slurm_scheduler: add support for per replica log files and API access

Author: d4l3k

scripts/slurmtest.sh

@@ -18,12 +18,12 @@ source "$VENV"/bin/activate
 python --version
 pip install "$REMOTE_WHEEL"
 
-APP_ID="$(torchx run --wait --scheduler slurm --scheduler_args partition=compute,time=10 utils.echo --image /tmp --num_replicas 3)"
+APP_ID="$(torchx run --wait --scheduler slurm --scheduler_args partition=compute,time=10 utils.echo --num_replicas 3)"
 torchx status "$APP_ID"
 torchx describe "$APP_ID"
-LOG_FILE="slurm-$(basename "$APP_ID").out"
-cat "$LOG_FILE"
-LINES="$(wc -l "$LOG_FILE" | cut -d' ' -f1)"
+sacct -j "$(basename "$APP_ID")"
+torchx log "$APP_ID"
+LINES="$(torchx log "$APP_ID" | wc -l)"
 
 if [ "$LINES" -ne 3 ]
 then

torchx/schedulers/local_scheduler.py

@@ -909,14 +909,19 @@ def __del__(self) -> None:
 
 class LogIterator:
     def __init__(
-        self, app_id: str, regex: str, log_file: str, scheduler: LocalScheduler
+        self,
+        app_id: str,
+        regex: str,
+        log_file: str,
+        scheduler: LocalScheduler,
+        should_tail: bool = True,
     ) -> None:
         self._app_id: str = app_id
         self._regex: Pattern[str] = re.compile(regex)
         self._log_file: str = log_file
         self._log_fp: Optional[TextIO] = None
         self._scheduler: LocalScheduler = scheduler
-        self._app_finished: bool = False
+        self._app_finished: bool = not should_tail
 
     def _check_finished(self) -> None:
         # either the app (already finished) was evicted from the LRU cache

torchx/schedulers/slurm_scheduler.py

@@ -15,10 +15,13 @@
 import shlex
 import subprocess
 import tempfile
+import warnings
 from dataclasses import dataclass
-from typing import Any, Dict, List, Mapping, Optional, Tuple
+from datetime import datetime
+from typing import Any, Dict, List, Mapping, Optional, Tuple, Iterable
 
-from torchx.schedulers.api import AppDryRunInfo, DescribeAppResponse, Scheduler
+from torchx.schedulers.api import AppDryRunInfo, DescribeAppResponse, Scheduler, Stream
+from torchx.schedulers.local_scheduler import LogIterator
 from torchx.specs import (
     NONE,
     AppDef,
@@ -100,12 +103,16 @@ def from_role(
             if resource.gpu > 0:
                 sbatch_opts.setdefault("gpus-per-task", str(resource.gpu))
 
+        srun_opts = {
+            "output": f"slurm-{macros.app_id}-{name}.out",
+        }
+
         return cls(
             name=name,
             entrypoint=role.entrypoint,
             args=list(role.args),
             sbatch_opts=sbatch_opts,
-            srun_opts={},
+            srun_opts=srun_opts,
             env=dict(role.env),
         )
 
@@ -176,7 +183,11 @@ class SlurmScheduler(Scheduler):
     resource allocations and args and then sbatch is used to launch all of them
     together.
 
-    Logs are written to the default slurm log file.
+    Logs are available in combined form via ``torchx log``, the programmatic API
+    as well as in the job launch directory as
+    ``slurm-<jobid>-<role>-<replica_id>.out``. If TorchX is running in a
+    different directory than where the job was created the logs won't be able to
+    be found.
 
     Some of the config options passed to it are added as SBATCH arguments to each
     replica. See https://slurm.schedmd.com/sbatch.html#SECTION_OPTIONS for info
@@ -203,9 +214,7 @@ class SlurmScheduler(Scheduler):
         type: scheduler
         features:
             cancel: true
-            logs: |
-                Logs are accessible via the default slurm log file but not the
-                programmatic API.
+            logs: true
             distributed: true
             describe: |
                 Partial support. SlurmScheduler will return job and replica
@@ -262,7 +271,7 @@ def _submit_dryrun(
                     app_id=macros.app_id,
                     replica_id=str(replica_id),
                 )
-                name = f"{app.name}-{role.name}-{replica_id}"
+                name = f"{role.name}-{replica_id}"
                 replica_role = values.apply(role)
                 replicas[name] = SlurmReplicaRequest.from_role(name, replica_role, cfg)
         req = SlurmBatchRequest(
@@ -308,13 +317,11 @@ def describe(self, app_id: str) -> Optional[DescribeAppResponse]:
             ), f"failed to translate slurm state {state} to torchx state"
             app_state = state_enum
 
-            name_parts = row["JobName"].split("-")
-            if len(name_parts) < 3:
+            role, _, replica_id = row["JobName"].rpartition("-")
+            if not replica_id:
                 # name should always have at least 3 parts but sometimes sacct
                 # is slow to update
                 continue
-            role = name_parts[-2]
-            replica_id = int(name_parts[-1])
             if role not in roles:
                 roles[role] = Role(name=role, num_replicas=0, image="")
                 roles_statuses[role] = RoleStatus(role, [])
@@ -331,6 +338,34 @@ def describe(self, app_id: str) -> Optional[DescribeAppResponse]:
             msg=msg,
         )
 
+    def log_iter(
+        self,
+        app_id: str,
+        role_name: str,
+        k: int = 0,
+        regex: Optional[str] = None,
+        since: Optional[datetime] = None,
+        until: Optional[datetime] = None,
+        should_tail: bool = False,
+        streams: Optional[Stream] = None,
+    ) -> Iterable[str]:
+        if since or until:
+            warnings.warn(
+                "since and/or until times specified for SlurmScheduler.log_iter."
+                " These will be ignored and all log lines will be returned"
+            )
+        if streams is not None and streams != Stream.COMBINED:
+            warnings.warn(
+                "streams specified for SlurmScheduler.log_iter."
+                " These will be ignored and all log lines will be returned"
+            )
+
+        log_file = f"slurm-{app_id}-{role_name}-{k}.out"
+
+        return LogIterator(
+            app_id, regex or ".*", log_file, self, should_tail=should_tail
+        )
+
 
 def create_scheduler(session_name: str, **kwargs: Any) -> SlurmScheduler:
     return SlurmScheduler(

torchx/schedulers/test/slurm_scheduler_test.py

@@ -4,8 +4,11 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import os
 import subprocess
+import tempfile
 import unittest
+from contextlib import contextmanager
 from unittest.mock import MagicMock, call, patch
 
 from torchx import specs
@@ -18,6 +21,18 @@
 )
 
 
+@contextmanager
+def tmp_cwd():
+    path = tempfile.TemporaryDirectory()
+    cwd = os.getcwd()
+    os.chdir(path.name)
+    try:
+        yield
+    finally:
+        os.chdir(cwd)
+        path.cleanup()
+
+
 class SlurmSchedulerTest(unittest.TestCase):
     def test_create_scheduler(self) -> None:
         scheduler = create_scheduler("foo")
@@ -40,12 +55,12 @@ def test_replica_request(self) -> None:
             ),
         )
         sbatch, srun = SlurmReplicaRequest.from_role(
-            "role-name", role, cfg={}
+            "role-0", role, cfg={}
         ).materialize()
         self.assertEqual(
             sbatch,
             [
-                "--job-name=role-name",
+                "--job-name=role-0",
                 "--ntasks-per-node=1",
                 "--cpus-per-task=2",
                 "--mem=10",
@@ -54,7 +69,13 @@ def test_replica_request(self) -> None:
         )
         self.assertEqual(
             srun,
-            ["--export=FOO=bar", "echo", "'hello slurm'", "test"],
+            [
+                '--output=slurm-"$SLURM_JOB_ID"-role-0.out',
+                "--export=FOO=bar",
+                "echo",
+                "'hello slurm'",
+                "test",
+            ],
         )
 
         # test nomem option
@@ -133,25 +154,25 @@ def test_dryrun_multi_role(self) -> None:
         self.assertEqual(req.cmd, ["sbatch", "--parsable"])
         self.assertEqual(
             set(req.replicas.keys()),
-            {"foo-a-0", "foo-a-1", "foo-b-0"},
+            {"a-0", "a-1", "b-0"},
         )
 
         script = req.materialize()
         self.assertEqual(
             script,
             """#!/bin/bash
-#SBATCH --job-name=foo-a-0 --ntasks-per-node=1
+#SBATCH --job-name=a-0 --ntasks-per-node=1
 #SBATCH hetjob
-#SBATCH --job-name=foo-a-1 --ntasks-per-node=1
+#SBATCH --job-name=a-1 --ntasks-per-node=1
 #SBATCH hetjob
-#SBATCH --job-name=foo-b-0 --ntasks-per-node=1
+#SBATCH --job-name=b-0 --ntasks-per-node=1
 
 # exit on error
 set -e
 
-srun echo 0 'hello '"$SLURM_JOB_ID"'' :\\
-     echo 1 'hello '"$SLURM_JOB_ID"'' :\\
-     echo
+srun --output=slurm-"$SLURM_JOB_ID"-a-0.out echo 0 'hello '"$SLURM_JOB_ID"'' :\\
+     --output=slurm-"$SLURM_JOB_ID"-a-1.out echo 1 'hello '"$SLURM_JOB_ID"'' :\\
+     --output=slurm-"$SLURM_JOB_ID"-b-0.out echo
 """,
         )
 
@@ -205,33 +226,43 @@ def test_cancel(self, run: MagicMock, describe: MagicMock) -> None:
     @patch("subprocess.run")
     def test_describe_completed(self, run: MagicMock) -> None:
         run.return_value.stdout = b"""
-JobID|JobName|Partition|Account|AllocCPUS|State|ExitCode|
-176+0|echo-echo-0|compute||1|COMPLETED|0:0|
-176+0.batch|batch|||1|COMPLETED|0:0|
-176+0.0|echo|||1|COMPLETED|0:0|
-176+1|echo-echo-1|compute||1|COMPLETED|0:0|
-176+1.0|echo|||1|COMPLETED|0:0|
-176+2|echo-echo-2|compute||1|COMPLETED|0:0|
-176+2.0|echo|||1|COMPLETED|0:0|
+JobID|JobName|Partition|Account|AllocCPUS|State|ExitCode
+1853+0|echo-0|compute||1|COMPLETED|0:0
+1853+0.batch|batch|||1|COMPLETED|0:0
+1853+0.0|echo|||1|COMPLETED|0:0
+1853+1|echo-1|compute||1|COMPLETED|0:0
+1853+1.0|echo|||1|COMPLETED|0:0
+1853+2|echo-2|compute||1|COMPLETED|0:0
+1853+2.0|echo|||1|COMPLETED|0:0
 """.strip()
 
         scheduler = create_scheduler("foo")
-        out = scheduler.describe(app_id="176")
+        out = scheduler.describe(app_id="1853")
 
         self.assertEqual(run.call_count, 1)
         self.assertEqual(
             run.call_args,
             call(
-                ["sacct", "--parsable2", "-j", "176"],
+                ["sacct", "--parsable2", "-j", "1853"],
                 stdout=subprocess.PIPE,
                 check=True,
             ),
         )
 
         self.assertIsNotNone(out)
-        self.assertEqual(out.app_id, "176")
+        self.assertEqual(out.app_id, "1853")
         self.assertEqual(out.msg, "COMPLETED")
         self.assertEqual(out.state, specs.AppState.SUCCEEDED)
+        self.assertEqual(
+            out.roles,
+            [
+                specs.Role(
+                    name="echo",
+                    image="",
+                    num_replicas=3,
+                )
+            ],
+        )
 
     @patch("subprocess.run")
     def test_describe_running(self, run: MagicMock) -> None:
@@ -253,3 +284,13 @@ def test_describe_running(self, run: MagicMock) -> None:
         self.assertEqual(out.app_id, "54")
         self.assertEqual(out.msg, "RUNNING")
         self.assertEqual(out.state, specs.AppState.RUNNING)
+
+    def test_log_iter(self) -> None:
+        scheduler = create_scheduler("foo")
+
+        with tmp_cwd():
+            with open("slurm-54-echo-1.out", "wt") as f:
+                f.write("hello\nworld\n")
+
+            logs = list(scheduler.log_iter("54", "echo", 1))
+            self.assertEqual(logs, ["hello", "world"])