Added 'Other Recommendations' to requirements page. Included details …

…about ranking and binding processes
geoschem · Aug 9, 2021 · a2c05cd · a2c05cd
1 parent 874f9d1
commit a2c05cd
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 2 deletions.
diff --git a/docs/requirements.txt b/docs/requirements.txt
@@ -2,4 +2,5 @@
 sphinx
 sphinx_rtd_theme
 sphinxcontrib-bibtex 
-recommonmark
+recommonmark
+docutils==0.16   # temporary fix for unordered lists not rendering properly
diff --git a/docs/source/getting-started/requirements.rst b/docs/source/getting-started/requirements.rst
@@ -70,7 +70,6 @@ Big Compute Recommendations
 These hardware recommendations are for users that are interested in tackling large bleeding-edge
 computational problems:
 
-* Hyper-threading can improve simulation throughput (reportedly up to 50%)
 * A high-performance-computing cluster (or a cloud-HPC service like AWS)
 
    * 1--50 nodes
@@ -79,3 +78,11 @@ computational problems:
 
 * Lots of storage. Several TB is sufficient, but tens or hundreds of TB is better.
 
+Other Recommendations
+^^^^^^^^^^^^^^^^^^^^^
+
+* Hyper-threading may improve simulation throughput, particularly at low core counts
+* MPI process should be bound sequentially across cores and nodes (e.g., a simulation with 48-processes with 24 processes per node 
+  should bind rank 0 to CPU L#0, rank 1 to CPU L#1, etc. on the first node, and rank 24 to CPU L#0, rank 1 to CPU L#1, etc. on the 
+  second node). This should be the default, but it's worth checking if your performance is lower than expected. With OpenMPI the
+  `--report-bindings` argument will show you how processes are ranked and binded.