Measure memory_threshold against the test process's own RSS

[davidanthoff]

Summary

memory_threshold, an experimental option that is off by default, now compares the test process's own current resident memory with total system memory. It used to look at system-wide memory use. A test process recycles itself once rss / Sys.total_memory() > threshold.

Why

The old check was 1 - Sys.free_memory()/Sys.total_memory() > threshold. On macOS, Sys.free_memory() stays close to zero even on an idle machine: on GitHub's macOS arm64 runners we measured 0.07–0.19 GB free out of 7 GB. So any threshold fired after every item, and every item paid for a process restart. A system-wide figure also recycles one process because of memory that other processes are using.

Recycling is the only mechanism that gives back memory a test process still holds live. Compiled code and caches grow over a process's lifetime, and no GC can free them. Measuring per process makes the option usable on every platform.

Changes

• New testprocess/TestItemServer/src/process_memory.jl with _current_rss(): current resident memory, not the peak that Sys.maxrss() reports.
  • Linux: /proc/self/statm.
  • macOS: task_info(MACH_TASK_BASIC_INFO). The struct layout was checked against the xnu headers.
  • Windows: K32GetProcessMemoryInfo with stdcall. SIZE_T is declared as Csize_t, so the struct is also correct in 32-bit Julia.
  • It uses only plain ccall and Julia 1.0-era Base, because the test server runs on Julia 1.0 and later. It returns nothing (never recycle) when it can't measure.
• _memory_over_threshold uses it. A threshold of 0.0 still recycles after every item, as the existing tests expect.
• The option is documented in the protocol, TestRunState, the test server and the execute_testrun docstring.
• New tests: _current_rss returns a plausible value, and a threshold of 1.0 never recycles.
• CHANGELOG entry. The option is experimental, so the change of meaning is noted rather than treated as breaking.

Testing

• test/test_worker_lifecycle.jl: 61/61 pass on Julia 1.12.7 (Windows). The "Julia 1.0/1.6 platform" items also pass, so the test server still loads there.
• _current_rss() checked directly on Windows with Julia 1.0.5, 1.6.7, 1.12.7 and 1.13.0:
  • It agrees with PowerShell's WorkingSet64 to within 0.6 MB.
  • It follows an allocation up, then back down after GC.gc(), while Sys.maxrss() stays at the peak.
• Only CI will exercise these paths: Linux, macOS and 32-bit Windows. The new _current_rss test item covers them.

Related

Follow-up to the arm64 GC investigation, alongside #121 (GC between items opt-in).

🤖 Generated with Claude Code