Design High Performing Architectures · Topic 2 of 7
Elastic Compute Options
Select EC2 instance families, placement groups, Graviton, Lambda and container platforms for performance, and match each compute choice to the workload.
- 5 min
- Match the workload to the family: C for compute, R and X for memory, I for storage I/O, P and G for GPU work.
- T instances are burstable and not designed for sustained high CPU.
- Graviton gives good price performance when the software supports Arm.
What it is
High performing compute means choosing a platform and size that fit the workload, then letting it scale with demand. AWS gives you virtual servers (EC2), functions (Lambda), containers (ECS, EKS, Fargate) and batch or high performance computing services.
The exam rarely asks for exact instance sizes. It asks you to match a description such as "memory hungry in memory cache" or "tightly coupled simulation" to a family or feature.
How it works
EC2 instance families
| Family type | Examples | Best for |
|---|---|---|
| General purpose | M, T | Balanced CPU and memory, web and application servers |
| Compute optimized | C | CPU bound work such as batch processing, gaming servers, media encoding |
| Memory optimized | R, X | In memory databases and caches, large analytics datasets |
| Storage optimized | I, D | High local disk I/O, such as NoSQL databases and data warehouses |
| Accelerated computing | P, G, plus AWS Inferentia and Trainium | Graphics, machine learning training and inference |
T family instances are burstable. They run at a baseline and earn CPU credits while idle, then spend them in bursts, so they suit light and spiky workloads but not constant high load.
Instances built on AWS Graviton (Arm based) processors often give better price performance for supported software, since the software must run on Arm. Enhanced networking, and the Elastic Fabric Adapter (EFA) for HPC and machine learning, give lower latency and higher network speed between nodes.
Placement groups
A placement group controls where instances sit on the underlying hardware.
| Strategy | Effect | Use it for |
|---|---|---|
| Cluster | Packs instances close together in one AZ | Lowest latency and highest network throughput, such as HPC |
| Spread | Puts each instance on separate hardware | A small number of critical instances that must not fail together |
| Partition | Groups instances into partitions on separate racks | Large distributed systems such as Hadoop, Cassandra and Kafka |
A cluster group improves performance but concentrates risk, because one hardware problem can hit all the instances. A spread group improves availability.
Elasticity
Elasticity means capacity follows demand. EC2 Auto Scaling does this for instances, and Application Auto Scaling does it for other services. Choose a scaling metric that reflects real load, such as requests per target or queue depth, instead of a metric that may not move.
Serverless platforms scale without an explicit scaling setup. Lambda scales by running more copies of your function, and Fargate runs the tasks you ask for without hosts to manage.
Other compute services
- AWS Batch: runs batch computing jobs by provisioning the right amount of compute for queued jobs, including on Spot Instances.
- AWS Elastic Beanstalk: deploys web applications from code, and manages the underlying EC2, load balancer and Auto Scaling for you.
- Amazon Lightsail: simple virtual servers for small, predictable projects.
- AWS Outposts: AWS infrastructure running in your own data center for low latency to on premises systems or local data needs.
- AWS Wavelength and Local Zones: place compute closer to mobile users or specific cities for very low latency.
Tuning Lambda
Lambda memory also sets CPU power, so raising memory can shorten the run time. A function in a VPC reaches private resources, but it needs a NAT gateway or VPC endpoint to reach the internet or AWS APIs. Use provisioned concurrency to reduce cold starts for latency sensitive APIs.
Choosing a platform
Pick Lambda for short, event driven tasks. Pick containers for long running services or when the team already ships container images. Pick EC2 when you need full control of the operating system, special hardware, a licensed product, or a persistent process that does not fit the others.
A phrase like "tightly coupled nodes needing the lowest possible latency" means a cluster placement group, often with EFA. "Critical instances isolated from each other's hardware failure" means a spread placement group.
What to remember for the exam
- Match the workload to the family: C for compute, R and X for memory, I for storage I/O, P and G for GPU work.
- T instances are burstable and not designed for sustained high CPU.
- Graviton gives good price performance when the software supports Arm.
- Cluster placement groups reduce latency, spread groups reduce correlated failure, and partition groups suit large distributed stores.
- EFA is for HPC and machine learning traffic between instances.
- Lambda memory setting also sets CPU, and Lambda has a 15 minute limit.
- AWS Batch runs queued batch jobs on managed compute, and Beanstalk deploys web apps with less setup.
- Scale on a metric that follows real load, and span multiple AZs.
Common mistakes
Picking a T instance for steady heavy load. After the credits run out, performance drops to the baseline (or extra charges apply in unlimited mode), so use an M or C instance instead.
Using a cluster group for high availability. It does the opposite, because the instances share hardware locality.
Forgetting that placement groups are mostly one AZ. A cluster group lives in a single AZ, so it needs a separate plan for AZ failure.
Choosing GPU instances for ordinary web work. Accelerated families cost more and help only with GPU workloads.
A cluster placement group can fail to launch when capacity is short. Launch all instances at once and use one instance type to improve the chance of success.
Check yourself
Which placement group gives the lowest network latency between instances?
A cluster placement group.
Which instance family fits an in memory cache with a large dataset?
A memory optimized family such as R or X.
Why are T family instances a poor choice for constant high CPU use?
They are burstable, and once the CPU credits are used up they fall back to the baseline performance, or cost extra in unlimited mode.
Topic 2 of 7 in Design High Performing Architectures
Test yourself on Design High Performing Architectures
The full AWS SAA-C03 question bank has 823 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.