundary leaves no room to absorb demand.
How Throughput per TiB Is Specified, Measured, and Validated
Treat the ratio as both a design calculation and an operational check. Label every value as documented, provisioned, benchmarked, or observed.
Specification and Documentation
- Build your model from Microsoft's service-level, resource-limit, and cost-model pages, plus release notes.
- Record the configuration behind each example: capacity, throughput allocation, pool, region, protocol, and workload.
- Keep rated maximums separate from tested values.
Measurement and Verification
Azure Monitor exposes these ANF volume metrics:
ReadThroughput,WriteThroughput,TotalThroughput(bytes/second)ReadIops,WriteIopsAverageReadLatency,AverageWriteLatency(milliseconds, measured inside storage and excluding network latency)ThroughputLimitReached(0 or 1, where 1 means QoS throttling)VolumeAllocatedSizeandVolumeLogicalSizefor provisioned versus used bytes
According to Microsoft's metrics documentation, ANF generally reports five-minute averages, which can mask short spikes. Convert bytes/second to MiB/s, note read versus write direction, and compare percentiles rather than only averages. Service-side latency is not end-to-end latency.

Workload Validation
- Test with production-like traffic: realistic file sizes, access patterns, concurrency, and peak periods.
- Use FIO, vdbench, or iometer to find VM-to-storage limits, but prefer the real application. Microsoft doesn't recommend
ddas a baseline. - Set FIO
randrepeat=0and use direct I/O to avoid cache distortion. - Compare calculated throughput per TiB with observed performance, then check client, VM, network, protocol, and application bottlenecks before blaming ANF.
Teams that also run Azure managed disks alongside ANF can use Lucidity on the block layer. Lumen analyzes IOPS, throughput, latency, and cost across disks, and AutoScaler resizes them with zero downtime.
Implications, Boundaries, and Common Misinterpretations
Running near or past a configuration's range creates latency and throughput problems even when capacity looks sufficient.
Performance and Cost Implications
Too little headroom means queueing and higher latency. Too much means paying for performance the workload never uses. On fixed tiers, billing follows provisioned pool capacity, including space not allocated to volumes.
The Flexible service level changes that trade-off. Microsoft's cost-model example needs 4 TiB of Ultra to deliver 512 MiB/s for 1 TiB of data, versus 1 TiB of Flexible. That removes capacity-driven overprovisioning, but you still size for workload behavior, quotas, latency, and availability. Check current pricing, since it varies by agreement.

Common Misinterpretations
- Treating the ratio as a guarantee. Microsoft calls 4,500 MiB/s an empirical regular-volume ceiling, and adding TiB doesn't scale throughput forever.
- Mixing units and scopes. Comparing MB/s with MiB/s, or a volume with a pool, gives false conclusions.
- Applying benchmark numbers everywhere. Aggregate results don't carry over to every volume, protocol, or region.
- Ignoring other limits. IOPS, latency, client caps, and concurrency often bind first.
- Over-specifying. Buying performance without weighing cost and utilization wastes budget.
Conclusion
Throughput per TiB is a useful governing metric for comparing and sizing ANF configurations. It is derived from capacity, throughput allocation, service-level behavior, and workload conditions.
Reliable design needs current documentation, explicit margins, workload-representative validation, and engineering judgment. No single headline number replaces them.
Frequently Asked Questions
What is throughput per TiB in Azure NetApp Files?
It's throughput divided by the relevant provisioned capacity, such as 64 MiB/s per TiB for Premium. It's a normalized comparison, not automatically a guaranteed application rate.
How is Azure NetApp Files throughput calculated?
With auto QoS, multiply volume quota in TiB by the tier rate. Identify the service level, volume or pool scope, and units, and note whether the value is provisioned or observed.
Does Azure NetApp Files throughput increase automatically when volume capacity increases?
On Standard, Premium, and Ultra with auto QoS, yes, up to volume ceilings. With manual QoS or Flexible, throughput is allocated independently, subject to current platform rules.
What is the difference between throughput, IOPS, and latency in ANF?
Throughput is data moved per second, IOPS is operations per second, and latency is time per operation. I/O size links the first two, and queueing at the ceiling raises latency.
How should I measure actual Azure NetApp Files throughput?
Use Azure Monitor ANF metrics, including ThroughputLimitReached, plus production-like tests. Keep units consistent, review percentiles, and check client-side bottlenecks.
Does the Flexible service level decouple throughput from capacity?
Largely. You buy them separately, but a 128 MiB/s pool baseline, a capacity-bounded maximum of 640 MiB/s per TiB, and a 24-hour decrease cooldown remain. Per TiB stays a handy derived comparison.