Translation Notice
This article was originally written in Chinese and translated into English with the assistance of AI. The translation has been reviewed and edited for clarity, accuracy, and readability. Please refer to the original source where necessary, especially for technical terms, commands, configuration details, and proper nouns.

A Basis for Measuring Cloud Resource Right-Sizing (Downscaling Strategy Design)#

Efficient use of cloud resources has always been a core cloud-platform operations concern. As government, private, industry, and hybrid clouds have expanded, wasted and insufficient resources often coexist. Improving utilization while keeping services stable has therefore become an important cloud-operations topic.

Resource allocation is inherently somewhat subjective and often depends on an operations engineer’s experience. Relying entirely on experience, however, leads to inconsistent standards. A more reasonable approach is to establish a utilization-based evaluation model as an important basis for human judgment.

1. Problems with Traditional Right-Sizing Strategies#

In practice, the most common right-sizing strategies are very simple:

For example:

Utilization peak below 30% → reduce to 1/2 of the original size
Utilization peak below 15% → reduce to 1/4 of the original size

This approach is simple and inexpensive to implement, and it can identify low-utilization resources effectively in many cases.

However, this single-metric strategy has clear weaknesses:

If users know the sizing rules, they may avoid downscaling simply by creating a short period of high load during the measurement window. For example:

Running a temporary stress test
Executing a short batch task
Deliberately causing a CPU spike

This can result in:

Resources remain underutilized for a long time, but a transient peak is misinterpreted as a reason to retain the configuration, causing waste.

Relying on a single peak metric is therefore unreliable; a more robust evaluation method is needed.

2. Multidimensional Utilization Model (Weighted Utilization Model)#

To avoid misjudgment caused by a single metric, introduce a multidimensional utilization model using three metrics commonly adopted by cloud providers:

Primary reference data:

Average utilization (Avg Utilization)
P95 utilization (95th Percentile)
Peak utilization (low weight)

Where:

P90 / P95 / P99 explained:

P90: 90% of the data is below this value

P95: 95% of the data is below this value

P99: 99% of the data is below this value

Percentile metrics effectively filter short-lived anomalies and are more useful than the absolute peak.

Compared with a single peak, P95 better reflects the workload level of real business traffic.

3. Weighted Utilization Calculation#

Calculate overall resource utilization using weights:

W₁ average-utilization weight: 80%
W₂ peak-utilization weight: 20%

Formula: Overall utilization = Avg × W₁ + Peak × W₂

Example data:

CPU	Mem	CPU Avg	CPU Peak	Mem Avg	Mem Peak	Assessment
16C	32G	5.34%	35.34%	15.34%	37.56%	No reduction under the traditional model

Looking only at the peak, CPU Peak is 35%, which may suggest that no reduction is needed.

With the weighted model:

Overall CPU utilization: 5.34 × 0.8 + 35.34 × 0.2 = 11.34%
Overall memory utilization: 15.34 × 0.8 + 37.56 × 0.2 = 19.78%

The result is:

Resource	Overall utilization	Recommendation
CPU	11.34%	Reduce to 1/4
Memory	19.78%	Reduce to 1/2

Final recommendation: 16C32G → 4C16G

Compared with a single-peak decision, this approach better reflects actual resource usage.

4. Handling Abnormal Peaks#

If historical data contains a CPU Peak of 100%:

determine whether it represents:

Sustained high load
An occasional burst
A scheduled task
Stress-testing activity

Additional signals can include:

Peak duration
Number of peak occurrences
The difference between P95 and Peak
Peak frequency

For example:

Peak = 100%
P95 = 18%

This indicates an occasional spike and should not be used as a basis for scaling up.

Peak = 100%
P95 = 75%

This may indicate a genuine lack of resources.

5. A More Reasonable Right-Sizing Recommendation#

When data granularity is insufficient, such as when only monthly statistics are available, use:

Multi-metric scoring: Score = Avg × 0.6 + P95 × 0.3 + Peak × 0.1

Recommended strategy:

Overall utilization	Recommendation
<15%	Reduce to 1/4
15–30%	Reduce to 1/2
30–50%	Keep the current size
>50%	Evaluate scaling up

This approach can:

Avoid human interference
Improve resource utilization
Make allocation more reasonable
Reduce the chance of misjudgment

6. Real-World Limitations#

It is important to understand that cloud-resource allocation can never be fully automated.

Utilization cannot capture factors such as:

An upcoming new service
Business growth trends
Tasks scheduled for specific times
Future capacity planning

For example: Previously: low utilization on 16 cores → possibly little business traffic

Now: high utilization on 4 cores → possibly growing business demand

Algorithms cannot fully judge such situations. Therefore, automated evaluation can only provide recommendations, not final decisions.

7. Conclusion#

Cloud-resource right-sizing is fundamentally a combination of data models and experience-based judgment.

The automated model provides:

Utilization analysis
Right-sizing recommendations
Risk alerts

People provide:

Business judgment
Risk assessment
The final decision

The goal is not to compress resources to the absolute minimum

but to maximize resource utilization within a safe operating range.

That is the true goal of cloud-operations optimization.

Music