# Resource Utilization

Resource utilization is the amount or proportion of available computing capacity that a workload uses over a given period.

Cloud concept. Updated 2026-10-29. Page: https://blowstack.com/tools/cloud-concepts-hub/resource-utilization

> Resource utilization shows how workloads consume CPU, memory, storage, network bandwidth, and other capacity. It informs rightsizing: selecting resource types, sizes, and quantities that meet performance and availability requirements without paying for unnecessary capacity.

Practice exams covering Resource Utilization: [CLF-C02 AWS Certified Cloud Practitioner](https://blowstack.com/practice-exams/aws-certified-cloud-practitioner), [SOA-C03 AWS Certified CloudOps Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-cloudops-engineer-associate), [DEA-C01 AWS Certified Data Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-data-engineer-associate), [DOP-C02 AWS Certified DevOps Engineer - Professional](https://blowstack.com/practice-exams/aws-certified-devops-engineer-professional), [AIP-C01 AWS Certified Generative AI Developer - Professional](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional), [MLA-C02 (beta) AWS Certified Machine Learning Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate), [SAA-C03 AWS Certified Solutions Architect - Associate](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate), [SAP-C03 AWS Certified Solutions Architect - Professional](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional).

## 6 key aspects of Resource Utilization

Effective utilization management combines representative measurements, workload-aware rightsizing, and sufficient headroom to control costs without compromising service requirements.

1. **Multidimensional measurement**: Evaluate CPU, memory, storage capacity, storage performance, and network demand separately. Low utilization in one dimension does not rule out a bottleneck in another.
2. **Evidence-based rightsizing**: Select resource types, sizes, and quantities using observed demand and application requirements. Rightsizing can mean reducing capacity, increasing it, or changing the resource configuration.
3. **Representative observation periods**: Include normal operation, traffic peaks, batch jobs, and seasonal events when assessing demand. Averages alone can hide short periods of saturation.
4. **Performance and resilience headroom**: Preserve enough spare capacity for bursts, scaling delays, maintenance, and failures. Appropriate utilization depends on service objectives rather than a universal target percentage.
5. **Coordination with elasticity**: Combine an appropriately sized baseline with demand-driven capacity changes where the application supports them. Validate scaling thresholds and startup times against real workload behavior.
6. **Continuous economic validation**: Reassess utilization after deployments and demand changes, and verify savings against actual billing. Treat purchasing discounts as separate from eliminating unnecessary capacity.

## Cloud services for Resource Utilization

Amazon CloudWatch supplies utilization and application telemetry, while AWS Compute Optimizer uses supported metrics and configurations to identify potential improvements. Engineers can implement validated changes through Amazon EC2 instance selection and use Amazon EC2 Auto Scaling to adjust fleet size as demand changes. AWS Cost Explorer helps compare spending before and after changes. Together, these services support a recurring process of measuring demand, selecting suitable capacity, preserving performance headroom, and checking financial outcomes. Recommendations and scaling policies still require workload-specific validation, especially for memory-intensive applications, irregular processing cycles, and strict availability requirements.

- **[Amazon CloudWatch](https://blowstack.com/tools/cloud-services-hub/amazon-cloudwatch)**: Collects infrastructure and application metrics used to understand utilization over time and configure alarms. Operating system memory metrics for EC2 instances typically require the CloudWatch agent or another custom metric publisher.
- **[AWS Compute Optimizer](https://blowstack.com/tools/cloud-services-hub/aws-compute-optimizer)**: Analyzes supported resource configurations and utilization metrics to provide optimization recommendations. Engineers should validate recommendations against workload cycles, performance requirements, and metrics that may not be available to the service.
- **[Amazon EC2](https://blowstack.com/tools/cloud-services-hub/amazon-ec2)**: Provides instance types and sizes with different CPU, memory, networking, and storage characteristics. Rightsizing can involve changing an instance's size or family, subject to compatibility and operational constraints.
- **[Amazon EC2 Auto Scaling](https://blowstack.com/tools/cloud-services-hub/amazon-ec2-auto-scaling)**: Adjusts the number of EC2 instances according to configured scaling policies and capacity limits. Target tracking policies can use metrics such as average CPU utilization to align fleet capacity with changing demand.
- **[AWS Cost Explorer](https://blowstack.com/tools/cloud-services-hub/aws-cost-explorer)**: Helps engineers examine cost and usage trends to assess the financial effects of rightsizing. Cost data complements utilization telemetry but does not independently establish whether an application has enough capacity.

## Resource Utilization exam questions

Real practice exam questions on Resource Utilization, with the correct answer and its explanation.

### Question 1 (CLF-C02)

**Question:**

A company runs a customer portal on two large Amazon EC2 instances in different Availability Zones. The portal must remain available if either Availability Zone fails. After a software improvement, testing confirms that one smaller instance can handle the portal's entire peak workload while meeting response-time requirements. The smaller instance has a lower On-Demand hourly price. Which change will reduce compute costs while preserving the availability requirement?

**Correct answer:** Replace both large instances with smaller instances, keeping one in each Availability Zone.

**Explanation:** Each smaller instance can handle the full peak workload, including when the other Availability Zone is unavailable. Reducing both instance sizes eliminates unnecessary capacity while preserving availability.

### Question 2 (MLA-C02 (beta))

**Question:**

A company hosts a conversational agent on Amazon Bedrock AgentCore Runtime. Users reuse the same runtime session for many conversation turns. Memory consumption rises during the business day, but model token usage per turn remains stable. Engineers need to determine whether individual sessions retain progressively more memory between turns or whether the increase is explained only by a larger number of concurrent sessions. They can add application instrumentation but must not change production session behavior.

Which monitoring approach best meets these requirements?

**Correct answer:** Emit structured memory measurements at consistent, quiescent points between turns, with session ID, turn number, timestamp, and in-flight operation count. Use CloudWatch Logs Insights to compare memory trends within individual sessions and correlate them with concurrent session counts.

**Explanation:** Session-level measurements taken when no operations are in flight distinguish retained memory from temporary processing demand. Comparing successive measurements within the same session reveals growth across turns, while concurrent-session counts help distinguish that pattern from an increase in the number of sessions. Session identifiers can remain log fields rather than creating a custom metric time series for every session.

### Question 3 (SAP-C03)

**Question:**

A company runs a latency-sensitive application as an Amazon ECS service on AWS Fargate. The service maintains six Linux tasks across three Availability Zones. Each task is configured with 4 vCPUs and 16 GiB of memory and contains an application container and a telemetry sidecar. The company must retain all six tasks to meet connection-distribution and availability requirements.

Monitoring over a complete business cycle, including peak processing, shows the following maximum resource consumption per task. The application and sidecar peaks occur simultaneously:

Container | CPU consumed | Memory consumed
Application | 0.9 vCPU | 7.5 GiB
Telemetry sidecar | 0.3 vCPU | 1.5 GiB

Load tests confirm that CPU and memory are the only capacity constraints. Company policy requires provisioned task CPU and memory to each be at least 1.25 times the corresponding maximum observed consumption. Tasks must use On-Demand capacity, and changes can be applied through a rolling deployment.

For the selected platform, valid Fargate task sizes include 1 vCPU with 2–8 GiB of memory, 2 vCPUs with 4–16 GiB, and 4 vCPUs with 8–30 GiB, in 1-GiB memory increments. CPU and memory are billed separately according to the task configuration.

Which change will minimize Fargate compute charges while meeting these requirements?

**Correct answer:** Configure each task with 2 vCPUs and 12 GiB of memory. Adjust container resource settings to fit within the new task configuration while preserving the required headroom.

**Explanation:** Combined maximum demand is 1.2 vCPUs and 9 GiB per task. Applying the required multiplier gives minimum capacities of 1.5 vCPUs and 11.25 GiB. The least expensive valid configuration that satisfies both dimensions is 2 vCPUs and 12 GiB. Reducing the task-level configuration reduces billed capacity while retaining all six tasks.

### Question 4 (SOA-C03)

**Question:**

A company runs a Linux connection-broker application on one Amazon EC2 instance that uses the Elastic Network Adapter (ENA). As the number of concurrent, long-lived TCP connections grows, new connections intermittently fail. ENA metrics published to Amazon CloudWatch show that conntrack_allowance_available falls to zero and conntrack_allowance_exceeded increases during the failures. CPU, memory, packet rate, and network bandwidth remain well below their limits.

The application must remain on one instance and must retain all legitimate connections with its current connection timeout settings. Which action should a CloudOps engineer take to resolve the bottleneck?

**Correct answer:** Replace the instance with an instance type whose documented connection-tracking allowance accommodates the peak concurrent connections with additional headroom.

**Explanation:** A zero available connection-tracking allowance accompanied by increasing allowance-exceeded metrics indicates that the instance cannot track additional connections. Selecting an instance type with sufficient connection-tracking capacity addresses this limit without dropping legitimate connections or changing timeout settings.

### Question 5 (AIP-C01)

**Question:**

A company serves a containerized LLM on Amazon EC2 instances, each with one GPU. The service meets its time-to-first-token requirement, but output tokens per second are insufficient during sustained peak traffic. The company must increase generation throughput without increasing the hourly fleet cost or changing model precision, context length, or response content.

Profiling during peak traffic shows:
• GPU compute utilization averages 28%.
• GPU device-memory bandwidth utilization remains above 95% during decoding.
• Model weights and peak inference allocations use 40 GiB of the GPU's 48 GiB of memory.
• Host CPU utilization remains below 25%, with no CPU throttling or storage or network bottlenecks.

The inference engine already uses continuous batching. Tests show that increasing active sequences further violates the latency requirement. The developer is evaluating replacement instance configurations with comparable hourly prices.

Which deployment change should the developer prioritize for representative load testing?

**Correct answer:** Select an instance with sufficient GPU memory for the measured peak allocation, substantially higher GPU memory bandwidth, and comparable GPU arithmetic throughput.

**Explanation:** The GPU has sufficient memory capacity, but decoding nearly saturates its memory bandwidth while compute utilization remains low. Prioritizing a higher-bandwidth GPU targets the constrained resource. Representative load testing must confirm token throughput, latency, memory headroom, and cost before production rollout.

### Question 6 (DEA-C01)

**Question:**

An AWS Glue Spark job uses a DynamicFrame JDBC reader to extract a large PostgreSQL table before transforming the data. The extraction stage accounts for most of the job's runtime. The Spark UI shows one JDBC input task running while the other executors remain idle. The database has sufficient capacity to handle additional concurrent queries. The table contains a numeric transaction_id column with evenly distributed values.

Which change should a data engineer make to parallelize the source extraction and reduce the job's runtime?

**Correct answer:** Configure the JDBC reader with hashfield set to transaction_id and hashpartitions set to an appropriate value, such as 16.

**Explanation:** AWS Glue supports parallel JDBC reads through hashfield and hashpartitions. These options divide the source data among concurrent read queries. An evenly distributed hash field helps balance the work, allowing the job to use its idle executors and the database's available capacity.

### Question 7 (DOP-C02)

**Question:**

A company runs an Amazon ECS service on EC2 instances supplied by an Auto Scaling group capacity provider. Managed scaling is enabled with a target capacity of 100%. The service uses only this capacity provider. The Auto Scaling group has a desired capacity of 4 and a maximum capacity of 20. Its mixed instances policy allows m5.large instances (2 vCPUs, 8 GiB of memory) and m5.2xlarge instances (8 vCPUs, 32 GiB of memory), without instance weights.

Each task requires 4 vCPUs and 16 GiB of memory. The four existing instances are m5.2xlarge instances and have no remaining capacity for additional tasks. After the service's desired count increases, additional tasks remain in PROVISIONING. The capacity provider's CapacityProviderReservation metric remains at its target, and no scale-out activity occurs. Subnet addresses, service quotas, and instance availability are not limiting factors.

The task resource requirements cannot be reduced. A DevOps engineer must correct the configuration so that managed scaling automatically provisions capacity for these tasks.

Which change meets this requirement?

**Correct answer:** Remove m5.large from the Auto Scaling group's allowed instance types, and ensure that every remaining allowed instance type can accommodate at least one task.

**Explanation:** ECS managed scaling checks whether provisioning tasks can fit on the smallest instance type allowed by the Auto Scaling group. Tasks that cannot fit are excluded from the scale-out calculation and can remain in PROVISIONING without triggering scaling. The m5.large cannot accommodate this task. Allowing only compatible instance types removes this constraint so that managed scaling can provision the required capacity.

### Question 8 (SAA-C03)

**Question:**

A company runs two EC2-based network appliances, one in each of two Availability Zones. Both appliances must remain running, and either appliance must be able to handle all traffic if the other Availability Zone fails. The appliances currently use large, general-purpose instances.

Amazon CloudWatch metrics and guest operating system metrics collected over a complete monthly workload cycle show that, when one appliance handles all traffic, it requires at most 3 vCPUs and 10 GiB of memory. Network throughput reaches 4 Gbps continuously during a 3-hour nightly transfer. The company has identified two less expensive instance configurations:

• Configuration A: 4 vCPUs, 16 GiB of memory, 1 Gbps baseline network bandwidth, and burst bandwidth of up to 10 Gbps. Testing shows that burst bandwidth lasts less than 1 hour during this workload.
• Configuration B: 4 vCPUs, 16 GiB of memory, and 5 Gbps of sustained network bandwidth. This configuration costs more than Configuration A but less than the current instances.

Testing confirms that both configurations meet all other appliance requirements. Which change will reduce costs while preserving performance and Availability Zone resilience?

**Correct answer:** Replace both current instances with Configuration B, retaining one appliance in each Availability Zone.

**Explanation:** Configuration B provides sufficient CPU, memory, and sustained network bandwidth for either appliance to handle the entire workload. Replacing both instances reduces costs while retaining the required two-Availability-Zone architecture.

## Practice Resource Utilization in a full exam

Questions on Resource Utilization in each practice exam, with every answer option and an explanation for each:

- [SAA-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate): 22 questions
- [DOP-C02](https://blowstack.com/practice-exams/aws-certified-devops-engineer-professional): 14 questions
- [MLA-C02 (beta)](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate): 11 questions
- [CLF-C02](https://blowstack.com/practice-exams/aws-certified-cloud-practitioner): 7 questions
- [SOA-C03](https://blowstack.com/practice-exams/aws-certified-cloudops-engineer-associate): 6 questions
- [AIP-C01](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional): 6 questions
- [DEA-C01](https://blowstack.com/practice-exams/aws-certified-data-engineer-associate): 2 questions
- [SAP-C03](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional): 1 question

## What is Resource Utilization

Resource utilization describes how much available capacity a workload actually uses. For example, a virtual machine might average 15 percent CPU utilization while using 85 percent of its memory. Utilization must be evaluated separately for each relevant resource because a workload can have spare processing capacity but still be constrained by memory, storage throughput, or network bandwidth. There is no single utilization percentage that accurately represents every aspect of an application.

In cloud economics, utilization is an important input to rightsizing. Rightsizing means matching a resource's type, size, and quantity to the workload's requirements. It can involve selecting a smaller virtual machine, moving to a memory-optimized configuration, increasing capacity to remove a bottleneck, or removing resources that are no longer needed. The objective is not simply to choose the cheapest option; it is to meet required performance and reliability at an appropriate cost.

Consider an application running on a virtual machine with eight virtual CPUs and 32 GiB of memory. If representative monitoring shows low CPU demand and a memory working set of 6 GiB, a smaller configuration may be suitable. However, an engineer should also examine peak traffic, scheduled jobs, storage performance, and response times before changing it. A smaller instance may have lower network or storage throughput limits even when its CPU and memory appear sufficient.

Measurements should cover a representative workload cycle. An hourly average can conceal brief saturation, while a quiet week may exclude month-end processing. Review utilization trends alongside latency, error rates, queue depth, and throughput. Where available, use higher-resolution metrics and percentile-based application measurements to understand bursts. Guest operating system memory metrics may require an agent or another telemetry mechanism rather than appearing automatically in basic infrastructure monitoring.

Rightsizing and elasticity are related but different. Rightsizing establishes an appropriate resource configuration; elasticity adjusts capacity as demand changes. An application might use smaller instances as its baseline and add instances during busy periods. Scaling policies need startup time, sensible thresholds, and enough spare capacity to absorb demand while new resources become ready. Some applications cannot scale horizontally without architectural changes.

High utilization is not automatically desirable. Running continuously near a resource limit can increase latency, cause throttling, or leave insufficient capacity during a failure. Conversely, low utilization is not always waste: redundant systems and disaster recovery capacity may intentionally remain lightly used. Assess efficiency against service objectives and resilience requirements rather than pursuing a universal utilization target.

For certification questions, distinguish rightsizing from purchasing discounts. A reservation or savings commitment can reduce the price of eligible usage, but it does not make an oversized resource appropriately sized. Likewise, low CPU utilization alone does not prove that an instance should be reduced. Measure multiple resource dimensions, make a controlled change, validate application behavior, and repeat the assessment as demand evolves. Stopping unused resources can reduce charges, but retained storage and other associated resources may continue to incur costs.

## Learn Resource Utilization in lessons

9 short lessons on Resource Utilization in 8 certification courses.

- CLF-C02: Resource Utilization: interpreting capacity use (Premium), https://app.blowstack.com/lessons/resource-utilization-interpreting-capacity-use-clf-c02
- MLA-C02 (beta): Resource Utilization: bottleneck-aware scaling signals (Premium), https://app.blowstack.com/lessons/resource-utilization-bottleneck-aware-scaling-signals-mla-c02
- SAP-C03: Resource Utilization: interpreting workload demand profiles (Premium), https://app.blowstack.com/lessons/resource-utilization-interpreting-workload-demand-profiles-sap-c03
- SOA-C03: Resource Utilization: interpreting compute capacity (Premium), https://app.blowstack.com/lessons/resource-utilization-interpreting-compute-capacity-soa-c03
- AIP-C01: Resource Utilization: LLM inference bottlenecks (Premium), https://app.blowstack.com/lessons/resource-utilization-llm-inference-bottlenecks-aip-c01
- DEA-C01: Resource Utilization: finding capacity bottlenecks (Premium), https://app.blowstack.com/lessons/resource-utilization-finding-capacity-bottlenecks-dea-c01
