LinuxPerf

Count one callback invocation with the explicit PerfCheckerLinuxPerf companion. First prepare its worker project using Prepare the explicit worker below, then start Julia in that project. The companion does not change kernel permissions or affinity.

using PerfChecker, PerfCheckerLinuxPerf
called = Ref(false)
result = measure_counters(; events=["instructions", "cpu-cycles"]) do
    called[] = true
    sum(1:10000)
end
println((result.status, result.reason, called[], length(result.records)))

Observed on Linux x86_64/Julia 1.13.1: (:unavailable, "permission_denied", false, 0). The actual receipt preserves dependency versions and environment hashes. The callback did not execute, so there is no counter plot or performance result. With permission, scope is the calling OS thread's user-space execution; cycles have count unit 1, not seconds.

Prepare the explicit worker

PerfCheckerLinuxPerf 1.0.1 is registered in General. Use a separate worker project:

using Pkg
Pkg.activate("perf/linuxperf-worker")
Pkg.add(["PerfChecker", "PerfCheckerLinuxPerf"])
Pkg.add(Pkg.PackageSpec(name="LinuxPerf", version="0.4.2"))

Start Julia with --project=perf/linuxperf-worker for the callback above. Install your measured target in that worker project separately. LinuxPerf 0.4.2 requires PrettyTables 2; the other companion requires PrettyTables 3. The kernel must permit perf events on a supported little-endian Linux x86_64, i686 or aarch64 ABI; architecture guards do not qualify positive hardware counts. Published Core has a known ARM import limitation.

Collection window

The suite integration runs operation and synchronization once, then stops counters before the oracle. It has no implicit warmup. LinuxPerf counts the calling OS thread's user-space execution, excluding other threads/processes; enabled/running times are separate nanosecond values. LIKWID measures selected CPUs within inherited affinity; access mode and PID configuration determine attribution and direct/daemon modes may include unrelated activity. LIKWID group time is in seconds; raw event units are explicit. Counts and CPU cycles use unit 1, not seconds. An unavailable result is missing evidence. LinuxPerf's counter_not_scheduled refusal can occur after the operation, so inspect its lifecycle before retrying.

Use run_counter_suite or an explicit counter_executor in a Julia suite runner. Default CLI, VS Code, Pluto and MCP selectors do not offer these companions.