Limit concurrent workers in batch

arc/job/pipe/pipe_coordinator.py

@@ -154,6 +154,7 @@ def submit_pipe_run(self, run_id: str, tasks: List[TaskSpec],
             tasks=tasks,
             cluster_software=cluster_software,
             max_workers=pipe_settings.get('max_workers', 100),
+            max_concurrent=pipe_settings.get('max_concurrent'),
             max_attempts=pipe_settings.get('max_attempts', 3),
             pipe_root=pipe_root,
         )

arc/job/pipe/pipe_run.py

@@ -17,6 +17,7 @@
 import stat
 import sys
 import time
+from numbers import Integral
 from typing import Dict, List, Optional
 
 import arc.parser.parser as parser
@@ -51,16 +52,30 @@ class PipeRun:
         run_id (str): Unique identifier for this pipe run.
         tasks (List[TaskSpec]): Task specifications to execute.
         cluster_software (str): Cluster scheduler type.
-        max_workers (int): Maximum number of concurrent array workers.
+        max_workers (int): Maximum total array worker slots (array size).
+        max_concurrent (Optional[int]): Max workers running simultaneously
+            (like PBS ``%N``). ``None`` disables throttling.
         max_attempts (int): Maximum retry attempts per task.
     """
 
+    @staticmethod
+    def _validate_max_concurrent(max_concurrent: Optional[int]) -> None:
+        """Accept ``None`` or a positive integer for throttle settings."""
+        if max_concurrent is None:
+            return
+        if isinstance(max_concurrent, bool) or not isinstance(max_concurrent, Integral):
+            raise ValueError('PipeRun max_concurrent must be None or a positive integer.')
+        if max_concurrent > 0:
+            return
+        raise ValueError('PipeRun max_concurrent must be None or a positive integer.')
+
     def __init__(self,
                  project_directory: str,
                  run_id: str,
                  tasks: List[TaskSpec],
                  cluster_software: str,
                  max_workers: int = 100,
+                 max_concurrent: Optional[int] = None,
                  max_attempts: int = 3,
                  pipe_root: Optional[str] = None,
                  ):
@@ -69,6 +84,8 @@ def __init__(self,
         self.tasks = tasks
         self.cluster_software = cluster_software
         self.max_workers = max_workers
+        self._validate_max_concurrent(max_concurrent)
+        self.max_concurrent = None if max_concurrent is None else int(max_concurrent)
         self.max_attempts = max_attempts
         self.pipe_root = pipe_root if pipe_root is not None \
             else os.path.join(project_directory, 'calcs', 'pipe_' + run_id)
@@ -103,6 +120,7 @@ def _save_run_metadata(self) -> None:
             'status': self.status.value,
             'cluster_software': self.cluster_software,
             'max_workers': self.max_workers,
+            'max_concurrent': self.max_concurrent,
             'max_attempts': self.max_attempts,
             'task_family': task_family,
             'engine': engine,
@@ -146,6 +164,7 @@ def from_dir(cls, pipe_root: str) -> 'PipeRun':
             tasks=tasks,
             cluster_software=data['cluster_software'],
             max_workers=data.get('max_workers', 100),
+            max_concurrent=data.get('max_concurrent'),
             max_attempts=data.get('max_attempts', 3),
             pipe_root=pipe_root,
         )
@@ -192,12 +211,41 @@ def _submission_resources(self):
         Derive resource settings from the homogeneous task list.
 
         Returns:
-            Tuple[int, int, int]: ``(cpus, memory_mb, array_size)``
+            Tuple[int, int, int, Optional[int]]:
+                ``(cpus, memory_mb, array_size, throttle)`` where ``throttle``
+                caps workers running simultaneously (clamped to ``array_size``),
+                or ``None`` if unthrottled.
         """
         cpus = self.tasks[0].required_cores if self.tasks else 1
         memory_mb = self.tasks[0].required_memory_mb if self.tasks else 4096
         array_size = min(self.max_workers, len(self.tasks)) if self.tasks else self.max_workers
-        return cpus, memory_mb, array_size
+        throttle = None
+        if self.max_concurrent is not None and self.max_concurrent > 0:
+            throttle = min(self.max_concurrent, array_size)
+        return cpus, memory_mb, array_size, throttle
+
+    def _render_throttle(self, array_size: int, throttle: Optional[int]) -> Dict[str, str]:
+        """
+        Render scheduler-specific array-range and extra-directives strings.
+
+        SLURM/PBS encode the throttle as an inline ``%K`` suffix on the range.
+        SGE uses a separate ``-tc`` directive. HTCondor uses ``max_materialize``
+        and takes a bare count (not a range) for ``queue``.
+        """
+        cs = 'sge' if self.cluster_software == 'oge' else self.cluster_software
+        if cs == 'htcondor':
+            array_range = str(array_size)
+            extra = f'max_materialize = {throttle}' if throttle else ''
+        elif cs == 'sge':
+            array_range = f'1-{array_size}'
+            extra = f'#$ -tc {throttle}' if throttle else ''
+        elif cs in ('slurm', 'pbs'):
+            suffix = f'%{throttle}' if throttle else ''
+            array_range = f'1-{array_size}{suffix}'
+            extra = ''
+        else:
+            raise NotImplementedError(f'No throttle rendering for {self.cluster_software}')
+        return {'array_range': array_range, 'extra_directives': extra}
 
     def write_submit_script(self) -> str:
         """
@@ -215,7 +263,8 @@ def write_submit_script(self) -> str:
             raise NotImplementedError(
                 f'No pipe submit template for cluster software: {self.cluster_software}. '
                 f'Available templates: {list(pipe_submit.keys())}')
-        cpus, memory_mb, array_size = self._submission_resources()
+        cpus, memory_mb, array_size, throttle = self._submission_resources()
+        throttle_fields = self._render_throttle(array_size, throttle)
         server = servers_dict.get('local', {})
         queue, _ = next(iter(server.get('queues', {}).items()), ('', None))
         engine = self.tasks[0].engine if self.tasks else ''
@@ -226,7 +275,8 @@ def write_submit_script(self) -> str:
             env_setup = f'{env_setup}\n{scratch_export}' if env_setup else scratch_export
         content = pipe_submit[template_key].format(
             name=f'pipe_{self.run_id}',
-            max_task_num=array_size,
+            array_range=throttle_fields['array_range'],
+            extra_directives=throttle_fields['extra_directives'],
             pipe_root=self.pipe_root,
             python_exe=sys.executable,
             cpus=cpus,

arc/job/pipe/pipe_run_test.py

@@ -11,6 +11,7 @@
 import tempfile
 import time
 import unittest
+import uuid
 
 from arc.job.adapters.mockter import MockAdapter
 from arc.job.pipe.pipe_state import TaskState, PipeRunState, TaskSpec, read_task_state, update_task_state
@@ -134,12 +135,14 @@ def setUp(self):
     def tearDown(self):
         shutil.rmtree(self.tmpdir, ignore_errors=True)
 
-    def _make_run(self, cluster_software, max_workers=10, n_tasks=None):
+    def _make_run(self, cluster_software, max_workers=10, n_tasks=None,
+                  max_concurrent=None, run_id=None):
         n = n_tasks if n_tasks is not None else max_workers
         tasks = [_make_spec(f't_{i}') for i in range(n)]
-        run = PipeRun(project_directory=self.tmpdir, run_id='sub_test',
+        run = PipeRun(project_directory=self.tmpdir,
+                      run_id=run_id or f'sub_test_{uuid.uuid4().hex[:8]}',
                       tasks=tasks, cluster_software=cluster_software,
-                      max_workers=max_workers)
+                      max_workers=max_workers, max_concurrent=max_concurrent)
         run.stage()
         return run
 
@@ -170,6 +173,104 @@ def test_htcondor_content(self):
             content = f.read()
         self.assertIn('queue 12', content)
 
+    def test_slurm_throttle(self):
+        run = self._make_run('slurm', max_workers=100, n_tasks=100, max_concurrent=8)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#SBATCH --array=1-100%8', content)
+
+    def test_pbs_throttle(self):
+        run = self._make_run('pbs', max_workers=50, n_tasks=50, max_concurrent=4)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#PBS -J 1-50%4', content)
+
+    def test_sge_throttle_uses_tc_directive(self):
+        run = self._make_run('sge', max_workers=20, n_tasks=20, max_concurrent=5)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#$ -t 1-20', content)
+        self.assertIn('#$ -tc 5', content)
+
+    def test_htcondor_throttle_uses_max_materialize(self):
+        run = self._make_run('htcondor', max_workers=12, n_tasks=12, max_concurrent=3)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('queue 12', content)
+        self.assertIn('max_materialize = 3', content)
+
+    def test_throttle_clamped_to_array_size(self):
+        # max_concurrent > array_size should clamp to array_size (no-op throttle).
+        run = self._make_run('slurm', max_workers=6, n_tasks=6, max_concurrent=99)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#SBATCH --array=1-6%6', content)
+
+    def test_unthrottled_has_no_throttle_markers(self):
+        """Regression guard: unthrottled submit scripts contain no throttle syntax."""
+        array_line_markers = {
+            'slurm':    '#SBATCH --array=',
+            'pbs':      '#PBS -J ',
+            'sge':      '#$ -t ',
+            'htcondor': 'queue ',
+        }
+        for cs, marker in array_line_markers.items():
+            with self.subTest(cluster_software=cs):
+                run = self._make_run(cs, max_workers=10, n_tasks=10, max_concurrent=None)
+                with open(run.write_submit_script()) as f:
+                    content = f.read()
+                array_line = next(line for line in content.splitlines() if marker in line)
+                self.assertNotIn('%', array_line)
+                self.assertNotIn('-tc ', content)
+                self.assertNotIn('max_materialize', content)
+
+    def test_oge_routes_to_sge_throttle(self):
+        """cluster_software='oge' should use the SGE throttle directive."""
+        run = self._make_run('oge', max_workers=15, n_tasks=15, max_concurrent=4)
+        with open(run.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#$ -t 1-15', content)
+        self.assertIn('#$ -tc 4', content)
+
+    def test_invalid_max_concurrent_values_raise(self):
+        """Only None and positive integers are accepted."""
+        for max_concurrent in (-1, 0, -2, 1.5, '3', True, False):
+            with self.subTest(max_concurrent=max_concurrent):
+                with self.assertRaisesRegex(ValueError, 'max_concurrent'):
+                    PipeRun(project_directory=self.tmpdir,
+                            run_id=f'invalid_{uuid.uuid4().hex[:8]}',
+                            tasks=[_make_spec('t_0')],
+                            cluster_software='slurm',
+                            max_concurrent=max_concurrent)
+
+    def test_render_throttle_branching_matrix(self):
+        """Unit-test _render_throttle directly across scheduler × throttle combos."""
+        run = self._make_run('slurm', max_workers=1, n_tasks=1)
+        cases = [
+            ('slurm',    100, None, {'array_range': '1-100',   'extra_directives': ''}),
+            ('slurm',    100, 8,    {'array_range': '1-100%8', 'extra_directives': ''}),
+            ('pbs',      50,  4,    {'array_range': '1-50%4',  'extra_directives': ''}),
+            ('sge',      20,  None, {'array_range': '1-20',    'extra_directives': ''}),
+            ('sge',      20,  5,    {'array_range': '1-20',    'extra_directives': '#$ -tc 5'}),
+            ('oge',      20,  5,    {'array_range': '1-20',    'extra_directives': '#$ -tc 5'}),
+            ('htcondor', 12,  None, {'array_range': '12',      'extra_directives': ''}),
+            ('htcondor', 12,  3,    {'array_range': '12',      'extra_directives': 'max_materialize = 3'}),
+        ]
+        for cs, array_size, throttle, expected in cases:
+            with self.subTest(cluster_software=cs, throttle=throttle):
+                run.cluster_software = cs
+                self.assertEqual(run._render_throttle(array_size, throttle), expected)
+
+    def test_from_dir_round_trip_preserves_max_concurrent(self):
+        """Persist max_concurrent through run.json so crash-recovered runs keep their throttle."""
+        run = self._make_run('slurm', max_workers=20, n_tasks=20, max_concurrent=7)
+        run.write_submit_script()
+        reloaded = PipeRun.from_dir(run.pipe_root)
+        self.assertEqual(reloaded.max_concurrent, 7)
+        with open(reloaded.write_submit_script()) as f:
+            content = f.read()
+        self.assertIn('#SBATCH --array=1-20%7', content)
+
     def test_overwrite_is_safe(self):
         run = self._make_run('slurm')
         p1 = run.write_submit_script()

arc/settings/settings.py

@@ -306,6 +306,7 @@
     'enabled': False,          # Set to True to enable pipe mode (it is off by default, use it for large compute campaigns).
     'min_tasks': 10,           # Minimum batch size to trigger pipe mode.
     'max_workers': 100,        # Upper bound on array worker slots per PipeRun.
+    'max_concurrent': None,    # Max array workers running simultaneously (like PBS `%N`). None = unthrottled.
     'max_attempts': 3,         # Retry budget per task before terminal failure.
     'lease_duration_hrs': 1,   # Worker lease duration in hours (default 1h).
     'env_setup': {},           # Engine-specific shell setup commands, e.g.,

arc/settings/submit.py

@@ -54,10 +54,10 @@
 #SBATCH -N 1
 #SBATCH -n {cpus}
 #SBATCH --mem={memory}
-#SBATCH --array=1-{max_task_num}
+#SBATCH --array={array_range}
 #SBATCH -o {pipe_root}/out_%a.txt
 #SBATCH -e {pipe_root}/err_%a.txt
-
+{extra_directives}
 {env_setup}
 WORKER_ID=$SLURM_ARRAY_TASK_ID
 
@@ -68,8 +68,8 @@
 #PBS -q {queue}
 #PBS -l ncpus={cpus}
 #PBS -l mem={memory}mb
-#PBS -J 1-{max_task_num}
-
+#PBS -J {array_range}
+{extra_directives}
 {env_setup}
 WORKER_ID="$PBS_ARRAY_INDEX"
 
@@ -80,10 +80,10 @@
 #$ -q {queue}
 #$ -pe smp {cpus}
 #$ -l h_vmem={memory}M
-#$ -t 1-{max_task_num}
+#$ -t {array_range}
 #$ -o {pipe_root}/out_$SGE_TASK_ID.txt
 #$ -e {pipe_root}/err_$SGE_TASK_ID.txt
-
+{extra_directives}
 {env_setup}
 WORKER_ID=$SGE_TASK_ID
 
@@ -96,7 +96,8 @@
 output = {pipe_root}/out_$(Process).txt
 error = {pipe_root}/err_$(Process).txt
 log = {pipe_root}/condor.log
-queue {max_task_num}
+{extra_directives}
+queue {array_range}
 """,
 }