Remove GPUs detection

Author: Blanca Fuentes Monjas

autodetect_config.py

@@ -1,11 +1,9 @@
 import reframe.core.config as config
 
 site_configuration = config.detect_config(
-    exclude_feats=['row*', 'c*-*', 'group*',
-                   'contbuild', 'startx', 'perf', 'cvmfs', 'gpumodedefault',
-                   'gpu'],
-    detect_containers=True,
-    detect_devices=False,
-    sched_options=['-A csstaff'],
-    filename='daint_config'
+    exclude_feats=['colum*'],
+    detect_containers=False,
+    sched_options=[],
+    time_limit=200,
+    filename='system_config'
 )

reframe/core/config.py

@@ -336,7 +336,7 @@ def sources(self):
     def subconfig_system(self):
         return self._local_system
 
-    def load_config_python(self, filename):
+    def load_config_python(self, filename, validate=True):
         try:
             mod = util.import_module_from_file(filename)
         except ImportError as e:
@@ -346,13 +346,12 @@ def load_config_python(self, filename):
                 f'could not load Python configuration file: {filename!r}'
             ) from e
 
-        print(dir(mod))
         if not hasattr(mod, 'site_configuration'):
             raise ConfigError(
                 f"not a valid Python configuration file: '{filename}'"
             )
 
-        if mod.site_configuration:
+        if validate:
             self._config_modules.append(mod)
             self.update_config(mod.site_configuration, filename)
 
@@ -669,7 +668,7 @@ def find_config_files(config_path=None, config_file=None):
     return res
 
 
-def load_config(*filenames):
+def load_config(*filenames, validate=True):
     ret = _SiteConfig()
     getlogger().debug('Loading the builtin configuration')
     ret.update_config(settings.site_configuration, '<builtin>')
@@ -681,7 +680,7 @@ def load_config(*filenames):
         getlogger().debug(f'Loading configuration file: {f!r}')
         _, ext = os.path.splitext(f)
         if ext == '.py':
-            ret.load_config_python(f)
+            ret.load_config_python(f, validate)
         elif ext == '.json':
             ret.load_config_json(f)
         else:
@@ -691,12 +690,11 @@ def load_config(*filenames):
 
 
 def detect_config(detect_containers: bool = False,
-                  detect_devices: bool = False,
                   exclude_feats: list = [],
                   sched_options: list = [],
-                  filename: str = 'system_config'
-                  ):
-    # job_cli_options, exclude_feats
+                  time_limit: int = 200,
+                  filename: str = 'system_config'):
+
     import reframe.core.runtime as rt
 
     # Initialize the Site Configuration object
@@ -710,7 +708,8 @@ def detect_config(detect_containers: bool = False,
     site_config.setdefault('name', '')
     site_config.setdefault('hostnames', [])
     hostname = ret._detect_system(detect_only=True)
-    site_config['hostnames'] += hostname
+    site_config['hostnames'] += [hostname]
+    print(site_config['hostnames'])
     site_config['name'] += hostname
     msg = color.colorize(
         f'Detected hostname: {hostname}', color.GREEN
@@ -737,15 +736,11 @@ def detect_config(detect_containers: bool = False,
     msg = color.colorize(f'Launcher set to {launcher_name}', color.GREEN)
     getlogger().info(msg)
 
-    # Initialize the RuntimeContext with builtin settings
-    rt.init_runtime(ret)
-
     site_config.setdefault('partitions', [])
     # Detect the context with the corresponding scheduler
     site_config['partitions'] = scheduler().build_context(
-        modules_system, launcher(), sched_options,
-        exclude_feats, rt.runtime().prefix,
-        detect_containers, detect_devices
+        modules_system, launcher(), sched_options, time_limit,
+        exclude_feats, rt.runtime().prefix, detect_containers
     )
 
     template_loader = FileSystemLoader(searchpath=os.path.join(

reframe/core/schedulers/__init__.py

@@ -757,13 +757,13 @@ class ReframeContext(abc.ABC):
     '''
 
     def __init__(self, modules_system, launcher, scheduler, prefix,
-                 detect_containers, detect_devices):
+                 time_limit, detect_containers):
         self.partitions = []
         self._modules_system = modules_system
         self._scheduler = scheduler
         self._launcher = launcher
+        self._time_limit = time_limit
         self._detect_containers = detect_containers
-        self._detect_devices = detect_devices
         self._p_n = 0  # System partitions counter
         self._keep_tmp_dir = False
         if prefix == '.':
@@ -775,7 +775,7 @@ def __init__(self, modules_system, launcher, scheduler, prefix,
                 prefix='reframe_config_detection_',
                 dir=prefix
             )
-        if detect_containers or detect_devices:
+        if detect_containers:
             getlogger().info(f'Stage directory: {self.TMP_DIR}')
 
     @abc.abstractmethod
@@ -812,9 +812,6 @@ def _check_gpus_count(self, node_devices_slurm: dict,
         for gpu_job in node_devices_job:
             devices.append({'type': 'gpu',
                             'model': gpu_job,
-                            # TODO
-                            'arch': c_d.nvidia_gpu_architecture.get(gpu_job) or
-                                    c_d.amd_gpu_architecture.get(gpu_job),
                             'num_devices': node_devices_job[gpu_job]})
             gpus_job_count += node_devices_job[gpu_job]
 
@@ -928,9 +925,6 @@ def _extract_info(self, job: Job):
         if job.detect_containers:
             job.container_platforms = self._parse_containers(file_path)
 
-        if job.detect_devices:
-            job.devices = self._parse_devices(file_path)
-
     def _create_detection_job(self, name: str, access_node: list,
                               access_options: list):
         '''Create the instance of the job for remote autodetection'''
@@ -942,8 +936,7 @@ def _create_detection_job(self, name: str, access_node: list,
             sched_access=access_node,
             sched_options=access_options
         )
-        # TODO: move this somewhere to be user defined, not here (maybe yaml)
-        remote_job.max_pending_time = 20
+        remote_job.max_pending_time = self._time_limit
         remote_job.time_limit = '2m'
         remote_job.container_platforms = []
         remote_job.devices = {}
@@ -954,14 +947,10 @@ def _generate_job_content(self, job):
         if job.detect_containers:
             job.content += [c_d.containers_detect_bash]
             job.content += ['\n\n\n']
-        if job.detect_devices:
-            job.content += [c_d.devices_detect_bash]
 
     def create_login_partition(self):
-        # TODO: move this somewhere to be user defined, not here (maybe yaml)
         max_jobs = 4
         time_limit = '2m'
-        # TODO: improve this (?)
         self.partitions.append(
             {'name':      'login',
              'scheduler':  'local',
@@ -971,34 +960,30 @@ def create_login_partition(self):
              'launcher':   'local'})
 
     def create_remote_partition(self, node_feats: tuple,
-                                job_cli_options):
+                                sched_options):
 
         node_features = list(node_feats)
-        _detect_devices = copy.deepcopy(self._detect_devices)
         _detect_containers = copy.deepcopy(self._detect_containers)
         self._p_n += 1  # Count the partition that is being created
-        access_options = copy.deepcopy(job_cli_options)
+        access_options = copy.deepcopy(sched_options)
         access_node = self._scheduler.feats_access_option(node_features)
         name = f'partition_{self._p_n}'
         getlogger().info(f'{name} : {node_feats}')
-        # TODO: move this somewhere to be user defined, not here (maybe yaml)
         max_jobs = 100
         time_limit = '10m'
         container_platforms = []
-        devices = []
 
+        # Try to get the devices from the scheduler config
+        _detect_devices = self._find_devices(node_features)
         if _detect_devices:
-            # If detection of remote devices is requested,
-            # try to get the devices from the scheduler config
-            _detect_devices = self._find_devices(node_features)
+            getlogger().info('GPUs were detected in this node type.')
 
         remote_job = None
-        if _detect_devices or _detect_containers:
+        if _detect_containers:
             self._keep_tmp_dir = True
             remote_job = self._create_detection_job(
                 name, access_node, access_options
             )
-            remote_job.detect_devices = _detect_devices
             remote_job.detect_containers = _detect_containers
             self._generate_job_content(remote_job)
             submission_error, access_node = self.submit_detect_job(
@@ -1041,22 +1026,8 @@ def create_remote_partition(self, node_feats: tuple,
                     f'partition "{name}"'
                 )
 
-            if remote_job.devices:
-                # Issue any warning regarding missconfigurations
-                # between Gres and the detected devices
-                getlogger().info(f"\nGPUs found in partition {name}")
-                devices = self._check_gpus_count(
-                    _detect_devices, remote_job.devices
-                )
-
-        elif not remote_job:
-            # No jobs were launched so we cannot check the access options
-            access_options += access_node
-        else:
-            # TODO check this
-            access_options = access_node
+        access_options += access_node
 
-        # TODO: improve this (?)
         # Create the partition
         self.partitions.append(
             {'name':      name,
@@ -1069,14 +1040,13 @@ def create_remote_partition(self, node_feats: tuple,
              'launcher':   self._launcher.name,
              'access':     access_options,
              'features':   node_features+['remote'],
-             'devices':    devices,
              'container_platforms': container_platforms}
         )
 
-    def create_partitions(self, job_cli_options):
+    def create_partitions(self, sched_options):
         # TODO: asynchronous
         for node in self.node_types:
-            self.create_remote_partition(node, job_cli_options)
+            self.create_remote_partition(node, sched_options)
         if not self._keep_tmp_dir:
             shutil.rmtree(self.TMP_DIR)
         else:

reframe/core/schedulers/slurm.py

@@ -586,17 +586,15 @@ def validate(cls) -> bool:
             return False
 
     def build_context(self, modules_system, launcher,
-                      job_cli_options, exclude_feats,
-                      prefix, detect_containers,
-                      detect_devices):
+                      sched_options, time_limit,
+                      exclude_feats, prefix, detect_containers):
         self._context = _SlurmContext(
-            modules_system, launcher, self, prefix,
-            detect_containers=detect_containers,
-            detect_devices=detect_devices
+            modules_system, launcher, self, prefix, time_limit,
+            detect_containers=detect_containers
         )
         self._context.search_node_types(exclude_feats)
         self._context.create_login_partition()
-        self._context.create_partitions(job_cli_options)
+        self._context.create_partitions(sched_options)
         return self._context.partitions
 
 
@@ -728,8 +726,7 @@ def satisfies(self, slurm_constraint: str):
         # Convert the Slurm constraint to a Python expression and evaluate it,
         # but restrict our syntax to accept only AND or OR constraints and
         # their combinations
-        # print(re.match(r'^[\w\d\(\)\|\&]*$', slurm_constraint))
-        if not re.match(r'^[\w\d\(\)\|\&]*$', slurm_constraint):
+        if not re.match(r'^[\w\d\(\)\|\&\-]*$', slurm_constraint):
             return False
 
         names = {grp[0]
@@ -741,16 +738,11 @@ def satisfies(self, slurm_constraint: str):
                         ] else key
             for key in slurm_constraint.split()])
 
-        # print(expr)
-        # print('mc' in self.active_features and '128gb' in self.active_features)
         vars = {n: True for n in self.active_features}
-        # print(vars)
         vars.update({n: False for n in names - self.active_features})
-        # print(eval(expr, {}, vars))
         try:
-            # print("returning", eval(expr, {}, vars))
             return eval(expr, {}, {'vars': vars})
-        except BaseException as e:
+        except BaseException:
             return False
 
     @property
@@ -792,12 +784,12 @@ def __str__(self):
 class _SlurmContext(sched.ReframeContext):
 
     def __init__(self, modules_system: str, launcher: str, scheduler: str,
-                 prefix: str, detect_containers: bool = True,
-                 detect_devices: bool = True, access_opt: list = []):
+                 prefix: str, time_limit: int, detect_containers: bool = True,
+                 access_opt: list = []):
 
         super().__init__(modules_system, launcher,
-                         scheduler, prefix,
-                         detect_containers, detect_devices)
+                         scheduler, prefix, time_limit,
+                         detect_containers)
         self.node_types = []
         self.default_nodes = []
         self.reservations = []
@@ -950,11 +942,11 @@ def _find_devices(self, node_feats: list) -> Union[dict, None]:
     def _get_access_partition(self, node_feats: list) -> Union[str, None]:
 
         nodes_info = self._scheduler.allnodes()
+        node_feats = "&".join(node_feats)
         nd_partitions = {tuple(n.partitions)
                          for n in nodes_info
                          if n.satisfies(node_feats)}
         nd_partitions = set(nd_partitions)
-        print(nd_partitions)
         if len(nd_partitions) > 1:
             return None
         else:

reframe/frontend/cli.py

@@ -879,7 +879,10 @@ def restrict_logging():
         printer.adjust_verbosity(calc_verbosity(site_config, options.quiet))
 
     if options.detect_configuration:
-        site_config = config.load_config(options.detect_configuration)
+        runtime.init_runtime(site_config)
+        site_config = config.load_config(
+            options.detect_configuration, validate=False
+        )
         sys.exit(0)
 
     # Now configure ReFrame according to the user configuration file

reframe/schemas/reframe_config_template.j2

@@ -39,21 +39,6 @@ site_configuration = {
                     'extras': {{partition.extras}},
                     # https://reframe-hpc.readthedocs.io/en/stable/config_reference.html#config.systems.partitions.env_vars
                     'env_vars': {{partition.env_vars}},
-                    {% if partition.devices %}
-                    # Check if any specific devices were found in this node type
-                    # The gpus found in slurm GRes will be specified here
-                    'devices': [
-                    {% for dev in partition.devices %}
-                        { 'type': '{{dev.type}}',
-                          'model': '{{dev.model}}',
-                        {% if dev.arch %}
-                          'arch': '{{dev.arch}}',
-                        {% endif %}
-                          'num_devices': {{dev.num_devices}}
-                                },
-                    {% endfor %}
-                    ],
-                    {% endif %}
                     {% if partition.container_platforms %}
                     # Check if any container platforms are available in these nodes and add them
                     # https://reframe-hpc.readthedocs.io/en/stable/config_reference.html#container-platform-configuration