Merge pull request #1464 from ESMCI/jgfouca/remove_total_cores

Remove TOTAL_CORES as it was inaccurate in many cases.
TOTAL_CORES was only being used by test_schedule in no-batch mode
in order to determine how many processors in the proc-pool should be
considered to be consumed by a specific test.

Users of other systems, specifically titan, were seeing this value
and wondering why it was so inaccurate.

Test_scheduler will now compute cores on the fly when needed.

COST_PES changed to take spare nodes into account.

Test suite: scripts_regression_tests
Test baseline:
Test namelist changes:
Test status: bit for bit

Fixes [CIME Github issue #]

User interface changes?: Removes TOTAL_CORES

Code review: @jedwards4b

scripts/lib/CIME/XML/env_mach_pes.py

@@ -45,7 +45,8 @@ def get_cost_pes(self, totaltasks, max_thread_count, machine=None):
         figure out the value of COST_PES which is the pe value used to estimate model cost
         """
         pespn = self.get_value("PES_PER_NODE")
-        num_nodes = self.get_total_nodes(totaltasks, max_thread_count)[0]
+        num_nodes, spare_nodes = self.get_total_nodes(totaltasks, max_thread_count)
+        num_nodes += spare_nodes
         # This is hardcoded because on yellowstone by default we
         # run with 15 pes per node
         # but pay for 16 pes per node.  See github issue #518

scripts/lib/CIME/case.py

@@ -824,9 +824,6 @@ def configure(self, compset_name, grid_name, machine_name=None,
                 if pio_typename in ("pnetcdf", "netcdf4p"):
                     self.set_value(key, "netcdf")
 
-        # Set TOTAL_CORES
-        self.set_value("TOTAL_CORES", self.total_tasks * self.cores_per_task )
-
         if input_dir is not None:
             self.set_value("DIN_LOC_ROOT", os.path.abspath(input_dir))
 

scripts/lib/CIME/case_setup.py

@@ -140,9 +140,6 @@ def _case_setup_impl(case, caseroot, clean=False, test_mode=False, reset=False):
             logger.info("Machine/Decomp/Pes configuration has already been done ...skipping")
 
             case.initialize_derived_attributes()
-
-            # Set TOTAL_CORES
-            case.set_value("TOTAL_CORES", case.total_tasks * case.cores_per_task )
         else:
             check_pelayouts_require_rebuild(case, models)
 
@@ -176,9 +173,6 @@ def _case_setup_impl(case, caseroot, clean=False, test_mode=False, reset=False):
                     logger.info("Writing %s script from input template %s" % (job, input_batch_script))
                     env_batch.make_batch_script(input_batch_script, job, case, pestot, tasks_per_node, num_nodes, thread_count)
 
-            # Set TOTAL_CORES
-            case.set_value("TOTAL_CORES", case.total_tasks * case.cores_per_task )
-
             # Make a copy of env_mach_pes.xml in order to be able
             # to check that it does not change once case.setup is invoked
             logger.info("Locking file env_mach_pes.xml")

scripts/lib/CIME/test_scheduler.py

@@ -593,8 +593,18 @@ def _get_procs_needed(self, test, phase, threads_in_flight=None, no_batch=False)
     ###########################################################################
         if phase == RUN_PHASE and (self._no_batch or no_batch):
             test_dir = self._get_test_dir(test)
-            out = run_cmd_no_fail("./xmlquery TOTAL_CORES -value", from_dir=test_dir)
-            return int(out)
+            total_pes = int(run_cmd_no_fail("./xmlquery TOTALPES --value", from_dir=test_dir))
+            threads = eval(run_cmd_no_fail("./xmlquery NTHRDS --value", from_dir=test_dir))
+            max_threads = 0
+            for item in threads:
+                _, comp_threads = item.split(":")
+                comp_threads = int(comp_threads)
+                if comp_threads > max_threads:
+                    max_threads = comp_threads
+
+            max_cores = total_pes * max_threads
+            return max_cores
+
         elif (phase == SHAREDLIB_BUILD_PHASE):
             # Will force serialization of sharedlib builds
             # TODO - instead of serializing, compute all library configs needed and build

scripts/tests/scripts_regression_tests.py

@@ -1385,7 +1385,7 @@ def test_cime_case_force_pecount(self):
             self.assertEqual(case.get_value("NTHRDS_CPL"), 8)
 
             expected_cores = 16 * case.cores_per_task
-            self.assertEqual(case.get_value("TOTAL_CORES"), expected_cores)
+            self.assertEqual(case.get_value("COST_PES"), expected_cores)
 
     ###########################################################################
     def test_cime_case_xmlchange_append(self):

src/drivers/mct/cime_config/config_component.xml

@@ -2037,7 +2037,6 @@
     <desc>Number of instances for each component</desc>
   </entry>
 
-
   <entry id="TOTALPES">
     <type>integer</type>
     <default_value>0</default_value>
@@ -2046,14 +2045,6 @@
     <desc>total number of tasks and threads (setup automatically - DO NOT EDIT)</desc>
   </entry>
 
-  <entry id="TOTAL_CORES">
-    <type>integer</type>
-    <default_value>1</default_value>
-    <group>mach_pes_last</group>
-    <file>env_mach_pes.xml</file>
-    <desc>total number of cores used (setup automatically - DO NOT EDIT)</desc>
-  </entry>
-
   <entry id="MAX_TASKS_PER_NODE">
     <type>integer</type>
     <default_value>0</default_value>