How long do file-intensive tasks take?¶

Motivation¶

pVisor host is close to direct execution for the tested repository checks and incremental builds. When VM isolation is needed and writes can merge before synchronization, guest writeback shortens publication waiting. VM file access still has additional costs: faster startup does not guarantee faster tasks, and larger working sets can time out.

The three-round publication median is 1.545 s for Native, 11.675 s for the default pVisor VM and 6.171 s for its writeback configuration. Default VM Ready is 101.75 ms. Task completion, startup and resource failures are reported separately.

Experiment design¶

Measure work plus final synchronization for repository inspection, incremental builds and publication as task completion time. It represents the wait imposed by the selected execution mode for file tasks. A lower time means faster completion within the same budget.

Nine modes on one host share static workloads, deterministic read-only inputs and Python/git/C tools, with images and disks prepared beforehand. Linux x86_64 / Ryzen 7 9700X / Btrfs; the entire process tree and private Docker share a two-core quota, 2 GiB and zero swap, with CPU affinity 0,1. VMs have 2 vCPUs/512 MiB; a parent-group cap for Native, host and containers is not equivalent to guest RAM. Tools come from a frozen Fedora 45 environment; the stock reference kernel is 7.2.11-300.fc45. pVisor uses the static release with matching embedded firmware.

Twelve mechanism workloads, four resource workloads and three workflows each use three fresh environments per mode, in randomized order, with three dependent rounds per environment. Sum the three completion times inside each environment before taking the median across environments. Host page caches remain warm; later rounds inherit verification-warmed pages. Ready runs from a shared pre-exec marker to payload Ready, excluding namespace preparation, task, exit and cleanup; it is separate from task clocks.

Each measured executor process has a 90-second limit and must exit normally, validate every round and pass an independent host final-state oracle. Staged modes must preserve the original workspace. After the first mode/workload failure, remaining planned observations for that cell are not run. Incomplete N=3 cells get no median or zero filling, and valid slow values remain. N=3 is exploratory, without reliable tails, cold-disk conclusions or cross-host rankings.

Data and analysis¶

Startup and complete tasks¶

Ready is in ms and tasks in s, medians across three fresh environments. Task columns sum three rounds per environment, excluding initial preparation and additional verification. Every build round edits a real translation unit and every publication creates a new directory. First image acquisition is excluded.

Host checks and builds stay close to Native; staging and VM file access add waiting. The pVisor VM becomes ready before the same-cohort stock reference VMs, but that does not offset repository and publication access costs. Writeback shortens publication; suitability still depends on complete task time.

Small writes and synchronization¶

Milliseconds, N=3 medians of three-round completion sums. Sequential writes total 4 MiB per round; synchronous writes total 1 MiB, fsyncing every 4 KiB. Synchronization semantics determine merge opportunities; write-return time alone omits completion costs.

Working set and concurrency¶

Seconds, N=3 medians of three-round completion sums. Every round writes a logical 1 GiB across four files; only working set and workers change. — means incomplete N=3, not zero time. Larger working sets also increase retained and synchronized bytes; four anchors do not establish a capacity knee.

Startup range¶

Configuration and coverage¶

The plan contains 19 workloads × 9 modes × 3 fresh environments = 513 environments. 505 ran: 501 passed and 4 failed; 8 further observations were not run under the registered rule. 167 / 171 cells have complete N=3 results, with 1503 valid round records.

Historical data¶

Earlier seven-tool, five-mode and nine-mode results retain their original scopes. They are not pooled with current samples or used for cross-cohort optimization ratios.

Evidence and reproduction¶

All workload levels · Ready · Failures and missing observations · Provenance

2026-10-25, source 9e1b8f15, based on main 92a4ce4e. Full raw observations, frozen tools/binaries/firmware, failed preflights and derived plots remain locally in benchmark/filesystem/.data/user-rebase-20261025-c/; the previous public materials are archived there too. Reproduction · Cache and asynchronous-write engineering analysis. Engineering attribution does not fill missing user results.