You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
lab: density, workspaces per host before the first login degrades, nested and metal #273
Context: a cost comparison by host type (dollars per workspace-hour) needs density, and it isn't measured anywhere. Placement and billing use the preset's starting size, not what a workspace grows to (spin docs/ROADMAP.md, "The machine").
Experiment for lab.yml: on one host, start 2 vCPU / 2 GiB workspaces one at a time (and with an idle and a CPU-bound load inside, e.g. a Go build), and record boot:ssh-first p50/p95 after each. Stop when p95 is 25% above the first. Run it on bench (nested, c8i.4xlarge), bench-metal (c8i.metal-48xl) and, when the fleet allows, a default runner type with local NVMe (c8id/m8id).
The result is workspaces per host for each type, which turns a per-hour price into a per-workspace-hour one, and feeds the spot break-even (how many workspaces one interruption moves).
The memory half of this is answered by #278 (boot:density, 8 idle 2 vCPU / 2 GiB machines, labs 44546329726 nested and 44546336558 metal): an idle workspace costs its host about 165-170 MiB, what its guest touched, plus 1-2.5 MiB of kernel, the same nested and on metal.
What's left here is the latency half, which is what decides workspaces per host for a cost model:
boot:ssh-first p95 as N grows, until it's 25% above N=1;
with a CPU-bound load in some of them (a Go build), not only idle for 30 s;
on bench, bench-metal, and a default runner type with local NVMe once the fleet has one (spin-stack/ci#21).
Note for whoever reads the result: spin places by the preset's starting size (2 GiB), not by touched memory, so idle density today is limited by that reservation (spin docs/ROADMAP.md, "The machine"), not by the ~170 MiB.
Activity
aledbf commentedon Nov 19, 2026
The memory half of this is answered by #278 (
boot:density, 8 idle 2 vCPU / 2 GiB machines, labs 44546329726 nested and 44546336558 metal): an idle workspace costs its host about 165-170 MiB, what its guest touched, plus 1-2.5 MiB of kernel, the same nested and on metal.What's left here is the latency half, which is what decides workspaces per host for a cost model:
boot:ssh-firstp95 as N grows, until it's 25% above N=1;bench,bench-metal, and a default runner type with local NVMe once the fleet has one (spin-stack/ci#21).Note for whoever reads the result: spin places by the preset's starting size (2 GiB), not by touched memory, so idle density today is limited by that reservation (spin docs/ROADMAP.md, "The machine"), not by the ~170 MiB.
🤖 Generated with Claude Code