Vultr Benchmark Orchestration

⚠️ Fair warning: This was vibe-coded from a plane using OpenClaw. It works, but don't expect enterprise-grade polish. PRs welcome, flames not so much. 🦀✈️

Instances are created with a model list baked into their userdata, run benchmarks autonomously, and self-destruct when done. Your laptop only needs to be running long enough to fire the vultr instance create calls (~10 seconds).

How It Works

Your laptop                    Vultr API              Vultr Instance
    |                              |                        |
    |-- instance create x N ------>|                        |
    |   (userdata = model list)    |                        |
    |<-- instance IDs -------------|                        |
    |  [laptop can close now]      |                        |
                                   |-- boots from snapshot ->|
                                                            |-- reads models from
                                                            |   metadata API
                                                            |-- uv run benchmark.py --register
                                                            |-- [Slack: started + claim URL]
                                                            |-- uv run benchmark.py --model ...
                                                            |-- uv run benchmark.py --model ...
                                                            |-- [Slack: done + result URLs]
                                                            |-- vultr instance delete $SELF

Each instance also schedules a safety-net self-destruct via at now + 5 hours at startup, so orphaned instances are cleaned up even if the runner crashes.

Running Benchmarks

uv run orchestrate_vultr.py --count 10

# Or override with an explicit list
uv run orchestrate_vultr.py --count 10 \
  --models \
  anthropic/claude-opus-4.5 \
  openai/gpt-4o \
  google/gemini-2.5-flash \
  ...

Models are distributed round-robin across instances (e.g. 30 models across 10 instances = 3 models per instance). The script exits as soon as all instances are created.

Options:

Option	Default	Description
--models	(optional)	Model IDs to benchmark (space-separated)
--models-file	default-models.yml	YAML file used when --models is not provided
--count	1	Number of instances; models distributed across them
--region	atl	Vultr region
--plan	vc2-1c-2gb	Vultr instance plan
--snapshot	(see VultrConfig in orchestrate_vultr.py)	Vultr snapshot ID — update after re-bootstrapping
--ssh-keys	a4b8f6d9-...	Vultr SSH key ID

Monitoring:

# Watch instances disappear as they finish
watch vultr instance list

# Tail logs on a running instance
ssh root@<ip> tail -f /var/log/bench-runner.log

# View systemd service output
ssh root@<ip> journalctl -u bench-runner -f

---

Bootstrapping a New Vultr Snapshot

See docs/bootstrapping-snapshot.md

---

Files

File	Purpose
orchestrate_vultr.py	Fire-and-forget launcher — creates instances and exits
bench_runner.sh	Runs on each instance; reads models from metadata, benchmarks, self-destructs
bench-runner.service	systemd unit that starts bench_runner.sh on first boot
bootstrap_instance.sh	One-shot setup script for building a new snapshot image
setup_snapshot.sh	Lighter alternative to bootstrap — installs just the runner files onto an existing instance
create_instance.sh	Convenience shell script with the full model list pre-filled
delete_instances.sh	Emergency cleanup: delete instances by ID