Lab notes¶
Engineering experiment records preserve performance measurements, mechanism analyses and failure reviews to explain past design choices. For current executor choices, use benchmarks and comparisons.
Numbers on each page belong to their frozen cohorts and artifact versions and do not represent the performance level of the current build. .data/ paths point to local archives that are not published with the repository or site; they exist for cohort provenance only.
Complete CSV/JSON, provenance summaries and raw records remain in each batch's local .data/ or publication directory. Previously public data files are archived under benchmark/.data/docs-assets-prune-20261018/; the migration inventory retains original paths, byte hashes and archive locations.
| Page | Contents |
|---|---|
| VM startup performance analysis | Startup-path breakdowns, regressions, and acceptance records |
| End-to-end task performance analysis | End-to-end measurements of real agent tasks across executors |
| VM memory performance analysis | Static curves and readback costs for deduplication, shared libraries, and pooling |
| OpenSandbox: startup, disk and memory | Native guest creation wait, compressed image inventory and complete service-group memory |
| V2 versus native Docker: storage, startup and memory | Same-digest native Docker inventories, CLI readiness and complete deployment memory |
| Filesystem performance analysis | Load measurements and retest records for OverlayFS configurations |
| Cache and asynchronous-write performance curves | Engineering curves for request size, concurrency, synchronization, and write coalescing |
| Benchmark methodology and evidence | Measurement scope and evidence inventory of historical cohorts |
| Memory offload during Agent waits | Fixed SWE replay, Host wait management and memory/latency tradeoffs |