source: trunk/documents/UserDoc/UsersGuides/ForApplicationDeveloper/html/Examples/parallel.html @ 1231

Last change on this file since 1231 was 1208, checked in by garnier, 15 years ago

CVS update
File size: 8.1 KB
Line 
1<html>
2<head>
3<title>Parallel Geant4 (ParGeant4)</title>
4</head>
5
6<body>
7<table WIDTH="100%" >
8<tr>
9<TD><a href="../../../../Overview/html/index.html">
10
11</a><a href="index.html">
12<img SRC="../../../../resources/html/IconsGIF/Contents.gif" ALT="Contents" height=16 width=59></a>
13<a href="ExtendedCodes.html">
14<img SRC="../../../../resources/html/IconsGIF/Previous.gif" ALT="Previous" height=16 width=59></a>
15<img SRC="../../../../resources/html/IconsGIF/NextGR.gif" ALT="Next" height=16 width=59></td>
16
17<td ALIGN=RIGHT><b><font color="#238E23"><font size=-1>Geant4 User's Guide</font></font></b>
18<br><b><font color="#238E23"><font size=-1>For Application Developers</font></font></b>
19<br><b><font color="#238E23"><font size=-1>Examples</font></font></b></td>
20</tr>
21</table>
22
23<center>
24<p><b><font color="#238E23"><font size=+3>Parallel Geant4 (ParGeant4)</font></font></b>
25<address>by Gene Cooperman (gene@ccs.neu.edu)<br>
26         and Viet Ha Nguyen (vietha@ccs.neu.edu)
27</address>
28</center>
29<p>
30<hr ALIGN="Center" SIZE="7%">
31<h2>What is ParGeant4 ?</h2>
32  <a target="ext" href="http://www.ccs.neu.edu/home/gene/pargeant4.html">ParGeant4</a>
33  is a parallel version of Geant4 that implements event-level
34  parallelism to simulate separate events on remote processors.
35  Typical simulations demonstrate a nearly linear speedup in running
36  time as the number of remote processors increases.  The needed
37  enhancements of Geant4 are included in the
38  examples/extended/parallel directory of the Geant4 distribution.
39</p>
40<p>
41<hr ALIGN="Center" SIZE="7%">
42<h2>Why is ParGeant4 useful?</h2>
43
44  When doing a large Geant4 simulation, one often wishes to run
45  on many processors to reduce the overall time.  Traditionally,
46  this has been done by splitting the events into multiple groups,
47  and running Geant4 independently on each processor for its own
48  group of events.  This requires restarting a run if a processor
49  goes down.  It also requires saving the histogram files from
50  each run, and merging the files prior to using the analysis tool.
51  The human effort in this is considerable.
52<p>
53  ParGeant4 provides a much simpler mechanism.  After setting up
54  ParGeant4 one links and runs the sequential Geant4 application
55  exactly as before, but additionally linking with some parallel libraries.
56  Upon execution, ParGeant4 on the console sends out events to
57  slave processes, collects all hits, and calls any analysis tool --
58  exactly as one would do in the sequential case.
59</p><p>
60  There is no need to split events into separate groups, track
61  whether one of the processors crashed, merge histogram files, etc.
62  If a slave processor crashes, ParGeant automatically re-sends the
63  events of that slave processor to a new slave processor for re-execution.
64</p>
65<p>
66<hr ALIGN="Center" SIZE="7%">
67<h2>What is the performance of ParGeant4?</h2>
68
69  As a rule of thumb, speedup will be nearly linear when each event
70  simulation lasts for at least several milliseconds.
71  ParGeant4 has been tested extensively on parallelizations of
72  examples/novice/N02 and of examples/advanced/underground_physics .
73  On N02, we see a speedup of 27 for 50 nodes and a speedup of
74  33 for 100 nodes.  When using the <code>--aggregated-tasks=50</code>
75  option (see below)
76  the speedup improves to 35 for 50 nodes and 60 for 100 nodes.
77<p>
78  In tests of underground_physics, events are longer and we see nearly
79  linear speedup (94 times speedup with 100 nodes).
80</p>
81<p>
82<hr ALIGN="Center" SIZE="7%">
83<h2>Getting started </h2>
84
85 Detailed information is under extended/parallel/ParN02/docs/000README.
86 There are four steps:
87 <ol>
88 <li> Install <a target="ext" href="http://www.ccs.neu.edu/home/gene/topc.html">TOP-C</a>.</li>
89 <li> Compile ParN02 by running <TT>gmake</TT>.</li>
90 <li> Make sure the "procgroup" file is correct and copy it to
91      directory of the executable binary file (for example,
92      <TT>$G4BIN/Linux-g++</TT>).</li>
93 <li> Run the parallel binary program.</li>
94 </ol> 
95</p>
96<p>
97<hr ALIGN="Center" SIZE="7%">
98<h2>What is involved in setting up ParGeant4?</h2>
99
100  To set up ParGeant4, one needs
101  <a target="ext" href="http://www.ccs.neu.edu/home/gene/topc.html">TOP-C</a> and
102  <a target="ext" href="http://www.ccs.neu.edu/home/gene/marshalgen.html">Marshalgen</a>
103  (free, open source software).
104  If one is parallelizing a new Geant4 application, one must then add/modify
105  approximately 20 lines of annotations (C++ comments
106  to indicate shallow vs. deep copying of pointers, etc.) in the .h files
107  for each hit type being defined
108  by the application. For details of the annotations, refer to the manual of the Marshalgen package.<br>
109
110  Finally, in the main routine of the application,
111  one replaces the call to the <TT>G4RunManager</TT> constructor by a call
112  to the <TT>ParRunManager</TT> constructor.  (<TT>ParRunManger</TT> is a derived class
113  of <TT>G4RunManager</TT>.)</p>
114<p>
115  After this, one invokes the already provided
116  GNUMakefile (a slightly modified version of the Geant4 example GNUMakefile)
117  to create the parallel application.  Finally, one writes a
118  "procgroup" file, which declares the names of the remote hosts
119  to use in the parallel computation.  Optionally, one may also
120  specify filenames (e.g. slave1.out, slave2.out, ...) to store
121  the printout from each slave process.  One then calls the
122  ParGeant4 binary exactly as one would call the Geant4 binary,
123  and the results appear as normal, only faster.
124</p>
125<p>
126<hr ALIGN="Center" SIZE="7%">
127<h2>Are there examples of using ParGeant4?</h2>
128  Yes.  ParGeant4 includes parallelizations of
129  other examples from the Geant4 distribution.  Specifically, ParGeant4
130  includes example parallelizations of novice/N02, novice/N04, and
131  advanced/underground_physics .
132</p>
133<p>
134<hr ALIGN="Center" SIZE="7%">
135<h2>What are some of the features of ParGeant4?</h2>
136  ParGeant4 includes all of the features of TOP-C.  In particular,
137  after building a binary, "parMySimulation",
138  one might call:<p>
139</p><dl><dt>
140    ./parMySimulation --TOPC-help
141</dt><dd>
142       Display command options, and then exit.
143<p>
144</p></dd><dt>
145    ./parMySimulation --TOPC-trace=0 <geant4 arguments="">
146</geant4></dt><dd>
147       By default, ParGeant4 traces each time a new event is sent to
148       a task.  This turns it off.
149<p>
150</p></dd><dt>
151    ./parMySimulation --TOPC-verbose=0 <geant4 arguments="">
152</geant4></dt><dd>
153       By default, ParGeant4 provides statistics indicating what
154       process was run, when it was run, what machine, the running
155       times and elapsed times of master and the average slave, and
156       other information.  This turns off the statistics.
157<p>
158</p></dd><dt>
159    ./parMySimulation --TOPC-aggregated-tasks=10 <geant4 arguments="">
160</geant4></dt><dd>
161       By default, ParGeant4 sends one event to one remote process before
162       turning to the next process.  This option sends 10 events
163       to a single remote process in one message.  This is useful
164       when events are relatively short, and the network latency
165       of sending a message is starting to dominate the running time.
166<p>
167</p></dd><dt>
168    ./parMySimulation --TOPC-slave-timeout=3600 <geant4 arguments="">
169</geant4></dt><dd>
170       By default, if a remote process has not communicated with the
171       master after 1800 seconds (a half hour), the slave process will
172       kill itself.  This prevents runaway processes that may be in an
173       infinite loop for some event, or may have lost their socket to
174       communicate with the master process.  In this example, we allow
175       7200 seconds (two hours) because we expect simulation of some
176       events to last up to (but not more than) 7200 seconds.
177</dd></dl>
178<p>
179  By default, ParGeant4 uses its own subset implementation of MPI (MPINU).
180  ParGeant4 adds approximately 50 KB to the "footprint" of the binary
181  executable.  By default, ParGeant4 uses "ssh" to set up remote processes.
182  Those who wish to use their own MPI (perhaps if a batch cluster requires
183  a specific MPI) may do so.  (See "Configuring a Different `MPI')"
184  in the <a target="ext" href="http://www.ccs.neu.edu/home/gene/topc/topc_toc.html">
185  TOP-C manual</a>.)
186
187</p>
188<hr><i><a href="../../../../Authors/html/subjectsToAuthors.html">About
189the authors</a></i>
190</body>
191</html>
Note: See TracBrowser for help on using the repository browser.