Restore optimized num nodes.

On titan, the aprun command is now correct but num_nodes was
incorrect due to not using the 'optimized' num nodes that task maker
was programmed to use with aprun long ago. This commit restores
that capability.

[BFB]

cime/utils/python/CIME/aprun.py

@@ -57,16 +57,14 @@ def _get_aprun_cmd_for_case_impl(ntasks, nthreads, rootpes, pstrids,
 
             c2 += 1
 
-    logger.info("total tasks is: %s" % total_tasks)
-
     # make sure all maxt values at least 1
     for c1 in xrange(0, total_tasks):
         if maxt[c1] < 1:
             maxt[c1] = 1
 
     # Compute task and thread settings for batch commands
-    tasks_per_node, task_count, thread_count, max_thread_count, aprun = \
-        0, 1, maxt[0], maxt[0], "aprun"
+    tasks_per_node, task_count, thread_count, max_thread_count, total_node_count, aprun = \
+        0, 1, maxt[0], maxt[0], 0, "aprun"
     for c1 in xrange(1, total_tasks):
         if maxt[c1] != thread_count:
             tasks_per_node = min(pes_per_node, max_tasks_per_node / thread_count)
@@ -83,6 +81,9 @@ def _get_aprun_cmd_for_case_impl(ntasks, nthreads, rootpes, pstrids,
 
             aprun += " -n %d -N %d -d %d %s :" % (task_count, tasks_per_node, thread_count, run_exe)
 
+            node_count = int(math.ceil(float(task_count) / tasks_per_node))
+            total_node_count += node_count
+
             thread_count = maxt[c1]
             max_thread_count = max(max_thread_count, maxt[c1])
             task_count = 1
@@ -99,6 +100,8 @@ def _get_aprun_cmd_for_case_impl(ntasks, nthreads, rootpes, pstrids,
 
     task_per_numa = int(math.ceil(tasks_per_node / 2.0))
 
+    total_node_count += int(math.ceil(float(task_count) / tasks_per_node))
+
     # Special option for Titan with intel compiler
     if machine == "titan" and tasks_per_node > 1:
         aprun += " -S %d" % task_per_numa
@@ -107,13 +110,13 @@ def _get_aprun_cmd_for_case_impl(ntasks, nthreads, rootpes, pstrids,
 
     aprun += " -n %d -N %d -d %d %s " % (task_count, tasks_per_node, thread_count, run_exe)
 
-    return aprun
+    return aprun, total_node_count
 
 ###############################################################################
 def get_aprun_cmd_for_case(case, run_exe):
 ###############################################################################
     """
-    Given a case, construct and return the aprun command
+    Given a case, construct and return the aprun command and optimized node count
     """
     models = case.get_values("COMP_CLASSES")
     ntasks, nthreads, rootpes, pstrids = [], [], [], []

cime/utils/python/CIME/case.py

@@ -132,14 +132,19 @@ def _initialize_derived_attributes(self):
         self.thread_count = env_mach_pes.get_max_thread_count(comp_classes)
         self.tasks_per_node = env_mach_pes.get_tasks_per_node(self.total_tasks, self.thread_count)
         logger.debug("total_tasks %s thread_count %s"%(self.total_tasks, self.thread_count))
-        self.num_nodes = env_mach_pes.get_total_nodes(self.total_tasks, self.thread_count)
+
         self.tasks_per_numa = int(math.ceil(self.tasks_per_node / 2.0))
         smt_factor = max(1,int(self.get_value("MAX_TASKS_PER_NODE") / pes_per_node))
 
         threads_per_node = self.tasks_per_node * self.thread_count
         threads_per_core = 1 if (threads_per_node <= pes_per_node) else smt_factor
         self.cores_per_task = self.thread_count / threads_per_core
 
+        if self.get_value("MACH") == "titan":
+            self.num_nodes = get_aprun_cmd_for_case(self, "acme.exe")[1]
+        else:
+            self.num_nodes = env_mach_pes.get_total_nodes(self.total_tasks, self.thread_count)
+
     # Define __enter__ and __exit__ so that we can use this as a context manager
     # and force a flush on exit.
     def __enter__(self):
@@ -1094,7 +1099,7 @@ def get_mpirun_cmd(self, job="case.run"):
 
         # special case for aprun
         if executable == "aprun":
-            return get_aprun_cmd_for_case(self, run_exe) + " " + run_misc_suffix
+            return get_aprun_cmd_for_case(self, run_exe)[0] + " " + run_misc_suffix
         else:
             mpi_arg_string = " ".join(args.values())
 

cime/utils/python/CIME/case_setup.py

@@ -163,7 +163,7 @@ def _case_setup_impl(case, caseroot, clean=False, test_mode=False, reset=False):
             # create batch files
             logger.info("Creating batch script case.run")
             env_batch = case.get_env("batch")
-            num_nodes = env_mach_pes.get_total_nodes(pestot, thread_count)
+            num_nodes = case.num_nodes
             tasks_per_node = env_mach_pes.get_tasks_per_node(pestot, thread_count)
             for job in env_batch.get_jobs():
                 input_batch_script  = os.path.join(case.get_value("MACHDIR"), env_batch.get_value('template', subgroup=job))