Wavefront User's Manual

Last edited: June 2013

Wavefront is Copyright (C) 2010- The University of Notre Dame.
All rights reserved.
This software is distributed under the GNU General Public License.
See the file COPYING for details.

Introduction

Wavefront( array R[x,0], array R[0,y], function F(x,y,d) )
returns matrix R where
R[i,j] = F( R[i-1,j], R[i,j-1], R[i-1,j-1] )
The Wavefront abstraction computes a two dimensional recurrence relation. You provide a function F that accepts the left (x), right (y), and diagonal (d). The abstraction then runs each of the functions in the order of dependency, handling load balancing, data movement, fault tolerance, and so on.

To use Wavefront, install the Cooperative Computing Tools and run the program named wavefront_master. You need to set up an input file with initial boundary values of R, and provide a function that computes a new cell from adjacent cells. Each line of the input file has the format: row column arg1,arg2,... In which row and column describe the position of a cell, with zero-based indices, and arg1,arg2,... is the list of arguments fed to F (that is, each R[i,j] is actually a tuple). To run wavefront, specifying the program that computes each cell, and the number of cells in each dimension, for example: wavefront_master ./func.exe 10 10 input.data output.data in which input.data is the file with the initial boundary conditions, and output.data are the results computed for a 10x10 matrix. The program func.exe may be written in any language that you like. For each cell of the result, the program will be invoked like this: ./func.exe x y x-file y-file d-file in which each of x-file, y-file and d-file are files that contain the data from the x, y, and diagonal neighbors, in the format: arg1,arg2,... These files are generated automatically by wavefront_master. Your function is required to parse them, and to print the result of the current cell to stdout, in the format: arg1,arg2,... wavefront_master is a Work Queue application, thus it does not perform any work by itself, but it relies on workers connecting to it. To launch a single worker: work_queue_worker mymachine.somewhere.edu 9123 in which 9123 is the default port for Work Queue applications. Work Queue provides convenient alternatives to launch many workers simultaneously for different batch systems (e.g. condor), which are explained in the section Project names in the Work Queue manual . Wavefront will check for a few error conditions, and then start to run, showing progress on the console like this: # elapsed time : waiting jobs / running jobs / complete jobs (percent complete) 0 : 0 / 1 / 0 (%0.00) 5 : 0 / 2 / 1 (%1.23) 10 : 0 / 3 / 3 (%3.70) 16 : 0 / 4 / 6 (%7.41) 21 : 0 / 4 / 8 (%9.88) ... When complete, your outputs will be stored in the output file specified, with the same format as for the input data. Here is a graph of a 100 by 100 problem run on a 64-core machine, where each F takes about five seconds to execute:

The -B option will write a bitmap progress image every second. Each pixel represents the state of one cell in the matrix: green indicates complete, blue currently running, yellow ready to run, and red not ready. Here is an example of the progress of a small ten by ten job using five CPUs:

Note that at the broadest part of the workload, there are not enough CPUs to run all cells at once, so some must wait. Also note that the wave does not need to run synchronously: cells may begin to compute as soon as their dependencies are satisfied.

For More Information

For the latest information about Wavefront, please visit our web site and subscribe to our mailing list.