AWS re:Invent 2025 - Managing Amazon EC2 capacity and availability (CMP331)

Managing Amazon EC2 Capacity and Availability

Overview

This presentation covers the challenges and trade-offs involved in managing Amazon EC2 capacity and availability. It explores various usage models, including on-demand instances, capacity reservations, capacity blocks for machine learning, and spot instances. The speakers also introduce AWS's new EC2 Capacity Manager tool to help organizations optimize their capacity utilization.

On-Demand Instances

• On-demand instances provide elasticity and a pay-as-you-go model, making them suitable for stateful workloads with scaling patterns or temporary/short-term needs.
• AWS does extensive forecasting and capacity planning to ensure on-demand availability, but there can still be instances where a specific instance type's pool runs low.
• Flexibility is key - being able to use different instance families, generations, sizes, or CPU manufacturers can help avoid capacity shortages.
• AWS provides tools like EC2 Fleet and Auto Scaling Groups to simplify managing this flexibility.

Capacity Reservations

• On-Demand Capacity Reservations (ODCRs) provide assured capacity for specific instance type and availability zone combinations.
• ODCRs are recommended for workloads that require deterministic performance or are so critical that capacity assurance is essential.
• ODCRs can be created on-demand or for a future date to ensure capacity is available when needed.
• Capacity reservation preferences in Auto Scaling Groups can help optimize ODCR utilization.
• Sharing ODCRs across accounts or with lower-priority workloads can improve efficiency.

Capacity Blocks for Machine Learning

• Capacity blocks provide reserved capacity for 1-26 weeks, primarily targeting GPU-accelerated or other specialized compute instances used for machine learning training.
• These instances are often less flexible due to the need to compile to specific hardware architectures and the large data requirements.
• Capacity blocks integrate with SageMaker to provide predictable access to accelerated compute resources.

Spot Instances

• Spot instances leverage AWS's spare data center capacity, offering up to 90% discounts but with the risk of interruption.
• Key strategies for using spot instances include managing the 2-minute termination notice, being flexible across instance types/sizes/generations, and leveraging spot placement scores.
• Spot placement scores help identify the most stable spot capacity based on your workload requirements.

EC2 Capacity Manager

• EC2 Capacity Manager is a new AWS tool that provides a unified view of an organization's EC2 capacity utilization across all usage models and accounts.
• It helps identify unused capacity reservations, track spot instance interruptions, and optimize capacity allocation.
• Capacity Manager can be set up at the account or organization level, providing visibility and management capabilities for large-scale EC2 deployments.

Key Takeaways

1. There is no one-size-fits-all EC2 usage model - organizations typically leverage a mix to optimize for cost, availability, and performance.
2. Flexibility, through techniques like attribute-based instance selection, is crucial for managing capacity constraints.
3. Capacity reservations can provide assured access but come at a higher cost, while spot instances offer significant discounts with interruption risk.
4. EC2 Capacity Manager simplifies capacity management at scale, enabling organizations to better optimize their EC2 utilization.