[Benchmarks] Pin benchmarks to small set of cores

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

@@ -79,14 +79,14 @@ runs:
     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
+      # UMF. Skip the first 3 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])
+          sub(/^0/, "3", b[1])
           print b[1]
         }')"
       echo "CPU core range to use: $CORES"

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):
@@ -167,6 +170,8 @@ 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,
@@ -268,6 +273,28 @@ 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
+        with the maximum frequency are pinned by default to satisfy multiple threads benchmarks.
+        """
+        available_cores = Process().cpu_affinity()
+        core_frequencies = []
+        for core in available_cores:  # type: ignore
+            with open(
+                f"/sys/devices/system/cpu/cpu{core}/cpufreq/cpuinfo_max_freq"
+            ) as f:
+                freq = int(f.read().strip())
+                core_frequencies.append((core, freq))
+        selected = core_frequencies[:4]  # first ones have highest frequency
+        if len({freq for _, freq in selected}) > 1:
+            log.warning(
+                f"Selected cores for pinning have differing max frequencies: {selected}"
+            )
+        cores_list = ",".join([str(core) for core, _ in selected])
+        return ["taskset", "-c", cores_list]
+
 
 class Suite(ABC):
     @abstractmethod

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