perf(infra): host-cache Tart VM disk reads (caching=cached)

What

Pass --root-disk-opts caching=cached to tart run for the ephemeral build VMs (tart-kubelet, internal/tart/tart.go) so the host page-caches the VM disk-image reads.

Why

The runner Mac minis (Scaleway M2-L = M2 Pro 10C / 16 GB) trail Namespace by ~60% on CAS-heavy Xcode builds. The investigation behind runners-benchmark#1 traced the gap to runner-side per-task overhead — both runners do identical work (5343 SwiftCompile + ~1600 CAS materializations, 0 cache errors), but tuist-macos is ~1.6× slower per task. Ruled out, with measurements:

• not the cache — the Kura node serves CAS reads at <1 ms p90;
• not the network — PN RTT to the node is ~2 ms (lower than Namespace’s WAN, yet we’re slower);
• not hit-rate — identical task/materialization counts, 0 errors;
• not CPU — ~even.

What’s left is the local read / page-in path: the compiler + Swift dylibs are paged in from the Tart sparse disk image on every one of thousands of short tasks, and our disk reads were slower/variable vs Namespace’s.

Tart’s default disk caching is automatic. An on-host A/B on an actual runner host (guest held constant) measured warm reads:

So the default leaves read throughput on the table, and caching=cached recovers ~1.8× — keeping the VM, no hardware change.

Why not also sync=none?

Evaluated and ruled out for this workload: the build issues no hard syncs (zero F_FULLFSYNC/fsync across the whole CLI; CAS/cache writes are plain data.write(to:), and macOS soft fsync is already fast on the runner). sync=none would be harmless for ephemeral VMs but wouldn’t move this build, so it’s left out to keep the change minimal.

Safety

These VMs are ephemeral — cloned per Pod, discarded on exit — so host read-caching is pure upside with no durability tradeoff.

Validation

• go test ./internal/tart/... passes; no test asserts the exact tart run args.
• On-host disk A/B above (1.8× warm reads on real runner hardware).
• The build-level effect (whether it narrows the benchmark gap) will be measured by the runner benchmark once this is live on the fleet. caching=cached is a strict read improvement regardless; it’s validated at the disk layer, not yet at the build layer.

Rollout

tart-kubelet runs as a launchd agent on each Mac mini, so this takes effect as the updated binary rolls to the hosts (not instant on merge) — gradual and reversible.

🤖 Generated with Claude Code

No GitHub comments yet.