Skip to content

start-stopped-results.md

Latest commit

 

History

History
57 lines (43 loc) · 1.67 KB

start-stopped-results.md

File metadata and controls

57 lines (43 loc) · 1.67 KB

Start-from-Stopped Results

Date: 2026-01-31 Instance Type: m7i.large AMI: Amazon Linux 2023 (same as cold baseline) Technique: Start pre-stopped instance (EBS/ENI already attached)

Summary

Phase Cold Baseline Start-Stopped Difference
API call 4797ms 1764ms -3033ms
Pending→Running 17811ms 16372ms -1439ms
Boot→SSH 3217ms 3042ms -175ms
Total 25825ms 21178ms -4647ms (18%)

Key Finding

Starting a stopped instance only saves ~4.6s (18%) vs cold launch.

The Pending→Running phase is nearly identical:

  • Cold: 17.8s
  • Start-stopped: 16.4s

This means the 69% of time spent in Pending→Running is NOT primarily:

  • ❌ EBS volume creation/attachment (already attached)
  • ❌ ENI creation/attachment (already attached)
  • ❌ Hypervisor placement/scheduling (already placed)

It IS likely:

  • ✅ Hypervisor cold-start of the guest VM
  • ✅ Hardware/firmware initialization
  • ✅ Network path establishment (even with ENI attached)

Implications

  1. Warm pools won't help much - Pre-provisioned stopped instances still take ~21s
  2. The bottleneck is deeper - Inside AWS infrastructure, not resource allocation
  3. Hibernate might help - Skips guest cold-start entirely
  4. Instance store might help - Different storage path

Raw Data

=== START-FROM-STOPPED ===
API call:     1764ms
Running:      18136ms (+16372ms)
SSH ready:    21178ms (+3042ms)
---
TOTAL:        21178ms

Next Steps

  • Test hibernate resume (skips kernel boot)
  • Test instance-store backed instances
  • Test different instance types (smaller might be faster)