Picking an instance size before you have run the workload is guesswork. The useful approach is to start deliberately small, measure what the application actually does, and resize once. That is faster than deliberating and more accurate than any rule of thumb.
Start from the workload shape
-
CPU-bound
Compilation, transcoding, simulation. More cores help; memory rarely does.
-
Memory-bound
Caches, in-memory analytics, large language runtimes. Memory is the binding constraint and swapping destroys performance.
-
I/O-bound
Databases and queues. Disk latency and throughput matter more than core count, and burst credits mislead.
-
Network-bound
Proxies and streaming. Bandwidth and packet rate are the ceiling.
A method that takes an afternoon
- Pick the smallest instance that plausibly runs the workload.
- Deploy the real application, not a synthetic benchmark.
- Apply load resembling your expected peak, sustained for at least thirty minutes.
- Record CPU, memory, disk latency and network throughput throughout.
- Identify which resource saturated first. That is your constraint.
- Resize along that axis only, and repeat once.
Two iterations almost always land close enough. Further tuning has diminishing returns until the workload itself changes.
Read the numbers carefully
| Observation | Likely meaning |
|---|---|
| CPU pinned at 100% | Genuinely CPU-bound, or a busy-wait bug. |
| Memory high but stable | Fine. Many runtimes hold memory deliberately. |
| Memory climbing steadily | Investigate before resizing; this may be a leak. |
| Low CPU, high latency | Look at disk or a downstream dependency. |
| Good for ten minutes, then poor | Burst capacity exhausted; test longer next time. |
Right-sizing is ongoing
Set a reminder to review utilisation monthly for the first quarter. Applications change shape as features land, and an instance chosen for last quarter’s workload is often either wasteful or straining.
Summary
Start small, load it properly, find the first resource to saturate, and resize along that axis. Measured sizing beats estimated sizing every time, and it costs an afternoon.
