Right-size CI runners to reduce waste and shuffle build order to avoid OOM

libeigen/eigen!2367

Co-authored-by: Rasmus Munk Larsen <rmlarsen@gmail.com>

ci/build.linux.gitlab-ci.yml

@@ -104,6 +104,8 @@ build:linux:docs:
     EIGEN_CI_C_COMPILER: clang
     EIGEN_CI_CXX_COMPILER: clang++
     EIGEN_CI_BEFORE_SCRIPT: ". ci/scripts/build_and_install_doxygen.sh Release_1_13_2"
+  tags:
+    - saas-linux-medium-amd64
   rules:
     - if: $CI_PIPELINE_SOURCE == "schedule" && $CI_PROJECT_NAMESPACE == "libeigen"
     - if: $CI_PIPELINE_SOURCE == "web" && $CI_PROJECT_NAMESPACE == "libeigen"
@@ -154,8 +156,10 @@ build:linux:x86-64:nvhpc-26.1:default:
     EIGEN_CI_INSTALL: ""
     EIGEN_CI_CROSS_INSTALL: ""
     # NVHPC's compiler frontend is very memory-hungry with template-heavy code.
-    # Limit parallelism to avoid OOM kills from the kernel.
-    EIGEN_CI_BUILD_JOBS: "8"
+    # The 2xlarge runner has 128 GB, so 16 jobs gives ~8 GB per process.
+    # The shuffled-batch build strategy spreads memory-hungry targets across
+    # batches, preventing simultaneous OOM-prone compilations.
+    EIGEN_CI_BUILD_JOBS: "16"
 
 ######## CUDA ##################################################################
 

ci/scripts/build.linux.script.sh

@@ -29,7 +29,62 @@ jobs=""
 if [[ -n "${EIGEN_CI_BUILD_JOBS}" ]]; then
   jobs="-j${EIGEN_CI_BUILD_JOBS}"
 fi
-cmake --build . ${target} -- -k0 ${jobs} || cmake --build . ${target} -- -k0 -j1
+
+# For phony meta-targets (e.g. buildtests), shuffle the dependency list and
+# build in batches so that memory-hungry compilations (like bdcsvd with
+# nvc++) are spread out instead of all running at once.  Ninja ignores the
+# command-line target order and schedules by its dependency graph, so we
+# must feed it small batches to actually influence scheduling.
+# Falls back to the normal build if the target is not a phony or if
+# ninja/shuf are not available.
+batch_size=${EIGEN_CI_BUILD_BATCH_SIZE:-48}
+shuffled=false
+if [[ -n "${EIGEN_CI_BUILD_TARGET}" ]] && command -v ninja >/dev/null 2>&1; then
+  # Suppress xtrace while extracting and shuffling the target list
+  # to avoid dumping ~1200 lines to the CI log.
+  { set +x; } 2>/dev/null
+  deps=$(ninja -t query "${EIGEN_CI_BUILD_TARGET}" 2>/dev/null \
+         | awk '/^  input:/{found=1; next} /^  outputs:/{found=0} found && /^    /{print $1}')
+  # Deterministic shuffle: hash each target name and sort by hash.
+  # Stable across runs (helps ninja's .ninja_log and build caches),
+  # portable (no shuf dependency), and spreads same-family targets apart.
+  # Uses Knuth's multiplicative hash (golden-ratio prime 2654435761) for
+  # good avalanche — similar names like bdcsvd_1..bdcsvd_51 get widely
+  # dispersed instead of clustering together.
+  shuffled_deps=$(echo "$deps" | awk '
+    BEGIN { for(i=0;i<128;i++) ord[sprintf("%c",i)]=i }
+    { h=0
+      for(i=1;i<=length($0);i++) h=(h+ord[substr($0,i,1)])*2654435761%2147483647
+      printf "%010d %s\n",h,$0 }' | sort | sed 's/^[^ ]* //')
+  if [[ -n "$shuffled_deps" ]]; then
+    ndeps=$(echo "$shuffled_deps" | wc -l)
+    echo "Building ${ndeps} targets in batches of ${batch_size}"
+    shuffled=true
+    # Build in batches: ninja parallelises within each batch, but batches
+    # run sequentially so memory-hungry targets from different families
+    # don't pile up simultaneously.  Track failures so we can report the
+    # right exit code at the end.
+    # Note: xtrace stays off to avoid dumping the full target list.
+    # Use process substitution so the while loop runs in the current
+    # shell and build_failed propagates.
+    batch_num=0
+    build_failed=false
+    while IFS= read -r batch; do
+      batch_num=$((batch_num + 1))
+      echo "=== Batch ${batch_num} ==="
+      ninja -k0 ${jobs} ${batch} || ninja -k0 -j1 ${batch} || build_failed=true
+    done < <(echo "$shuffled_deps" | xargs -n "${batch_size}")
+    if [[ "$build_failed" == "true" ]]; then
+      echo "Some batches failed."
+      exit 1
+    fi
+  fi
+  set -x
+fi
+
+if [[ "$shuffled" != "true" ]]; then
+  cmake --build . ${target} -- -k0 ${jobs} || cmake --build . ${target} -- -k0 -j1
+fi
 
 # Return to root directory.
 cd ${rootdir}

ci/scripts/build.windows.script.ps1

@@ -5,8 +5,8 @@ $VS_INSTALL_DIR = &"${Env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\v
 # http://allen-mack.blogspot.com/2008/03/replace-visual-studio-command-prompt.html
 cmd.exe /c "`"${VS_INSTALL_DIR}\VC\Auxiliary\Build\vcvarsall.bat`" $EIGEN_CI_MSVC_ARCH -vcvars_ver=$EIGEN_CI_MSVC_VER & set" |
   foreach {
-    if ($_ -match "=") {
-      $v = $_.split("="); set-item -force -path "ENV:\$($v[0])" -value "$($v[1])"
+    if ($_ -match "^([^=]+)=(.*)$") {
+      set-item -force -LiteralPath "ENV:\$($Matches[1])" -value "$($Matches[2])"
     }
   }
 

ci/scripts/test.linux.script.sh

@@ -13,11 +13,16 @@ elif [[ ${EIGEN_CI_CTEST_LABEL} ]]; then
   target="-L ${EIGEN_CI_CTEST_LABEL}"
 fi
 
+exclude=""
+if [[ -n "${EIGEN_CI_CTEST_EXCLUDE}" ]]; then
+  exclude="-E ${EIGEN_CI_CTEST_EXCLUDE}"
+fi
+
 set +x
 
 EIGEN_CI_CTEST_PARALLEL=${EIGEN_CI_CTEST_PARALLEL:-${NPROC}}
 EIGEN_CI_CTEST_REPEAT=${EIGEN_CI_CTEST_REPEAT:-3}
-ctest_cmd="ctest ${EIGEN_CI_CTEST_ARGS} --parallel ${EIGEN_CI_CTEST_PARALLEL} --output-on-failure --no-compress-output --build-noclean ${target}"
+ctest_cmd="ctest ${EIGEN_CI_CTEST_ARGS} --parallel ${EIGEN_CI_CTEST_PARALLEL} --output-on-failure --no-compress-output --build-noclean ${target} ${exclude}"
 
 echo "Running initial tests..."
 if ${ctest_cmd} -T test; then

ci/test.linux.gitlab-ci.yml

@@ -266,6 +266,10 @@ test:linux:cuda-12.2:clang-12:
     EIGEN_CI_CROSS_TARGET_TRIPLE: arm-linux-gnueabihf
     EIGEN_CI_CTEST_ARGS: --timeout 2000
     EIGEN_CI_CTEST_PARALLEL: "4"
+    # Thread pool tests are too slow under qemu emulation.
+    EIGEN_CI_CTEST_EXCLUDE: tensor_thread_pool
+  tags:
+    - saas-linux-large-amd64
 
 .test:linux:arm:gcc-10:default:
   extends: .test:linux:arm
@@ -350,6 +354,10 @@ test:linux:aarch64:clang-14:default:unsupported:
     EIGEN_CI_CROSS_TARGET_TRIPLE: powerpc64le-linux-gnu
     EIGEN_CI_CTEST_ARGS: --timeout 2000
     EIGEN_CI_CTEST_PARALLEL: "4"
+    # Thread pool tests are too slow under qemu emulation.
+    EIGEN_CI_CTEST_EXCLUDE: tensor_thread_pool
+  tags:
+    - saas-linux-large-amd64
 
 .test:linux:ppc64le:gcc-14:default:
   extends: .test:linux:ppc64le
@@ -393,6 +401,10 @@ test:linux:ppc64le:clang-16:default:unsupported:
     EIGEN_CI_CROSS_INSTALL: g++-14-loongarch64-linux-gnu qemu-user-static
     EIGEN_CI_CTEST_ARGS: --timeout 2000
     EIGEN_CI_CTEST_PARALLEL: "4"
+    # Thread pool tests are too slow under qemu emulation.
+    EIGEN_CI_CTEST_EXCLUDE: tensor_thread_pool
+  tags:
+    - saas-linux-large-amd64
 
 # GCC-14 (Ubuntu 24)
 .test:linux:loongarch64:gcc-14:default: