From cae6b7311a71150fd198a25bcc8c8387ff8602e8 Mon Sep 17 00:00:00 2001
From: Joseph H Kennedy <me@jhkennedy.org>
Date: Tue, 21 May 2019 13:20:36 -0400
Subject: [PATCH] Set default NTEST_PARALLEL_JOBS=MAX_MPITASKS_PER_NODE and
 limit E3SM machines

This resets the default value of NTEST_PARALLEL_JOBS to MAX_MPITASKS_PER_NODE
so as to not make any behavioral changes to CESM.

Warning: This is not a safe value on machine with batch systems who's
login nodes are more limited than the compute nodes and therefore
NTEST_PARALLEL_JOBS should be set on these systems.

@jgfouca found via E3SM testing that limiting to 4 parallel jobs was
required for many of the testing machines with batch systems to prevent
hammering login nodes. Therefore, we set that value for these E3SM
machines:
* cori-haswell
* cori-knl
* blues
* anvil
* bebop
* theta
* titan
* summit

Warning: Non test machines for E3SM that have a batch system may still
oversubscribe parallel test jobs.
---
 config/e3sm/machines/config_machines.xml | 8 ++++++++
 scripts/lib/CIME/test_scheduler.py       | 2 +-
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/config/e3sm/machines/config_machines.xml b/config/e3sm/machines/config_machines.xml
index f5a08077164..ba9f6f95cbb 100644
--- a/config/e3sm/machines/config_machines.xml
+++ b/config/e3sm/machines/config_machines.xml
@@ -218,6 +218,7 @@
     <CCSM_CPRNC>/project/projectdirs/acme/tools/cprnc.cori/cprnc</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_developer</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>nersc_slurm</BATCH_SYSTEM>
     <SUPPORTED_BY>e3sm</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>32</MAX_TASKS_PER_NODE>
@@ -356,6 +357,7 @@
     <CCSM_CPRNC>/project/projectdirs/acme/tools/cprnc.cori/cprnc</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_developer</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>nersc_slurm</BATCH_SYSTEM>
     <SUPPORTED_BY>e3sm</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>128</MAX_TASKS_PER_NODE>
@@ -1038,6 +1040,7 @@
     <CCSM_CPRNC>/home/ccsm-data/tools/cprnc</CCSM_CPRNC>
     <GMAKE_J>4</GMAKE_J>
     <TESTS>e3sm_integration</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>pbs</BATCH_SYSTEM>
     <SUPPORTED_BY>acme</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>16</MAX_TASKS_PER_NODE>
@@ -1141,6 +1144,7 @@
     <CCSM_CPRNC>/lcrc/group/acme/tools/cprnc/cprnc</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_integration</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>slurm</BATCH_SYSTEM>
     <SUPPORTED_BY>E3SM</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>36</MAX_TASKS_PER_NODE>
@@ -1255,6 +1259,7 @@
     <CCSM_CPRNC>/lcrc/group/acme/tools/cprnc/cprnc</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_integration</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>slurm</BATCH_SYSTEM>
     <SUPPORTED_BY>E3SM</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>36</MAX_TASKS_PER_NODE>
@@ -1570,6 +1575,7 @@
     <CCSM_CPRNC>/projects/ccsm/acme/tools/cprnc/cprnc</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_developer</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>cobalt_theta</BATCH_SYSTEM>
     <SUPPORTED_BY>E3SM</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>128</MAX_TASKS_PER_NODE>
@@ -2177,6 +2183,7 @@
     <CCSM_CPRNC>/lustre/atlas1/cli900/world-shared/cesm/tools/cprnc/cprnc.titan</CCSM_CPRNC>
     <GMAKE_J>8</GMAKE_J>
     <TESTS>e3sm_developer</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>pbs</BATCH_SYSTEM>
     <ALLOCATE_SPARE_NODES>TRUE</ALLOCATE_SPARE_NODES>
     <SUPPORTED_BY>E3SM</SUPPORTED_BY>
@@ -3065,6 +3072,7 @@
     <CCSM_CPRNC>/gpfs/alpine/cli115/world-shared/e3sm/tools/cprnc.summit/cprnc</CCSM_CPRNC>
     <GMAKE_J>32</GMAKE_J>
     <TESTS>e3sm_developer</TESTS>
+    <NTEST_PARALLEL_JOBS>4</NTEST_PARALLEL_JOBS>
     <BATCH_SYSTEM>lsf</BATCH_SYSTEM>
     <SUPPORTED_BY>e3sm</SUPPORTED_BY>
     <MAX_TASKS_PER_NODE>84</MAX_TASKS_PER_NODE>
diff --git a/scripts/lib/CIME/test_scheduler.py b/scripts/lib/CIME/test_scheduler.py
index e2f8bf1a830..bd22f97f27c 100644
--- a/scripts/lib/CIME/test_scheduler.py
+++ b/scripts/lib/CIME/test_scheduler.py
@@ -194,7 +194,7 @@ def __init__(self, test_names, test_data=None,
         if parallel_jobs is None:
             mach_parallel_jobs = self._machobj.get_value("NTEST_PARALLEL_JOBS")
             if mach_parallel_jobs is None:
-                mach_parallel_jobs = 3
+                mach_parallel_jobs = self._machobj.get_value("MAX_MPITASKS_PER_NODE")
             self._parallel_jobs = min(len(test_names), mach_parallel_jobs)
         else:
             self._parallel_jobs = parallel_jobs