This code provides an out-of-core large data sort implementation for
distributed memory systems.  It is tailored to read input from disk
that is generated via the gensort utility:

www.ordinal.com/gensort.html

----------------------
Software Requirements
----------------------

Building the code requires the following:

(1) Autotools (autoconf, automake, etc.)
(2) MPI compiler for C++ with OpenMP support
(3) GRVY, https://red.ices.utexas.edu/projects/software/wiki/GRVY
(4) Boost C++ headers, http://www.boost.org
(5) sort_dist (the underlying sort utility code). A copy of the
    sort_dist distribution that has been run on Cray and InfiniBand,
    Lustre-based supercomputers is included in the sort_dist/
    subdirectory.

The code base uses an autotools based build system. A basic
configuration and build should be achievable using something similar
to the following:

$ ./bootstrap
$ ./configure --with-grvy=<GRVY-install-path> --with-boost=<Boost-install-path> --with-parsort=sort_dist
$ make -j 4

See ./configure --help for more information.

----------------
Runtime Control
----------------

Runtime configuration is controlled via a keyword driven input file
(see input.dat for an example). Use this to control the location of
input files, desired paths for temporary and final output, number of
dedicated IO readers, number of threads per sort host,
etc. Alternatively, several runtime options can be specified directly
as command-line arguments:

./testdev <input-file> <numfiles> <numIO_hosts> <num_threads> <num_sort_groups>

-----------
References 
-----------

More information on the algorithm and example performance measurements
obtained are available in the following paper:

Hari Sundar, Dhairya Malhotra, and Karl W. Schulz. 2013. "Algorithms
for high-throughput disk-to-disk sorting". 

In Proceedings of SC13: International Conference for High Performance
Computing, Networking, Storage and Analysis (SC '13). ACM, New York,
NY, USA, , Article 93 , 10 pages. 

http://doi.acm.org/10.1145/2503210.2503259