Add description of shared memory to the docs

docs/generate_documentation.sh

@@ -131,6 +131,9 @@ for f in $(find ../../umbridge/ -name 'README.md'); do
    $NAME
 EOF
 done
+
+# Extra docs. Added manually
+echo "   ../shared_memory" >> docs_output/source/umbridge/index.rst
 echo "   ../mpi_container" >> docs_output/source/umbridge/index.rst
 echo "   ../math_description" >> docs_output/source/umbridge/index.rst
 

docs/source/shared_memory.rst

@@ -0,0 +1,20 @@
+.. _shared-memory:
+
+=============================================
+Client-Server Communication via Shared Memory
+=============================================
+
+UM-Bridge uses HTTP to communicate between an UQ client and a model server
+by default. However, this is not the most performant option (in HPC) as it 
+carries around 1 millisecond of overhead per request. We provide a quicker
+alternative that uses shared memory (RAM) to transfer data. This approach 
+works by using allocated memory buffers in RAM as the transfer medium for 
+UM-Bridge inputs and outputs. To prevent data races in a parallel setting, 
+each memory buffer is tagged with the ID of the thread making the request 
+in the parallel client, and this ID is shared through a JSON entry so that 
+the model can access the correct buffer. 
+
+Currently, only the C++ and Python implementations are supported, and this
+feature is enabled automatically when available. The code will first 
+perform a test transfer using shared memory when the client first connects 
+to the server; it reverts back to HTTP if this test fails.