Cloned SEACAS for EXODUS library with extra build files for internal package management.
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

438 lines
16 KiB

<!-------- @HEADER
!
! !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
!
! Zoltan Toolkit for Load-balancing, Partitioning, Ordering and Coloring
! Copyright 2012 Sandia Corporation
!
! Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
! the U.S. Government retains certain rights in this software.
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions are
! met:
!
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
!
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
!
! 3. Neither the name of the Corporation nor the names of the
! contributors may be used to endorse or promote products derived from
! this software without specific prior written permission.
!
! THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
! EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
! IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
! CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
! EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
! PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
! PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
! LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
! NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
! SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
!
! Questions? Contact Karen Devine kddevin@sandia.gov
! Erik Boman egboman@sandia.gov
!
! !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
!
! @HEADER
------->
<!DOCTYPE html PUBLIC "-//w3c//dtd html 4.0 transitional//en">
<html>
<head>
<meta http-equiv="Content-Type"
content="text/html; charset=iso-8859-1">
<meta name="GENERATOR"
content="Mozilla/4.76 [en] (X11; U; Linux 2.4.2-2smp i686) [Netscape]">
<meta name="sandia.approval_type" content="formal">
<meta name="sandia.approved" content="SAND2007-4748W">
<meta name="author" content="Zoltan PI">
<title>Zoltan User's Guide: ParMETIS Interface</title>
</head>
<body bgcolor="#ffffff">
<div align="right"><b><i><a href="ug.html">Zoltan User's Guide</a>&nbsp;
|&nbsp; <a href="ug_alg_ptscotch.html">Next</a>&nbsp; |&nbsp; <a
href="ug_graph_vs_hg.html">Previous</a></i></b></div>
<h2>
<a name="ParMETIS"></a>ParMETIS - Parallel Graph Partitioning</h2>
<a href="https://www-users.cs.umn.edu/%7Ekarypis/metis/parmetis/">ParMETIS</a>
is a parallel library for graph partitioning (for static load
balancing)
and repartitioning (for dynamic load balancing) developed at the
University
of Minnesota by Karypis, Schloegel and Kumar [<a
href="ug_refs.html#parmetis">Karypis
and Kumar</a>]. Note that ParMetis must be obtained separately and has a different license than Zoltan (see page 31 of <a
href="https://glaros.dtc.umn.edu/gkhome/fetch/sw/parmetis/manual.pdf">
this manual</a>). ParMETIS is strictly speaking not a method but
rather a collection of methods. In the Zoltan context, ParMETIS is a
method
with many sub-methods. Zoltan provides an interface to all the ParMETIS
(sub-)methods.&nbsp; The user selects which ParMETIS method to use
through
the parameter PARMETIS_METHOD. Most of the ParMETIS methods are based
on
either multilevel Kernighan-Lin partitioning or a diffusion algorithm.
The names of the ParMETIS methods used by Zoltan are identical to those
in the ParMETIS library. For further information about the various <a
href="https://www-users.cs.umn.edu/%7Ekarypis/metis/parmetis/">ParMETIS</a>
methods and parameters, please consult the <a
href="https://www-users.cs.umn.edu/%7Ekarypis/metis/parmetis/">ParMETIS</a>
User's Guide.
<p>Graph partitioning is a useful abstraction for load balancing. The
main
idea is to represent the computational application as a weighted graph.
The nodes or vertices in the graph correspond to objects in
Zoltan.&nbsp;
Each object may have a weight that normally represents the amount of
computation.
The edges or arcs in the graph usually correspond to communication
costs.
In graph partitioning, the problem is to find a partition of the
graph
(that is,&nbsp; each vertex is assigned to one out of <i>k</i>
possible
sets called parts) that minimizes the cut size (weight) subject to
the parts having approximately equal size (weight). In
repartitioning,
it is assumed that a partition already exists. The problem is to
find
a good partition that is also "similar" in some sense to the
existing
partition. This keeps the migration cost low. All the problems
described
above are NP-hard so no efficient exact algorithm is known, but
heuristics work well in practice.<br>
</p>
<p>We give only a brief summary of the various ParMETIS methods here;
for
more details see the <a
href="https://www-users.cs.umn.edu/%7Ekarypis/metis/parmetis/">ParMETIS</a>
documentation. The methods fall into three categories:
</p>
<ol>
<li>
Part* - Perform graph partitioning without consideration of the initial
distribution. (LB_APPROACH=partition)<br>
</li>
<li>
AdaptiveRepart - Incremental
algorithms with small migration cost. (LB_APPROACH=repartition)</li>
<li>
Refine* - Refines a given partition (balance).&nbsp; Can be applied
multiple times to reduce the communication cost (cut weight) if
desired. (LB_APPROACH=refine)</li>
</ol>
As a rule of thumb, use one of the Part* methods if you have a poor
initial
balance and you are willing to spend some time doing migration. One
such
case is static load balancing; that is, you need to balance only once.
Use AdaptiveRepart or the Repart* methods when you have a reasonably
good
load balance that you wish to update incrementally. These methods are
well
suited&nbsp; for dynamic load balancing (for example,&nbsp; adaptive
mesh
refinement). A reasonable strategy is to call PartKway once to obtain a
good initial balance and&nbsp; later update this balance using
AdaptiveRepart.
<p>
Zoltan supports the multiconstraint partitioning
feature of ParMETIS through multiple object weights (see <i><a
href="ug_param.html#General_Parameters">OBJ_WEIGHT_DIM</a></i>).
</p>
<p>The graph given to Zoltan/ParMETIS must be symmetric. Any self edges
(loops) will be ignored. Multiple edges between a pair of vertices is
not allowed. All weights must be non-negative. The graph does not have to
be connected.
</p>
<p>Zoltan is currently compatible with ParMETIS version 3.1 and 4.0.x.
The ParMETIS source code can be obtained from
the <a href="https://glaros.dtc.umn.edu/gkhome/views/metis/parmetis">
ParMETIS home page</a>. ParMETIS has a separate license:
<i>'PARMETIS can be freely used for educational and research purposes by non-profit institutions and US government agencies only. Other organizations are allowed to
use PARMETIS only for evaluation purposes, and any further uses will require prior approval from <a href="https://www.license.umn.edu/Products/ParMETIS---Mesh-Graph-Partitioning-Algorithm__Z09041.aspx">the technology transfer office at the University of Minnesota</a>'
</i> <br>
If you do not wish to install ParMETIS, it
is possible to compile Zoltan without any references to ParMETIS;
(when you 'make' Zoltan, comment out the PARMETIS_LIBPATH variable in
the
configuration file <i><a href="ug_usage.html#Building%20the%20Library">Utilities/Config/Config.&lt;platform&gt;</a></i>).
</p>
<br>
&nbsp;
<table nosave="" width="100%">
<tbody>
<tr>
<td valign="top"><b>Value of LB_METHOD:</b></td>
<td><b>GRAPH</b></td>
</tr>
<tr>
<td valign="top"><b>Value of GRAPH_PACKAGE:</b></td>
<td><b>Parmetis</b></td>
</tr>
<tr>
<td><b>Parameters:</b></td>
<td><br>
</td>
</tr>
<tr>
<td valign="top"><i>&nbsp;&nbsp;&nbsp; LB_APPROACH<br>
</i></td>
<td>The load balancing approach:. <br>
<i>PARTITION</i>&nbsp; - partition from scratch, not taking
the current data distribution into account<br>
<i>REPARTITION</i>&nbsp; - partition but try to stay close to the
current partition/distribution
<br>
<i>REFINE</i>&nbsp; - refine the current partition/distribution;
assumes only small changes
</td>
</tr>
<tr>
<td valign="top"><i>&nbsp;&nbsp;&nbsp; PARMETIS_METHOD</i></td>
<td>The specific ParMETIS method to be used (see below).
Note: See also <a href="ug_alg.html#LB_APPROACH">LB_APPROACH</a>,
which is a simpler way to specify the overall load balance approach.
Only use <i>PARMETIS_METHOD</i> if you really need a specific
implementation.<br>
<i>PartKway</i> - multilevel Kernighan-Lin partitioning&nbsp;
<br>
<i>PartGeom</i> - space filling curves (coordinate based)&nbsp;
<br>
<i>PartGeomKway</i> - hybrid method based on PartKway and
PartGeom
(needs both&nbsp; graph data and coordinates)&nbsp;
<br>
<i>AdaptiveRepart - </i>adaptive repartitioning (only in
ParMETIS 3.0
and higher)
<br>
<i>RefineKway</i> - refine the current partition (balance)</td>
</tr>
<tr>
<td><br>
</td>
<td>The method names are case insensitive.</td>
</tr>
<tr>
<td valign="top">&nbsp;&nbsp;&nbsp; <i>PARMETIS_OUTPUT_LEVEL</i></td>
<td>Amount of output the load-balancing algorithm should
produce.&nbsp;
<br>
0 = no output, 1 = print timing info. Turning on more bits displays
more information (for example, 3=1+2, 5=1+4, 7=1+2+4).</td>
</tr>
<tr>
<td>&nbsp;&nbsp;&nbsp; <i>PARMETIS_COARSE_ALG</i></td>
<td>Coarse algorithm for PartKway. 1 = serial, 2 = parallel.
(ParMETIS
2 only)</td>
</tr>
<tr>
<td>&nbsp;&nbsp;&nbsp; <i>PARMETIS_SEED</i></td>
<td>Random seed for ParMETIS.</td>
</tr>
<tr nosave="" valign="top">
<td>&nbsp;&nbsp;&nbsp; <i>PARMETIS_ITR</i></td>
<td nosave="">Ratio of interprocessor communication time to
redistribution
time. A high value will emphasize reducing the edge cut, while a small
value will try to keep the change in the new partition (distribution)
small.
This parameter is used only by AdaptiveRepart. A value of between 100
and
1000 is good for most problems.</td>
</tr>
<tr nosave="" valign="top">
<td>&nbsp;&nbsp;&nbsp; <i>CHECK_GRAPH</i></td>
<td nosave="">Level of error checking for graph input: 0 = no
checking, 1
= on-processor checking, 2 = full checking. (CHECK_GRAPH==2 is very
slow
and should be used only during debugging).</td>
</tr>
<tr nosave="" valign="top">
<td>&nbsp;&nbsp;&nbsp; <i>SCATTER_GRAPH</i></td>
<td nosave="">Scatter graph data by distributing contiguous
chunks of objects
(vertices) of roughly equal size to each processor before calling the
partitioner.
0 = don't scatter;
1 = scatter only if all objects are on a single processor;
2 = scatter if at least one processor owns no objects;
3 = always scatter.&nbsp;</td>
</tr>
<tr nosave="" valign="top">
<td>&nbsp;&nbsp;&nbsp; <i>GRAPH_SYMMETRIZE</i></td>
<td nosave="">How to symmetrize the graph:
NONE = graph is symmetric and no symmetrization is needed <br/>
TRANSPOSE = if M is adjacency matrix of the input graph, output will be the graph representation of M+M<sup>T</sup> <br/>
BIPARTITE = graph is symmetrized in a bipartite way : [[ 0 M ][M<sup>t</sup> 0]]
</td>
</tr>
<tr nosave="" valign="top">
<td>&nbsp;&nbsp;&nbsp; <i>GRAPH_SYM_WEIGHT</i></td>
<td nosave="">How edge weights are handled during symmetrization:
ADD = weights of each arc are added <br/>
MAX = only the heaviest arc weight is kept <br/>
<p>See more informations about graph build options on this <a href="ug_graph_build.html">page</a></p>
<!-- ERROR = fail if complementary arcs don't have the same weight. -->
</td>
</tr>
<!-- <tr nosave="" valign="top"> -->
<!-- <td>&nbsp;&nbsp;&nbsp; <i>GRAPH_BIPARTITE_TYPE</i></td> -->
<!-- <td nosave=""> In the case of a bipartite symmetrization, -->
<!-- NONE = graph is symmetric and no symmetrization is needed <br/> -->
<!-- TRANSPOSE = if M is adjacency matrix of the input graph, output will be the graph representation of M+M<sup>T</sup> <br/> -->
<!-- BIPARTITE = graph is symmetrized in a bipartite way : [[ 0 M ][M<sup>t</sup> 0]] -->
<!-- </td> -->
<!-- </tr> -->
<tr>
<td valign="top"><b>Default values:</b></td>
<td><br>
</td>
</tr>
<tr>
<td><br>
</td>
<td><i>LB_APPROACH</i> = Repartition</td>
</tr>
<tr>
<td><br>
</td>
<td><i>PARMETIS_METHOD</i> = AdaptiveRepart</td>
</tr>
<tr>
<td><br>
</td>
<td><i>PARMETIS_OUTPUT_LEVEL</i> = 0</td>
</tr>
<tr>
<td><br>
</td>
<td><i>PARMETIS_COARSE_ALG </i>= 2</td>
</tr>
<tr>
<td><br>
</td>
<td><i>PARMETIS_SEED </i>= 15</td>
</tr>
<tr>
<td><br>
</td>
<td><i>PARMETIS_ITR </i>= 100</td>
</tr>
<tr>
<td><br>
</td>
<td><i>USE_OBJ_SIZE </i>= 1</td>
</tr>
<tr>
<td><br>
</td>
<td><i>CHECK_GRAPH</i> = 1</td>
</tr>
<tr>
<td><br>
</td>
<td><i>SCATTER_GRAPH </i>= 1</td>
</tr>
<tr>
<td><br>
</td>
<td><i>GRAPH_SYMMETRIZE </i>= NONE</td>
</tr>
<tr>
<td><br>
</td>
<td><i>GRAPH_SYM_WEIGHT </i>= ADD</td>
</tr>
<tr>
<td valign="top"><b>Required Query Functions:</b></td>
<td><br>
</td>
</tr>
<tr>
<td>For all submethods:</td>
<td><b><a href="ug_query_lb.html#ZOLTAN_NUM_OBJ_FN">ZOLTAN_NUM_OBJ_FN</a></b></td>
</tr>
<tr>
<td><br>
</td>
<td><b><a href="ug_query_lb.html#ZOLTAN_OBJ_LIST_FN">ZOLTAN_OBJ_LIST_FN</a></b>
</td>
</tr>
<tr>
<td>Only PartGeom &amp; PartGeomKway:</td>
<td><b><a href="ug_query_lb.html#ZOLTAN_NUM_GEOM_FN">ZOLTAN_NUM_GEOM_FN</a></b></td>
</tr>
<tr>
<td><br>
</td>
<td>
<b><a href="ug_query_lb.html#ZOLTAN_GEOM_MULTI_FN">ZOLTAN_GEOM_MULTI_FN</a></b>
or <b><a href="ug_query_lb.html#ZOLTAN_GEOM_FN">ZOLTAN_GEOM_FN</a></b>
</td>
</tr>
<tr nosave="" valign="top">
<td>All but PartGeom:</td>
<td nosave="">
<b><a href="ug_query_lb.html#ZOLTAN_NUM_EDGES_MULTI_FN">ZOLTAN_NUM_EDGES_MULTI_FN</a></b>
or
<b><a href="ug_query_lb.html#ZOLTAN_NUM_EDGES_FN">ZOLTAN_NUM_EDGES_FN</a></b>
<br>
<b><a href="ug_query_lb.html#ZOLTAN_EDGE_LIST_MULTI_FN">ZOLTAN_EDGE_LIST_MULTI_FN</a></b>
or
<b><a href="ug_query_lb.html#ZOLTAN_EDGE_LIST_FN">ZOLTAN_EDGE_LIST_FN</a></b>
</td>
</tr>
<tr>
<td valign="top"><b>Optional Query Functions:</b></td>
<td><br>
</td>
</tr>
<tr>
<td><br>
</td>
<td>
<b><a href="ug_query_mig.html#ZOLTAN_OBJ_SIZE_MULTI_FN">ZOLTAN_OBJ_SIZE_MULTI_FN</a></b>
or
<b><a href="ug_query_mig.html#ZOLTAN_OBJ_SIZE_FN">ZOLTAN_OBJ_SIZE_FN</a></b> for <i>PARMETIS_METHOD</i>=<i>AdaptiveRepart</i>
</td>
</tr>
<tr>
<td><br>
</td>
<td>
<b><a href="ug_query_lb.html#ZOLTAN_PART_MULTI_FN">ZOLTAN_PART_MULTI_FN</a></b>
or
<b><a href="ug_query_lb.html#ZOLTAN_PART_FN">ZOLTAN_PART_FN</a></b> for part remapping in ParMETIS.
</td>
</tr>
</tbody>
</table>
</p>
<hr width="100%">[<a href="ug.html">Table of Contents</a>&nbsp; | <a
href="ug_alg_ptscotch.html">Next:&nbsp;
PT-Scotch</a>&nbsp; |&nbsp; <a href="ug_graph_vs_hg.html">Previous:&nbsp;
Graph vs. Hypergraph Partitioning</a>&nbsp; |&nbsp; <a href="https://www.sandia.gov/general/privacy-security/index.html">Privacy and Security</a>]
</body>
</html>