[ML] Tune test parallelism: fix low-core and improve mid-core machines

Two changes to the test parallelism formula:

1. On <=4 cores (macOS CI Orka VMs), use numCpus-1 instead of numCpus.
   Full CPU saturation caused CKMostCorrelatedTest/testScale to fail
   because wall-clock complexity assertions became unreliable.

2. On >4 cores, use ceil(numCpus/2) instead of ceil(numCpus/3).
   The /3 divisor was too conservative on 8-core Linux aarch64 CI
   (52 min with -j 3 vs 39.5 min with -j 8). The /2 divisor gives
   -j 4 on 8 cores — a better balance of parallelism vs contention.

Also adds diagnostic logging of CPU count and parallelism settings.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: edsavage

build.gradle

@@ -116,9 +116,12 @@ project.ext.numCpus = Runtime.runtime.availableProcessors()
 // Each suite spawns ctest --parallel <all_cpus> internally, so running too
 // many suites simultaneously causes resource contention. These values were
 // determined empirically (see PR #2900).
-// On low-core machines (<=4, e.g. macOS CI Orka VMs), use all cores since
-// CTest internal parallelism is modest enough that contention is minimal.
-project.ext.testParallel = isWindows ? 2 : (numCpus <= 4 ? numCpus : Math.max(2, (int) Math.ceil(numCpus / 3.0)))
+// On low-core machines (<=4, e.g. macOS CI Orka VMs with 4 cores),
+// cap at numCpus-1 to leave headroom for timing-sensitive tests
+// (e.g. CKMostCorrelatedTest/testScale). For higher core counts,
+// ceil(numCpus/2) balances parallelism vs contention — ceil(numCpus/3)
+// was too conservative on 8-core machines (52 min vs 39.5 min).
+project.ext.testParallel = isWindows ? 2 : (numCpus <= 4 ? Math.max(2, numCpus - 1) : Math.max(2, (int) Math.ceil(numCpus / 2.0)))
 project.ext.makeEnvironment = [ 'CPP_CROSS_COMPILE': cppCrossCompile,
                                 'VERSION_QUALIFIER': versionQualifier,
                                 'SNAPSHOT': (isSnapshot ? 'yes' : 'no'),

dev-tools/docker/docker_entrypoint.sh

@@ -68,13 +68,15 @@ if [ "x$1" = "x--test" ] ; then
     echo passed > build/test_status.txt
     # Each test suite spawns ctest --parallel <nproc> internally, so limit
     # the number of suites running concurrently to avoid resource contention.
-    # On low-core machines (<=4), use all cores since CTest internal
-    # parallelism is modest enough that contention is minimal.
+    # On low-core machines (<=4), cap at nproc-1 to leave headroom for
+    # timing-sensitive tests (e.g. CKMostCorrelatedTest/testScale).
+    # For higher core counts, ceil(nproc/2) balances parallelism vs
+    # contention — ceil(nproc/3) was too conservative on 8-core machines.
     NCPUS=$(nproc)
     if [ "$NCPUS" -le 4 ]; then
-        TEST_PARALLEL=$NCPUS
+        TEST_PARALLEL=$(( NCPUS > 2 ? NCPUS - 1 : 2 ))
     else
-        TEST_PARALLEL=$(( (NCPUS + 2) / 3 ))
+        TEST_PARALLEL=$(( (NCPUS + 1) / 2 ))
     fi
     echo "Test parallelism: nproc=${NCPUS}, TEST_PARALLEL=${TEST_PARALLEL} (cmake --build -j ${TEST_PARALLEL})"
     cmake --build cmake-build-docker ${CMAKE_VERBOSE} -j ${TEST_PARALLEL} -t test_individually || echo failed > build/test_status.txt