Skip to content

cold-baseline-results.md

Latest commit

 

History

History
75 lines (57 loc) · 2.38 KB

cold-baseline-results.md

File metadata and controls

75 lines (57 loc) · 2.38 KB

Cold Launch Baseline Results

Date: 2026-02-01 00:53:43 UTC Instance Type: m7i.large AMI: ami-0fcee47b1475c1af3 (Amazon Linux 2023) Technique: cloud-init (default)

Summary

Phase Description Average Notes
T0→T1 API call 4797ms RunInstances returns
T1→T2 Pending→Running 17811ms Hypervisor/scheduler
T2→T4 Boot→SSH ready 3217ms Kernel + cloud-init + sshd
Total End-to-end 25825ms (25.8s)

Individual Runs

Run API Pending Boot→SSH Total
1 4626ms 17779ms 3250ms 25655ms
2 4800ms 17805ms 3200ms 25805ms
3 4965ms 17851ms 3201ms 26017ms

Analysis

Where Does the Time Go?

  1. API Call (T0→T1): ~4797ms

    • AWS receives request, validates, schedules
    • Returns instance ID immediately

  2. Pending→Running (T1→T2): ~17811ms

    • Instance scheduled to hypervisor
    • ENI attached, EBS attached
    • Hypervisor starts guest

  3. Boot→SSH (T2→T4): ~3217ms

    • Linux kernel boot
    • systemd initialization
    • cloud-init runs (fetches metadata, injects SSH keys)
    • sshd starts and accepts connections

Where the Time Goes (Breakdown)

Phase Time % of Total
API 4.8s 19%
Pending→Running 17.8s 69%
Boot→SSH 3.2s 12%

Key Finding: Pending→Running dominates at 69%. This is AWS infrastructure time (hypervisor scheduling, ENI/EBS attachment). We have limited control here.

Optimization Opportunities

  • API (~5s): Minor. Could pre-warm with describe calls but won't help much.
  • Pending→Running (~18s): AWS infrastructure time - largely outside our control. Possible improvements:
    • Warm pools (out of scope)
    • Pre-allocated ENI (out of scope)
    • Better instance type placement
  • Boot→SSH (~3s): Small but fully controllable:
    • Skip cloud-init (bake keys into AMI)
    • Minimal systemd or custom init
    • Faster sshd startup

Realistic target: Even with perfect boot optimization (0s boot), we'd still be at ~22s due to AWS overhead. The 5-10s target from specs/requirements.md may be unrealistic without warm pools or pre-allocation.

Raw Data

API_TIMES: 4626 4800 4965
RUNNING_TIMES: 17779 17805 17851
SSH_TIMES: 3250 3200 3201
TOTAL_TIMES: 25655 25805 26017