Data from "Replicable Parallel Branch and Bound Search" submitted to the
Journal of Parallel and Distributed Computing.

Authors: Blair Archibald, Patrick Maier, Ciaran McCreesh, Robert Stewart
and Phil Trinder

Results are stored in space-separated value (ssv) format.

The plots from the paper may be reproduced using the analysis scripts in
the scripts directory. A Makefile is provided to automate the process
and show the required commands. The scripts depend on R with the ggplot2
dplyr, grid and cowplot packages available.

Setup
=====

All experiments are run on a the University of Glasgow GPG cluster with
the following spec:

1.  20 Nodes (Max 17 used)
2.  16 Intel cores/node (2 * Intel Xeon E5-2640 2GHz)
3.  64Gb RAM per node, i.e. 4Gb RAM/core.
4.  Local Disk: 300Gb/node
5.  10Gb Ethernet Interconnect
6.  Ubuntu 14.04, Linux 3.19

Source code for the application is available at:

http://dx.doi.org/10.5281/zenodo.254088

File Information
================

maxclique.dat, knapsack.dat, tsp.dat
------------------------------------

Main data files for testing the properties in Section 7.

  Field        Type    Desc
  ------------ ------- --------------------------------------------------------------------------------------
  variant      str     Ordered or Unordered - Which skeleton are we using
  benchmark    str     instance name
  hosts        int     Number of distributed hosts we run on
  processes    int     Number of OS Processes we spawn. Round robin distributed over hosts
  schedulers   int     Number of *schedulers* assigned to HdpH. i.e Worker threads
  cores        int     Number of cores assigned to the GHC runtime
  workers      int     Total number of workers used (schedulers * processes)
  runtime      float   Time in seconds to perform the search (no setup time) as reported by the application

maxclique-sequential-haskell.dat
--------------------------------

Haskell Data for comparison with state of the art C++ maximum clique
implementation (Table 6).

Uses the Ordered skeleton with a single worker and no spawning to
achieve an ordering that matches sequential while also using the generic
skeleton routines.

  Field       Type    Desc
  ----------- ------- -----------------
  benchmark   str     instance name
  runtime     float   Time in seconds

maxclique-sequential-cpp.dat
----------------------------

C++ Data for maximum clique sequential comparison (Table 6).

  Field       Type    Desc
  ----------- ------- -----------------
  benchmark   str     instance name
  runtime     float   Time in seconds