Self-hosted runners on Google Cloud: speed and cost per job

Speed

1.5×

faster than GitHub-hosted (ubuntu-latest)

5th of 9 by speed

Cost per job

−59%

vs GitHub-hosted (ubuntu-latest)

6th of 9; the same at 1,000 to 100,000 jobs a month

Lowest median time

1 / 12

workloads

flask

Failed runs

1 of 360

measured runs

vite 1 of 30. Counted, not replaced.

Speed on each workload

× faster than GitHub-hosted (ubuntu-latest) on the same workload

Burst: 20 jobs at once

1.7×

cli

1.2×

codex-lint

1.5×

container

1.3×

1.4×

flask

2.1×

guava

1.4×

monorepo

1.4×

pydantic

1.5×

vite

1.5×

vue

1.6×

Configuration

Self-hosted · 2 vCPU / 8 GiB

Measured configuration: Self-hosted: GCE n4d-standard-2 always-on pool (on-demand, U = 50%, all-in)
Size: 2 vCPU / 8 GiB
Label or setup: self-hosted label of the deployment
Price as measured: all-in: what the jobs use in your cloud account, plus fixed monthly costs spread over the volume; measured per job (cost rules)
Notes: All-in (METHOD.md, Cost): run time / U x (VM + boot disk + public IPv4), plus internet egress measured per job (cost_gce.py, Cloud Monitoring sent bytes). No controller host.
Definition: providers/gce-pool.json

Per workload

Green marks the lowest median time on the workload. The burst row covers all 20 jobs.

Workload	Median time	vs GitHub-hosted (ubuntu-latest)	p95 time	Mean cost	vs GitHub-hosted (ubuntu-latest)	Successful runs	Rank
Burst: 20 jobs at once	153 s	1.7x faster	165 s	$0.135	73% lower	30 of 30	3rd
cli	310 s	1.2x faster	326 s	$0.017	56% lower	30 of 30	5th
codex-lint	581 s	1.5x faster	625 s	$0.064	27% lower	30 of 30	6th
container	178 s	1.3x faster	201 s	$0.053	100% higher	30 of 30	3rd
fd	64 s	1.4x faster	72 s	$0.0035	71% lower	30 of 30	3rd
flask	11 s	2.1x faster	32 s	$0.0011	82% lower	30 of 30	1st
guava	820 s	1.4x faster	854 s	$0.045	59% lower	30 of 30	5th
monorepo	278 s	1.4x faster	297 s	$0.016	60% lower	30 of 30	6th
pydantic	119 s	1.5x faster	129 s	$0.0065	67% lower	30 of 30	4th
ts	533 s	1.5x faster	582 s	$0.029	63% lower	30 of 30	6th
vite	401 s	1.5x faster	457 s	$0.023	62% lower	29 of 30	5th
vue	76 s	1.6x faster	109 s	$0.0044	69% lower	30 of 30	4th

Burst queue, p95: 3 s, against 31 s for GitHub-hosted (ubuntu-latest)

Indexes and 95% intervals

Time index: 0.667 (95% interval 0.658 to 0.692)
Cost index: 0.409 (0.386 to 0.440) at 100,000 jobs a month

Compare with…

Head-to-head pages: speed, cost, failures and each workload

Vitko Runners vs Self-hosted runners on Google Cloud Blacksmith vs Self-hosted runners on Google Cloud Ubicloud vs Self-hosted runners on Google Cloud RunsOn vs Self-hosted runners on Google Cloud Self-hosted runners on AWS vs Self-hosted runners on Google Cloud Self-hosted runners on Google Cloud vs GitHub-hosted larger runner (4 vCPU)Self-hosted runners on Google Cloud vs GitHub-hosted (ubuntu-latest)