SW-824: Resolve SW-824

Closes SW-824

LCS/PyCommon/cep4_utils.py

@@ -53,7 +53,9 @@ def wrap_command_in_cep4_available_cpu_node_with_lowest_load_ssh_call(cmd, via_h
     :param bool via_head: when True, route the cmd first via the cep4 head node
     :return: the same subprocess cmd list, but then wrapped with cep4 ssh calls
     '''
-    lowest_load_node_nr = get_cep4_available_cpu_node_with_lowest_load()
+    lowest_load_node_nr = get_cep4_cpu_node_with_lowest_load()
+    if lowest_load_node_nr is None:
+        raise RuntimeError("No cpu node available to run the cmd on. cmd=%s" %(cmd,))
     return wrap_command_in_cep4_cpu_node_ssh_call(cmd, lowest_load_node_nr, via_head=via_head)
 
 def wrap_command_in_cep4_cpu_node_ssh_call(cmd, cpu_node_nr, via_head=True):
@@ -99,29 +101,50 @@ def wrap_command_for_docker(cmd, image_name, image_label='', mount_dirs=['/data'
     dockerized_cmd += cmd
     return dockerized_cmd
 
-def get_cep4_available_cpu_nodes(include_allocated_nodes=False):
+# a selection of slurm states relevant for lofar usage
+SLURM_AVAILABLE_STATES = {'idle','mix'}
+SLURM_IN_USES_STATES = {'alloc','mix'}
+SLURM_UNAVAILABLE_STATES = {'down','drain','drng','resv','maint'}
+SLURM_STATES = SLURM_AVAILABLE_STATES | SLURM_IN_USES_STATES | SLURM_UNAVAILABLE_STATES
+
+# lofar cep4 default slurm partitions
+SLURM_CPU_PARTITION = 'cpu'
+SLURM_GPU_PARTITION = 'gpu'
+SLURM_PARTITIONS = {SLURM_CPU_PARTITION, SLURM_GPU_PARTITION}
+
+def get_cep4_slurm_nodes(states, partition: str) -> []:
     '''
-    get a list of cep4 cpu nodes which are currently up and running according to slurm
-    :return: a list of cpu node numbers (ints) for the up and running cpu nodes
+    get a list of cep4 nodes from the given partition which have (one of) the given states according to slurm
+    states: set/list of slurm states, see predefined sets: SLURM_AVAILABLE_STATES, SLURM_IN_USES_STATES, SLURM_UNAVAILABLE_STATES
+    partition: one of the known SLURM_PARTITIONS
+    :return: a list of node numbers (ints) of the nodes in the partition
     '''
-    available_cep4_nodes = []
+    slurm_nodes = []
 
     try:
-        logger.debug('determining available cep4 cpu nodes')
+        # filter out unknown states
+        _states = [s for s in states if s in SLURM_STATES]
+        if not _states:
+            raise ValueError("the given states:%s are not valid slurm states:%s" % (states, SLURM_STATES))
+
+        if partition not in SLURM_PARTITIONS:
+            raise ValueError("the given partition:%s is not one of the valid cep4 slurm partition:%s" % (partition, SLURM_PARTITIONS))
+
+        logger.debug('determining available cep4 nodes states:%s, partition:%s', _states, partition)
 
         # find out which nodes are available
-        cmd = ['sinfo -p cpu -t idle,mix%s' % (',alloc' if include_allocated_nodes else '')]
+        cmd = ['sinfo -p %s -t %s' % (partition, ','.join(_states))]
         cmd = wrap_command_in_cep4_head_node_ssh_call(cmd)
 
         logger.debug('executing command: %s', ' '.join(cmd))
         out = check_output_returning_strings(cmd)
         lines = out.split('\n')
-        for state in ['idle', 'mix', 'alloc']:
+        for state in _states:
             try:
                 line = next(l for l in lines if state in l).strip()
                 # get nodes string part of line:
                 nodes_part = line.split(' ')[-1]
-                available_cep4_nodes += convert_slurm_nodes_string_to_node_number_list(nodes_part)
+                slurm_nodes += convert_slurm_nodes_string_to_node_number_list(nodes_part)
 
             except StopIteration:
                 pass  # no line with state in line
@@ -129,13 +152,34 @@ def get_cep4_available_cpu_nodes(include_allocated_nodes=False):
     except Exception as e:
         logger.exception(e)
 
-    available_cep4_nodes = sorted(list(set(available_cep4_nodes)))
-    logger.debug('available cep4 cpu nodes: %s', ','.join(str(x) for x in available_cep4_nodes))
+    slurm_nodes = sorted(list(set(slurm_nodes)))
+    logger.debug('cep4 nodes with states:%s in partition:%s are: %s', _states, partition, slurm_nodes)
+    return slurm_nodes
+
+def get_cep4_available_cpu_nodes():
+    '''
+    get a list of cep4 cpu nodes which are currently up and running  (either totally free, or partially free) according to slurm
+    :return: a list of cpu node numbers (ints)
+    '''
+    available_cep4_nodes = get_cep4_slurm_nodes(SLURM_AVAILABLE_STATES, SLURM_CPU_PARTITION)
+
     if not available_cep4_nodes:
         logger.warning('no cep4 cpu nodes available')
 
     return available_cep4_nodes
 
+def get_cep4_up_and_running_cpu_nodes():
+    '''
+    get a list of cep4 cpu nodes which are currently up and running (either totally free, partially free, or totally used) according to slurm
+    :return: a list of cpu node numbers (ints)
+    '''
+    cep4_nodes = get_cep4_slurm_nodes(SLURM_AVAILABLE_STATES | SLURM_IN_USES_STATES, SLURM_CPU_PARTITION)
+
+    if not cep4_nodes:
+        logger.warning('no cep4 cpu nodes up and running')
+
+    return cep4_nodes
+
 def convert_slurm_nodes_string_to_node_number_list(slurm_string):
     ''' converts strings like: cpu[01-03,11-12]' to [1,2,3,11,12]
     or 'cpu01' to [1]
@@ -168,15 +212,15 @@ def convert_slurm_nodes_string_to_node_number_list(slurm_string):
         result.append(node)
     return result
 
-def get_cep4_cpu_nodes_loads(node_nrs=None, normalized=False, include_allocated_nodes=False):
+def get_cep4_cpu_nodes_loads(node_nrs=None, normalized=False):
     '''
     get the 5min load for each given cep4 cpu node nr
-    :param node_nrs: optional list of node numbers to get the load for. If None, then all available nodes are queried.
+    :param node_nrs: optional list of node numbers to get the load for. If None, then all up-and-running nodes are queried.
     :param bool normalized: when True, then normalize the loads with the number of cores.
     :return: dict with node_nr -> load mapping
     '''
     if node_nrs == None:
-        node_nrs = get_cep4_available_cpu_nodes(include_allocated_nodes=include_allocated_nodes)
+        node_nrs = get_cep4_up_and_running_cpu_nodes()
 
     procs = {}
     loads = {}
@@ -224,17 +268,17 @@ def get_cep4_cpu_nodes_loads(node_nrs=None, normalized=False, include_allocated_
                 ', '.join('cpu%02d:%.3f' % (nr, loads[nr]) for nr in sorted(loads.keys())))
     return loads
 
-def get_cep4_available_cpu_nodes_sorted_ascending_by_load(max_normalized_load=0.33, min_nr_of_nodes=0, node_nrs=None, include_allocated_nodes=False):
+def get_cep4_available_cpu_nodes_sorted_ascending_by_load(max_normalized_load=0.33, min_nr_of_nodes=0, node_nrs=None):
     '''
     get the cep4 available cpu node numbers sorted ascending by load (5min).
     :param float max_normalized_load: filter available nodes which are at most max_normalized_load
     :param int min_nr_of_nodes: do return this minimum number of nodes, even if their load is higher than max_normalized_load
                                 If not enough nodes are up, then of course it cannot be guaranteed that we return this amount.
-    :param list node_nrs: optional list of node numbers to apply the filtering on. If None, then all available nodes are queried.
+    :param list node_nrs: optional list of node numbers to apply the filtering on. If None, then all up-and-running nodes are queried.
     :return: sorted list of node numbers.
     '''
     if not node_nrs:
-        node_nrs = get_cep4_available_cpu_nodes(include_allocated_nodes=include_allocated_nodes)
+        node_nrs = get_cep4_up_and_running_cpu_nodes()
     loads = get_cep4_cpu_nodes_loads(node_nrs, normalized=True)
     load_tuples_list = [(cpu_nr,load) for cpu_nr,load in list(loads.items())]
     sorted_load_tuples_list = sorted(load_tuples_list, key=lambda x: x[1])
@@ -249,15 +293,21 @@ def get_cep4_available_cpu_nodes_sorted_ascending_by_load(max_normalized_load=0.
                  max_normalized_load, min_nr_of_nodes, sorted_node_nrs)
     return sorted_node_nrs
 
-def get_cep4_available_cpu_node_with_lowest_load(max_normalized_load=0.33, include_allocated_nodes=False):
+def get_cep4_cpu_node_with_lowest_load(max_normalized_load=0.33):
     '''
-    get the cep4 cpu node which is available and has the lowest (5min) load of them all.
-    :param float max_normalized_load: filter available nodes which a at most max_normalized_load
+    get the cep4 cpu node which has the lowest (5min) load of them all. Preferably a node which is not fully used yet.
+    :param float max_normalized_load: filter nodes which a at most max_normalized_load
     :return: the node number (int) with the lowest load.
     '''
+    # first see if there are any not-fully-used nodes available
+    nodes = get_cep4_available_cpu_nodes()
+
+    if not nodes: # if not, then just query all up and running nodes
+        nodes = get_cep4_up_and_running_cpu_nodes()
+
     node_nrs = get_cep4_available_cpu_nodes_sorted_ascending_by_load(max_normalized_load=max_normalized_load,
                                                                      min_nr_of_nodes=1,
-                                                                     include_allocated_nodes=include_allocated_nodes)
+                                                                     node_nrs=nodes)
     if node_nrs:
         logger.debug('cep4 cpu node with lowest load: %s', node_nrs[0])
         return node_nrs[0]