[Benchmarks] Pin benchmarks to small set of cores (#20403)

For better results stability, pin benchmark binaries to four cores with
the maximum available frequency.

---------

Co-authored-by: Łukasz Ślusarczyk <[email protected]>

Author: PatKamin

devops/actions/run-tests/benchmark/action.yml

@@ -82,25 +82,13 @@ runs:
       python3 ./devops/scripts/benchmarks/presets.py query "$PRESET"
       [ "$?" -ne 0 ] && exit 1  # Stop workflow if invalid preset
       echo "PRESET=$PRESET" >> $GITHUB_ENV
-  - name: Compute CPU core range to run benchmarks on
+  - name: Set NUMA node to run benchmarks on
     shell: bash
     run: |
-      # Compute the core range for the first NUMA node; second node is used by
-      # UMF. Skip the first 4 cores as the kernel is likely to schedule more
-      # work on these.
-      CORES="$(lscpu | awk '
-        /NUMA node0 CPU|On-line CPU/ {line=$0}
-        END {
-          split(line, a, " ")
-          split(a[4], b, ",")
-          sub(/^0/, "4", b[1])
-          print b[1]
-        }')"
-      echo "CPU core range to use: $CORES"
-      echo "CORES=$CORES" >> $GITHUB_ENV
-
-      ZE_AFFINITY_MASK=0
-      echo "ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK" >> $GITHUB_ENV
+      # Set CPU and GPU affinity for the first NUMA node; second node is used by UMF
+      NUMA_NODE=0
+      echo "ZE_AFFINITY_MASK=$NUMA_NODE" >> $GITHUB_ENV
+      echo "NUMA_NODE=$NUMA_NODE" >> $GITHUB_ENV
 
   # Compute-benchmarks relies on UR static libraries, cmake config files, etc.
   # DPC++ doesn't ship with these files. The easiest way of obtaining these
@@ -243,7 +231,8 @@ runs:
       WORKDIR="$(realpath ./llvm_test_workdir)"
       if [ -n "$WORKDIR" ] && [ -d "$WORKDIR" ] && [[ "$WORKDIR" == *llvm_test_workdir* ]]; then rm -rf "$WORKDIR" ; fi
 
-      taskset -c "$CORES" ./devops/scripts/benchmarks/main.py "$WORKDIR" \
+      numactl --cpunodebind "$NUMA_NODE" --membind "$NUMA_NODE" \
+      ./devops/scripts/benchmarks/main.py "$WORKDIR" \
         --sycl "$(realpath ./toolchain)" \
         --ur "$(realpath ./ur/install)" \
         --adapter "$FORCELOAD_ADAPTER" \

devops/scripts/benchmarks/benches/base.py

@@ -6,15 +6,18 @@
 import os
 import shutil
 import subprocess
-from pathlib import Path
+from abc import ABC, abstractmethod
 from enum import Enum
-from utils.result import BenchmarkMetadata, BenchmarkTag, Result
+from pathlib import Path
+
+from psutil import Process
+
 from options import options
-from utils.utils import download, run
-from abc import ABC, abstractmethod
-from utils.unitrace import get_unitrace
 from utils.flamegraph import get_flamegraph
 from utils.logger import log
+from utils.result import BenchmarkMetadata, BenchmarkTag, Result
+from utils.unitrace import get_unitrace
+from utils.utils import download, run
 
 
 class TracingType(Enum):
@@ -151,6 +154,8 @@ def run_bench(
             log.debug(f"FlameGraph perf data: {perf_data_file}")
             log.debug(f"FlameGraph command: {' '.join(command)}")
 
+        command = self.taskset_cmd() + command
+
         try:
             result = run(
                 command=command,
@@ -252,6 +257,16 @@ def get_metadata(self) -> dict[str, BenchmarkMetadata]:
             )
         }
 
+    def taskset_cmd(self) -> list[str]:
+        """Returns a list of strings with taskset usage for core pinning.
+        Pin compute benchmarks to a CPU cores set to ensure consistent results
+        and non-zero CPU count measurements (e.g. avoid E-cores). Exactly 4 cores
+        are pinned by default to satisfy multiple threads benchmarks. It is assumed
+        that they have the maximum, or at least similar, frequency.
+        """
+        selected_cores = [str(core) for core in Process().cpu_affinity()[:4]]  # type: ignore
+        return ["taskset", "-c", ",".join(selected_cores)]
+
     @staticmethod
     def get_adapter_full_path():
         for libs_dir_name in ["lib", "lib64"]:

devops/scripts/benchmarks/benches/compute.py

@@ -3,7 +3,6 @@
 # See LICENSE.TXT
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-import copy
 import csv
 import io
 import math

devops/scripts/benchmarks/requirements.txt

@@ -3,3 +3,4 @@ mpld3==0.5.10
 dataclasses-json==0.6.7
 PyYAML==6.0.1
 Mako==1.3.0
+psutil>=7.0.0

devops/scripts/install_build_tools.sh

@@ -28,7 +28,8 @@ apt update && apt install -yqq \
       libzstd-dev \
       linux-tools-generic \
       linux-tools-common \
-      time 
+      time \
+      numactl
 
 # To obtain latest release of spriv-tool.
 # Same as what's done in SPRIV-LLVM-TRANSLATOR: