Virtual Machines

How virtual machine and scale set recommendations are sized, what the VM toggles do, and how to read the VM card. Select a VM or scale set on the Recommendations Page to open its card.

Azure VMs and VM scale sets with collected performance metrics are analyzed, producing VM SKU Downsize and VMSS SKU Downsize recommendations. Analysis uses daily Azure Monitor metrics kept for up to 365 days, plus 30 days of hourly CPU. Each VM needs 7 days of data before its first recommendation by default; set the Virtual Machines Minimum Data Window to 0 to include every VM.

Statistics Used

CPU

The higher of the P95 of daily averages (up to 365 days) and the P95 of hourly averages (last 30 days). Hourly is used only when at least 24 hours of hourly data exist.

Memory

From the Azure Monitor Available Memory Bytes platform metric; no agent is needed. The higher of an average-based statistic (P99 with 100+ days of data, otherwise P95) and the P95 of daily peak usage (needs 7+ days).

Disk IOPS

VM-level uncached and cached IOPS-consumed P95s when available, otherwise the OS and data disk IOPS-consumed percentage.

When a VM doesn't report Available Memory Bytes, memory shows "Data not available", the recommended SKU keeps at least the current memory, and only CPU-based downsizing is possible.

Workload Profile

Each VM is classified from its CPU, memory and disk pressure. The profile decides which families are considered when cross-family SKUs are allowed.

Compute Intensive

F-series preferred

Memory Intensive

E-series preferred

Balanced

D-series preferred

Burstable

B-series preferred

Oversized

Consistently low on everything

Candidate Checks

A candidate SKU that fails any of these checks is never recommended:

• Enough vCPU and memory for the workload at the aggressiveness ceiling.
• Enough uncached and cached IOPS capacity.
• Supports Premium storage and accelerated networking when the VM uses them, and at least as many data disks as are attached.
• Has a price at the VM's pricing tier. A Spot-only price doesn't qualify for a regular VM.
• B-series only when Allow B-series (Burstable) is on, and only when the SKU's baseline CPU can sustain the VM's average CPU load. B-series SKUs with no known baseline are skipped.

VM Toggles

These five toggles live in the Virtual Machines section of Toggle Options. All are off by default, which keeps recommendations in the same family and current generation.

Allow Feature Changes

Allows CPU architecture and storage feature changes, such as AMD to ARM. When off, SKU sub-type suffixes like "as", "pls" and "ds" are preserved.

Allow Cross-Family SKUs

Allows moving to another VM family, for example D-series to E-series for a memory-heavy workload.

Allow Older Generations

Allows older SKU generations, such as v5 to v4, for more savings. Independent of cross-family.

Allow B-series (Burstable)

Allows burstable B-series SKUs.

Exclude Restricted SKUs

Skips SKUs with zone or location restrictions for the VM's subscription and shows the cheapest unrestricted option. When off, a restricted SKU can be recommended and the card shows Availability Restrictions.

Reading the VM Card

The VM SKU Downsize Recommendation card is used for both VMs and scale sets. Alongside the badges, rationale and estimated savings described in How Sizing Works, it shows:

Slot badge

Which toggle produced the pick: Cross-Family, Older Generation, Burstable, Feature Change, a combination such as Cross-Family + Older Gen, or Multi-Toggle. No badge means the default alternative.

Current and Recommended

SKU, vCPU, memory (GB) and monthly cost for each.

Pricing based on

The OS (Windows or Linux) and priority (Spot or Pay-as-you-go) detected from the VM. Candidates are priced and ranked at this pricing tier.

All Pricing Options

Per-instance current, recommended and savings prices for Linux and Windows, Pay-as-you-go and Spot.

Availability Restrictions

Shown only when the recommended SKU is restricted for this subscription, for example in one zone.

Utilization Statistics

P95, Average and Maximum for CPU and memory over the analysis window. P95 (Hourly) appears only when hourly CPU data exists.

Workload Analysis

The workload profile, family, analysis days and CPU, memory and disk pressure. An info icon next to Family appears when the workload's suggested family differs from the current one; it names the toggle to enable.

Scale Sets

Scale set cost and savings cover the whole scale set: the per-instance price times the current instance count. Realized savings are recorded on a SKU change and compared at today's instance count, so instance-count changes alone don't count as savings.

Confidence

VM confidence uses five factors. Levels and thresholds are described in How Sizing Works.

Data Sufficiency

30%

Workload Stability

25%. Highly bursty intraday workloads, with hourly peaks far above daily averages, score lower and get a consideration.

Resource Headroom

20%

Workload Pattern

10%

Memory Data Availability

15%. Without memory data, confidence is capped at High.

Troubleshooting

My VM has run for a week and has no recommendation

Memory says Data not available

The recommended SKU shows Availability Restrictions

Workload Analysis suggests another family but the recommendation stays in the same family

My deallocated VM's recommendation disappeared