Final submisssion.

Author: Damian Rouson

body.tex

@@ -53,7 +53,7 @@ \subsection{Motivation and Background}
 \end{figure*}
 
 \subsection{Objectives}
-This evaluates alternatives for compiling, linking, and executing one
+This paper evaluates alternatives for compiling, linking, and executing one
 \gls{mini-app} \gls{caf} source code using the following technologies:
 \begin{itemize}
   \item \gls{mpi}~\cite{mpiforum2016mpi} and OpenSHMEM~\cite{openshmem2016} communication layers,
@@ -70,10 +70,7 @@ \subsection{Objectives}
   \item Puts: an image stores data in memory managed by
         another image without the receiving image's involvement.
 \end{enumerate}
-Fortran semantics necessitates that gets block.  Puts are non-blocking.
-
-%Our results will inform our future decisions around the choice of compiler, communication
-%layer, platform, and data access patterns.
+We also explore the performance and scalability of multi- versus many-core processors.
 
 \section{Methodology}
 \subsection{Physics and numerics}
@@ -105,7 +102,7 @@ \subsection{Compilers, runtimes, and hardware}
 Cheyenne uses a Mellanox EDR Infiniband interconnect with a partial 9D Enhanced Hypercube single-plane topology.
 We compiled coarray-\gls{icar} on the \gls{nersc} systems using the Cray Fortran compiler version 8.6.0.  We compiled
 at \gls{ncar} using the \gls{gcc} version 6.3 Fortran front end, which uses the
-OpenCoarrays \gls{abi}~\cite{fanfarillo2014opencoarrays} to support \gls{caf}.  We tested two OpenCoarrays parallel runtime libraries:
+OpenCoarrays \gls{abi}~\cite{fanfarillo2014opencoarrays} to support \gls{caf}.  We tested two parallel runtime libraries that implement the OpenCoarrays \gls{abi}:
 \begin{enumerate}
   \item The default \gls{mpi} library using the SGI MPT \gls{mpi},
   \item The recently released OpenSHMEM library.
@@ -138,7 +135,7 @@ \section{Discussion of Results}
 To aid interpretation, the scaling results are reported as a fraction of the ideal scaling for each machine in the top left.
 These results show that the Cray compiler+system scales better than the gfortran+SGI system,
 with 75\% efficiency at >10k cores, while on Cheyenne, only 55\% of ideal was achieved.
-This also shows that the KNL system scales well out to large core counts (60\% ideal with \num{19200} cores),
+This also shows that the KNL system scales well out to large core counts (60\% of the ideal with \num{19200} cores),
 but that the total runtimes are significantly slower than the equivalent runtimes on Xeons (top right).
 The KNL performance might be improved in the future by implementing OpenMP threaded parallelism within a node.
 

main.pdf

@@ -94,8 +94,8 @@
 We examine the scalability and performance of an open-source, \gls{caf} \gls{mini-app} that solves
 several parallel, numerical algorithms known to dominate the execution of \gls{icar}~\cite{gutmann2016intermediate} model,
 a package developed at the \gls{ncar}.
-The \gls{mini-app} uses standard Fortran 2008, including several Fortran 2008 implementations of the collective
-subroutines defined in the Committee Draft of the upcoming Fortran 2015 standard.  The ability of \gls{caf} to run atop various
+The \gls{mini-app} uses standard Fortran 2008, including one Fortran 2008 implementation of a collective
+subroutine defined in the Committee Draft of the upcoming Fortran 2018 standard.  The ability of \gls{caf} to run atop various
 communication layers and the increasing \gls{caf} compiler availability facilitated evaluating several compilers,
 runtime libraries and hardware platforms.  Results are presented for the GNU and Cray compilers, each of which offers
 different parallel runtime libraries employing one or more communication layers, including \gls{mpi}, OpenSHMEM, and proprietary