Add default benchmark scheduler

gematria/datasets/pipelines/benchmark_cpu_scheduler.py

@@ -13,6 +13,9 @@
 # limitations under the License.
 
 import abc
+from collections.abc import Collection
+import os
+import re
 from typing_extensions import override
 
 
@@ -66,3 +69,82 @@ def setup_and_get_benchmark_core(self) -> int | None:
   @override
   def verify(self):
     pass
+
+
+def _get_neighboring_threads(cpu_index: int) -> list[int]:
+  with open(
+      f'/sys/devices/system/cpu/cpu{cpu_index}/topology/thread_siblings_list'
+  ) as thread_sibling_list_handle:
+    neighboring_threads_strings = re.split(
+        '[-,]+', thread_sibling_list_handle.read().strip()
+    )
+    neighboring_threads = [
+        int(cpu_index_str) for cpu_index_str in neighboring_threads_strings
+    ]
+  return neighboring_threads
+
+
+class DefaultBenchmarkScheduler(BenchmarkScheduler):
+  """A BenchmarkScheduler that schedules processes separately.
+
+  DefaultBenchmarkScheduler schedules the main process and the benchmark
+  subprocess on separate cores, making sure to reserve the second hyperthread
+  on the benchmarking core to prevent interference. It expects that the main
+  process is initially given a CPU Mask with three active threads, additionally
+  assuming that two of the threads are neighboring (part of the same core).
+  Errors are raised if these conditions are not met. The benchmarking core
+  returned is one of the two neighboring threads. The main process has its
+  CPU mask limited to the thread that neighbors neither of the other threads.
+  """
+
+  def __init__(self):
+    self._cpu_mask = []
+
+  def _get_aux_core_and_hyperthread_pair(
+      self,
+      cpu_mask: Collection[int],
+  ) -> tuple[int, list[int]]:
+    for cpu_index in cpu_mask:
+      neighboring_threads = _get_neighboring_threads(cpu_index)
+      if len(neighboring_threads) != 2:
+        raise ValueError('Expected two hyperthreads per CPU.')
+
+      if (
+          neighboring_threads[0] not in cpu_mask
+          or neighboring_threads[1] not in cpu_mask
+      ):
+        continue
+
+      cpus = list(cpu_mask)
+      cpus.remove(neighboring_threads[0])
+      cpus.remove(neighboring_threads[1])
+
+      if not cpus:
+        continue
+
+      return (cpus[0], [neighboring_threads[0], neighboring_threads[1]])
+    raise ValueError(
+        'Expected a pair of neighboring hyperthreads and an aux thread in the'
+        ' CPU mask.'
+    )
+
+  @override
+  def setup_and_get_benchmark_core(self) -> int | None:
+    cpu_mask = os.sched_getaffinity(0)
+
+    if len(cpu_mask) != 3:
+      raise ValueError('Expected to have three CPUs.')
+
+    aux_core, hyperthread_pair = self._get_aux_core_and_hyperthread_pair(
+        cpu_mask
+    )
+    os.sched_setaffinity(0, [aux_core])
+    self._cpu_mask = [aux_core]
+
+    return hyperthread_pair[0]
+
+  @override
+  def verify(self):
+    cpu_mask = list(os.sched_getaffinity(0))
+    if self._cpu_mask != cpu_mask:
+      raise ValueError('Expected the CPU mask to not change.')

gematria/datasets/pipelines/benchmark_cpu_scheduler_test.py

@@ -12,18 +12,124 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
+from collections.abc import Iterable
+
 from absl.testing import absltest
 
 from gematria.datasets.pipelines import benchmark_cpu_scheduler
 
 
 class BenchmarkSchedulerTests(absltest.TestCase):
 
+  def _set_normal_affinity(self):
+    cpu_mask = os.sched_getaffinity(0)
+    cpu_mask_list = list(cpu_mask)
+
+    original_cpu_mask = list(cpu_mask_list)
+    self.addCleanup(lambda: os.sched_setaffinity(0, original_cpu_mask))
+
+    # We first select a CPU to use for the hyperthread pair and then remove
+    # both cores in the pair from the mask set to ensure that we do not
+    # accidentally select an auxiliary CPU that is part of the pair.
+    hyperthread_pair_part = next(iter(cpu_mask))
+    hyperthread_pair = benchmark_cpu_scheduler._get_neighboring_threads(
+        hyperthread_pair_part
+    )
+    cpu_mask.remove(hyperthread_pair[0])
+    cpu_mask.remove(hyperthread_pair[1])
+    aux_cpu = cpu_mask.pop()
+    new_cpu_mask = [aux_cpu, *hyperthread_pair]
+
+    os.sched_setaffinity(0, new_cpu_mask)
+    return (aux_cpu, hyperthread_pair, cpu_mask_list)
+
   def test_no_scheduling(self):
     scheduler = benchmark_cpu_scheduler.NoSchedulingBenchmarkScheduler()
     self.assertIsNone(scheduler.setup_and_get_benchmark_core())
     scheduler.verify()
 
+  def test_default_scheduler_get_neighboring_threads(self):
+    neighboring_threads = benchmark_cpu_scheduler._get_neighboring_threads(0)
+
+    # Just check that we get two CPU ids back that are not the same. We cannot
+    # do much more without knowing more about the system topology, and this
+    # should be a reasonable enough test.
+    self.assertLen(neighboring_threads, 2)
+    self.assertNotEqual(neighboring_threads[0], neighboring_threads[1])
+
+  def test_default_scheduler_get_cores(self):
+    expected_aux_cpu, expected_hyperthread_pair, _ = self._set_normal_affinity()
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    cpu_mask = os.sched_getaffinity(0)
+    aux_cpu, hyperthread_pair = scheduler._get_aux_core_and_hyperthread_pair(
+        cpu_mask
+    )
+    self.assertEqual(aux_cpu, expected_aux_cpu)
+    self.assertContainsSubsequence(hyperthread_pair, expected_hyperthread_pair)
+
+  def test_default_scheduler_get_cores_no_neighboring_threads(self):
+    cpu_mask = os.sched_getaffinity(0)
+
+    # If we have less than five CPUs, that means we are guaranteed to get
+    # a pair when selecting three of them. Skip this test in those
+    # instances.
+    if len(cpu_mask) < 5:
+      self.skipTest('Not enough cores to complete setup properly.')
+
+    three_cores = [cpu_mask.pop(), cpu_mask.pop(), cpu_mask.pop()]
+
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    with self.assertRaisesRegex(
+        ValueError,
+        'Expected a pair of neighboring hyperthreads and an aux thread in the'
+        ' CPU mask.',
+    ):
+      scheduler._get_aux_core_and_hyperthread_pair(three_cores)
+
+  def test_default_scheduler_setup(self):
+    expected_aux_cpu, expected_hyperthread_pair, _ = self._set_normal_affinity()
+
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    benchmark_core = scheduler.setup_and_get_benchmark_core()
+    self.assertIn(benchmark_core, expected_hyperthread_pair)
+    set_cpu_mask = os.sched_getaffinity(0)
+    self.assertSequenceEqual(
+        set_cpu_mask,
+        {
+            expected_aux_cpu,
+        },
+    )
+
+  def test_default_scheduler_not_three_cpus(self):
+    old_cpu_mask = os.sched_getaffinity(0)
+    cpu_mask_list = list(old_cpu_mask)
+    os.sched_setaffinity(0, cpu_mask_list[0:2])
+
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    with self.assertRaises(ValueError):
+      scheduler.setup_and_get_benchmark_core()
+
+    os.sched_setaffinity(0, old_cpu_mask)
+
+  def test_default_scheduler_verify(self):
+    self._set_normal_affinity()
+
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    scheduler.setup_and_get_benchmark_core()
+    scheduler.verify()
+
+  def test_default_scheduler_verify_mask_changed(self):
+    _, _, old_cpu_mask = self._set_normal_affinity()
+
+    scheduler = benchmark_cpu_scheduler.DefaultBenchmarkScheduler()
+    scheduler.setup_and_get_benchmark_core()
+
+    cpu_mask_list = list(old_cpu_mask)
+    os.sched_setaffinity(0, cpu_mask_list[1:3])
+    with self.assertRaises(ValueError):
+      scheduler.verify()
+
 
 if __name__ == '__main__':
   absltest.main()