METIS Manual

METIS Manual
M ETIS∗
A Software Package for Partitioning Unstructured
Graphs, Partitioning Meshes, and Computing
Fill-Reducing Orderings of Sparse Matrices
Version 5.0
George Karypis
Department of Computer Science & Engineering
University of Minnesota
Minneapolis, MN 55455
[email protected]
August 4, 2011
Metis [MEE tis]: ‘Metis’ is the Greek word for wisdom. Metis was a titaness in Greek mythology. She was the consort
of Zeus and the mother of Athena. She presided over all wisdom and knowledge.
∗ M ETIS is copyrighted by the regents of the University of Minnesota.
http://www.cs.umn.edu/˜karypis
1
Related papers are available via WWW at URL:
Contents
1
Introduction
2
What is new in version 5.0
2.1 Changes in the command-line programs
2.1.1 Migration issues . . . . . . . .
2.2 Changes in the API routines . . . . . .
2.2.1 Migration issues . . . . . . . .
3
4
5
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Overview of M ETIS
3.1 Partitioning a graph . . . . . . . . . . . . . . . . . . . . . . .
3.2 Alternate partitioning objectives . . . . . . . . . . . . . . . .
3.3 Support for multi-phase and multi-physics computations . . .
3.4 Partitioning a mesh . . . . . . . . . . . . . . . . . . . . . . .
3.5 Partitioning for heterogeneous parallel computing architectures
3.6 Computing a fill-reducing ordering of a sparse matrix . . . . .
3.7 Converting a mesh into a graph . . . . . . . . . . . . . . . . .
M ETIS’ stand-alone programs
4.1 Input file formats . . . . . . . . . .
4.1.1 Graph file . . . . . . . . . .
4.1.2 Mesh file . . . . . . . . . .
4.1.3 Target partition weights file
4.2 Output file formats . . . . . . . . .
4.2.1 Partition file . . . . . . . . .
4.2.2 Ordering file . . . . . . . .
4.3 Programs . . . . . . . . . . . . . .
gpmetis . . . . . . . . . . .
mpmetis . . . . . . . . . . .
ndmetis . . . . . . . . . . .
m2gmetis . . . . . . . . . .
graphchk . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
5
5
5
5
.
.
.
.
.
.
.
6
7
8
8
8
9
9
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
9
10
12
12
12
12
12
13
15
16
17
18
M ETIS’ API
5.1 Header files . . . . . . . . . . . . . . . .
5.2 Use of NULL parameters . . . . . . . . .
5.3 C/C++ and Fortran Support . . . . . . . .
5.4 Options array . . . . . . . . . . . . . . .
5.5 Graph data structure . . . . . . . . . . . .
5.6 Mesh data structure . . . . . . . . . . . .
5.7 Partitioning objectives . . . . . . . . . .
5.8 Graph partitioning routines . . . . . . . .
METIS PartGraphRecursive .
METIS PartGraphKway . . . . .
5.9 Mesh partitioning routines . . . . . . . .
METIS PartMeshDual . . . . .
METIS PartMeshNodal . . . . .
5.10 Sparse Matrix Reordering Routines . . . .
METIS NodeND . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
19
19
19
19
19
22
23
23
25
25
25
27
27
29
30
30
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
5.11 Mesh-to-graph conversion routines . . .
METIS MeshToDual . . . . . .
METIS MeshToNodal . . . . .
5.12 Utility routines . . . . . . . . . . . . .
METIS SetDefaultOptions
METIS Free . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
31
32
33
33
33
6
System requirements and contact information
34
7
Copyright & license notice
34
3
1
Introduction
Algorithms that find a good partitioning of highly unstructured graphs are critical for developing efficient solutions for
a wide range of problems in many application areas on both serial and parallel computers. For example, large-scale
numerical simulations on parallel computers, such as those based on finite element methods, require the distribution
of the finite element mesh to the processors. This distribution must be done so that the number of elements assigned
to each processor is the same, and the number of adjacent elements assigned to different processors is minimized.
The goal of the first condition is to balance the computations among the processors. The goal of the second condition
is to minimize the communication resulting from the placement of adjacent elements to different processors. Graph
partitioning can be used to successfully satisfy these conditions by first modelling the finite element mesh by a graph,
and then partitioning it into equal parts.
Graph partitioning algorithms are also used to compute fill-reducing orderings of sparse matrices. These fillreducing orderings are useful when direct methods are used to solve sparse systems of linear equations. A good
ordering of a sparse matrix dramatically reduces both the amount of memory as well as the time required to solve
the system of equations. Furthermore, the fill-reducing orderings produced by graph partitioning algorithms are particularly suited for parallel direct factorization as they lead to high degree of concurrency during the factorization
phase.
Graph partitioning is also used for solving optimization problems arising in numerous areas such as design of very
large scale integrated circuits (VLSI), storing and accessing spatial databases on disks, transportation management,
and data mining.
2
What is new in version 5.0
Version 5.0 represents nearly a complete re-write of the code-base whose purpose was to streamline and unify the
stand-alone programs and API, provide better support for 64 bit architectures, enhance its functionality, and reduce
the memory requirements by re-factoring its internal memory management routines. As a result, both the stand-alone
programs and API routines have changed, making it incompatible with the earlier versions of M ETIS. However, in
order to minimize the code changes that the revised API will require, the new API relies heavily on reasonable default
values for most of the new parameters that it introduced.
The following represents a list of some of the major functionality-related changes and enhancements that are accessible by both the command-line programs and the API routines.
• Multi-constraint partitioning can be used in conjunction with minimization of the total communication volume.
• All graph and mesh partitioning routines take as input the target sizes of the partitions, which among others,
allow them to compute partitioning solutions that are well-suited for parallel architectures with heterogeneous
computing capabilities.
• When multi-constraint partitioning is used, the target sizes of the partitions are specified on a per partitionconstraint pair.
• The multilevel k-way partitioning algorithms can compute a partitioning solution in which each partition is
contiguous.
• All partitioning and ordering routines can compute multiple different solutions and select the best as the final
solution.
• The mesh partitioning and mesh-to-graph conversion routines can operate on mixed element meshes.
• The command-line programs provide full access to the entire set of capabilities provided by M ETIS’ API.
4
Operation
4.x stand-alone program
5.x stand-alone program
Partition a graph
pmetis
kmetis
gpmetis
Partition a mesh
partnmesh
partdmesh
mpmetis
Compute a fill-reducing
ordering of a sparse matrix
oemetis
onmetis
ndmetis
Convert a mesh into a graph
mesh2nodal
mesh2dual
m2gmetis
Graph format checker
graphchk
graphchk
Table 1: Mapping between the old 4.x and the new 5.x command-line programs.
2.1
Changes in the command-line programs
Table 1 shows how the v4.x command-line programs map to the new set of command-line programs provided by
the 5.0 version of M ETIS. As part of these changes, none of the functionality offered by the old programs has been
removed. On the contrary, the new programs have been extended to substantially increase the type of partitionings
and orderings that can be computed by them. This was achieved by expanding the number of optional parameters that
these programs can take, which allow users to have a complete access to all of M ETIS’ functionality. Prior to this
release, if users wanted to access some of the advance features of M ETIS, they had to write their own programs based
on the supplied API.
2.1.1
Migration issues
In order to support the enhanced functionality offered by the new command-line programs, the format of the graph
mesh files has changed (Section 4.1). In the case of the graph file, the new format is backwards compatible, so graphs
written for the earlier format will work correctly when used by gpmetis and ndmetis. However, the new mesh file
format is not backward compatible, and as such they should not be used with mpmetis and m2gmetis. Fortunately,
they can be easily converted to the new format by a slightly modifying their header line.
2.2
Changes in the API routines
Table 2 shows how the v4.x API routines map to the new set of APIs provided by the 5.0 version of M ETIS. The
number of distinct core functions has been reduced to seven by expanding the calling sequence of the new routines
to provide the functionally offered by the old specialized routines. In most cases, the functionality provided by the
new API routines is a superset of that offered by the old routines, especially in the areas related to partitioning for
heterogeneous computing architectures, multi-constraint partitioning, and communication volume-based partitioning
objectives.
2.2.1
Migration issues
Since the calling sequence of all the API routines and in some cases their names has changed, migrating to the new
API will require code modifications. To ensure that these modifications are minimal, the new API routines allow users
to provide NULL as the argument to many of the parameters for which there are reasonable defaults. Thus, we expect
the migration to the new API will be rather straightforward, as long as the application does not want to take advantage
of the newly added capabilities.
5
Operation
4.x routine
5.x routine
Partition a graph
METIS PartGraphRecursive
METIS mCPartGraphRecursive
METIS WPartGraphRecursive
METIS PartGraphRecursive
METIS
METIS
METIS
METIS
METIS
METIS PartGraphKway
PartGraphKway
mCPartGraphKway
WPartGraphKway
PartGraphVKway
WPartGraphVKway
Partition a mesh
METIS PartMeshNodal
METIS PartMeshDual
METIS PartMeshNodal
METIS PartMeshDual
Compute a fill-reducing
ordering of a sparse matrix
METIS EdgeND
METIS NodeND
METIS NodeWND
METIS NodeND
Convert a mesh into a
graph
METIS MeshToNodal
METIS MeshToDual
METIS MeshToNodal
METIS MeshToDual
Utility routines
METIS EstimateMemory
Deprecated
New utility routines
METIS SetDefaultOptions
METIS Free
Table 2: Mapping between the old 4.x and the new 5.x API.
3
Overview of M ETIS
M ETIS is a serial software package for partitioning large irregular graphs, partitioning large meshes, and computing
fill-reducing orderings of sparse matrices. M ETIS has been developed at the Department of Computer Science &
Engineering at the University of Minnesota and is freely distributed. Its source code can downloaded directly from
http://www.cs.umn.edu/˜metis, and is also included in numerous software distributions for Unix-like operating systems
such as Linux and FreeBSD.
The algorithms implemented in M ETIS are based on the multilevel graph partitioning paradigm [4, 3, 2], which
has been shown to quickly produce high-quality partitionings and fill-reducing orderings. The multilevel paradigm,
illustrated in Figure 1, consists of three phases: graph coarsening, initial partitioning, and uncoarsening. In the graph
coarsening phase, a series of successively smaller graphs is derived from the input graph. Each successive graph is
constructed from the previous graph by collapsing together a maximal size set of adjacent pairs of vertices. This
process continues until the size of the graph has been reduced to just a few hundred vertices. In the initial partitioning
phase, a partitioning of the coarsest and hence, smallest, graph is computed using relatively simple approaches such
as the algorithm developed by Kernighan-Lin [5]. Since the coarsest graph is usually very small, this step is very fast.
Finally, in the uncoarsening phase, the partitioning of the smallest graph is projected to the successively larger graphs
by assigning the pairs of vertices that were collapsed together to the same partition as that of their corresponding collapsed vertex. After each projection step, the partitioning is refined using various heuristic methods to iteratively move
vertices between partitions as long as such moves improve the quality of the partitioning solution. The uncoarsening
phase ends when the partitioning solution has been projected all the way to the original graph.
M ETIS uses novel approaches to successively reduce the size of the graph as well as to refine the partition during
the uncoarsening phase. During coarsening, M ETIS employs algorithms that make it easier to find a high-quality
partition at the coarsest graph. During refinement, M ETIS focuses primarily on the portion of the graph that is close
to the partition boundary. These highly tuned algorithms allow M ETIS to quickly produce high-quality partitions and
6
Operation
Stand-alone program
API routine
Partition a graph
gpmetis
METIS PartGraphRecursive
METIS PartGraphKway
Partition a mesh
mpmetis
METIS PartMeshNodal
METIS PartMeshDual
Compute a fill-reducing
ordering of a sparse matrix
ndmetis
METIS NodeND
Convert a mesh into a
graph
m2gmetis
METIS MeshToNodal
METIS MeshToDual
Table 3: An overview of M ETIS’ command-line and library interfaces.
fill-reducing orderings for a wide variety of irregular graphs, unstructured meshes, and sparse matrices.
M ETIS provides a set of stand-alone command-line programs for computing partitionings and fill-reducing orderings as well as an application programming interface (API) that can be used to invoke its various algorithms from
C/C++ or Fortran programs. The list of stand-alone programs and API routines of M ETIS is shown in Table 3. The
API routines allow the user to alter the behaviour of the various algorithms and provide additional routines that partition graphs into unequal-size partitions and compute partitionings that directly minimize the total communication
volume.
3.1
Partitioning a graph
M ETIS can partition an unstructured graph into a user-specified number k of parts using either the multilevel recursive
bisection [4] or the multilevel k-way partitioning [3] paradigms. Both of these methods are able to produce high
quality partitions. However, M ETIS’s multilevel k-way partitioning algorithm provides additional capabilities (e.g.,
Multilevel Partitioning
GO
G1
G1
G2
G2
G3
Uncoarsening Phase
Coarsening Phase
GO
G3
G4
Initial Partitioning Phase
Figure 1: The three phases of multilevel k-way graph partitioning. During the coarsening phase, the size of the graph is successively decreased. During the initial partitioning phase, a k-way partitioning is computed, During the multilevel refinement (or
uncoarsening) phase, the partitioning is successively refined as it is projected to the larger graphs. G0 is the input graph, which is
the finest graph. Gi+1 is the next level coarser graph of Gi . G4 is the coarsest graph.
7
minimize the resulting subdomain connectivity graph, enforce contiguous partitions, minimize alternative objectives,
etc.) and as such it may be preferable than multilevel recursive bisection. M ETIS’ stand-alone program for partitioning
a graph is gpmetis and the functionality that it provides is achieved by the METIS PartGraphRecursive and
METIS PartGraphKway API routines.
3.2
Alternate partitioning objectives
The objective of the traditional graph partitioning problem is to compute a k-way partitioning such that the number of
edges (or in the case of weighted graphs the sum of their weights) that straddle different partitions is minimized. This
objective is commonly referred to as the edge-cut. When partitioning is used to distribute a graph or a mesh among
the processors of a parallel computer, the objective of minimizing the edge-cut is only an approximation of the true
communication cost resulting from the partitioning [1].
The communication cost resulting from a k-way partitioning generally depends on the following factors: (i) the
total communication volume, (ii) the maximum amount of data that any particular processor needs to send and receive; and (iii) the number of messages a processor needs to send and receive. M ETIS’ multilevel k-way partitioning
approaches can be used to directly minimize the total communication volume resulting from the partitioning (first factor). In addition, M ETIS also provides support for minimizing the third factor (which essentially reduces the number
of startups) and indirectly (up to a point) reduces the second factor.
3.3
Support for multi-phase and multi-physics computations
The traditional graph partitioning problem formulation is limited in the types of applications that it can effectively
model because it specifies that only a single quantity be load balanced. Many important types of multi-phase and multiphysics computations require that multiple quantities be load balanced simultaneously. This is because synchronization
steps exist between the different phases of the computations, and so, each phase must be individually load balanced.
That is, it is not sufficient to simply sum up the relative times required for each phase and to compute a partitioning
based on this sum. Doing so may lead to some processors having too much work during one phase of the computation
(and so, these may still be working after other processors are idle), and not enough work during another. Instead, it is
critical that every processor have an equal amount of work from each phase of the computation.
M ETIS includes partitioning routines that can be used to partition a graph in the presence of such multiple balancing
constraints. Each vertex is now assigned a vector of m weights and the objective of the partitioning routines is
to minimize the edge-cut subject to the constraints that each one of the m weights is equally distributed among the
domains. For example, if the first weight corresponds to the amount of computation and the second weight corresponds
to the amount of storage required for each element, then the partitioning computed by the new algorithms will balance
both the computation performed in each domain as well as the amount of memory that it requires. The multi-constraint
partitioning algorithms and their applications are further described in [2]. The gpmetis program invokes the multiconstraint partitioning routines whenever the input graph specifies more that one set of vertex weights, and all of its
graph partitioning API routines allow for the specification of multiple balancing constraints.
3.4
Partitioning a mesh
M ETIS provides the mpmetis program for partitioning meshes arising in finite element or finite volume methods.
This program take as input the element-node array of the mesh and compute a k-way partitioning for both its elements
and its nodes. This program first converts the mesh into either a dual graph (i.e., each element becomes a graph vertex)
or a nodal graph (i.e., each node becomes a graph vertex), and then uses the graph partitioning API routines to partition
this graph. M ETIS utilizes a flexible approach for creating a graph for a finite element mesh, which allows it to handle
meshes with different and possibly mixed element types (e.g., triangles, tetrahedra, hexahedra, etc.). The functionality
provided by mpmetis is achieved by the METIS PartMeshNodal and METIS PartMeshDual API routines.
8
3.5
Partitioning for heterogeneous parallel computing architectures
Heterogeneous computing platforms containing processing nodes with different computational and memory capabilities are becoming increasingly more common. M ETIS’ graph and mesh partitioning programs and API routines are
designed to partition a graph into k parts such that each part contains a pre-specified fraction of the total number
of vertices/elements/nodes. In addition, in the case of multi-constraint partitioning, these pre-specified fractions are
provided for each one of the vertex weights. By matching the weights specified for each partition to the relative computational and memory capabilities of the various processors, these routines can be used to compute partitionings that
balance the computations on heterogeneous architectures.
3.6
Computing a fill-reducing ordering of a sparse matrix
M ETIS provides the ndmetis program and its associated METIS NodeND API routine for computing fill-reducing
orderings of sparse matrices based on the multilevel nested dissection paradigm [4]. The nested dissection paradigm
is based on computing a vertex-separator for the the graph corresponding to the matrix. The nodes in the separator are
moved to the end of the matrix, and a similar process is applied recursively for each one of the other two parts. The
multilevel nested dissection paradigm is quite effective in producing re-orderings that incur low fill-in.
3.7
Converting a mesh into a graph
M ETIS provides the m2gmetis program for converting a mesh into the graph format used by M ETIS. This program
can generate either the nodal or dual graph of the mesh. The corresponding API routines are METIS MeshToNodal
and METIS MeshToDual. Since M ETIS does not provide API routines that can directly compute a multi-constraint
partitioning of a mesh, these routines can be used to first convert the mesh into a graph, which can then be used as
input to M ETIS’ graph partitioning routines to obtain such partitionings.
4
M ETIS’ stand-alone programs
M ETIS provides a variety of programs that can be used to partition graphs, partition meshes, compute fill-reducing
orderings of sparse matrices, as well as programs to convert meshes into graphs appropriate for M ETIS’s graph partitioning programs.
4.1
Input file formats
The various programs in M ETIS require as input either a file storing a graph or a file storing a mesh. The format of
these files are described in the following sections.
4.1.1
Graph file
The primary input of the partitioning and fill-reducing ordering programs in M ETIS is the undirected graph to be
partitioned or ordered. This graph is stored in a file and is supplied to the various programs as one of the command
line parameters.
A graph G = (V, E) with n vertices and m edges is stored in a plain text file that contains n + 1 lines (excluding
comment lines). The first line, referred to as the header line contains information about the size and the type of the
graph, while the remaining n lines contain information for each vertex of G. Any line that starts with ‘%’ is a comment
line and is skipped.
The header line contains either two (n, m), three (n, m, fmt), or four (n, m, fmt, ncon) parameters. The first two
parameters (n, m) are the number of vertices and the number of edges, respectively. Note that in determining the
number of edges m, an edge between any pair of vertices v and u is counted only once and not twice (i.e., we do not
count the edge (v, u) separately from (u, v)). For example, the graph in Figure 2 contains 11 vertices.
The fmt parameter is used to specify if the graph file contains information about vertex sizes, vertex weights, and
edge weights. The fmt parameter is a three-digit binary number. If the least significant bit is set to 1 (i.e., the 1st
bit from right to left), then the graph file provides information about the weights of the edges. If the second least
significant bit is set to 1 (i.e., the 2nd bit from right to left), then the graph file provides information about the weights
9
of the vertices. Finally, if the third lest significant bit is set to 1 (i.e., the 3rd bit from right to left), then the graph file
provides information of the sizes of the vertices. For example, if fmt is 011, then the graph file provides information
about both vertex weights and edge weights. Note that when the fmt parameter is not provided, it is assumed that the
vertex sizes, vertex weights, and edge weights are all equal to 1.
The ncon parameter specifies the number of vertex weights associated with each vertex of the graph. The value of
this parameter determines whether or not M ETIS will use the multi-constraint partitioning algorithms (Section 3.3). If
this parameter is omitted, then the vertices of the graph are assumed to have a single weight. Note that if ncon is greater
than 0, then the file should contain the required vertex weights and the fmt parameter should be set appropriately (i.e.,
the 2nd bit from right to left should be set to 1).
The remaining n lines store information about the actual structure of the graph. In particular, the ith line (excluding
comment lines) contains information that is relevant to the ith vertex. Depending on the value of the fmt and ncon
parameters, the information stored at each line is somewhat different. In the most general form (when fmt = 11 and
ncon > 1) each line will have the following structure (all elements are integer):
s w1 w2 . . . wncon v1 e1 v2 e2 . . . vk ek
where s is the size of the vertex, w1 , w2 , . . . , wncon are the ncon vertex weights associated with this vertex, v1 , . . . , vk
are the vertices adjacent to this vertex, and e1 , . . . , ek are the weights of these edges. The vertices (i.e., the various
vi entries) are numbered starting from 1 (not from 0 as is often done in C). Furthermore, the vertex-sizes and vertexweights must be integers greater or equal to 0, whereas the edge-weights must be strictly greater than 0.
Vertex size versus vertex weights The graph format allows for the specification of both vertex sizes and vertex
weights. These two quantities are used by M ETIS for two entirely different purposes. The vertex weights are used for
ensuring that the computed partitionings satisfy the specified balancing constraints (e.g., the sum of the weights of the
vertices assigned to each partition is the same across the partitions). On the other hand, the vertex sizes are used for
determining the total communication volume, when gpmetis and mpmetis are invoked with the -objtype=vol
option. Additional details on how the vertex size is used to determine the communication volume are provided in
Section 5.7, which provides the precise formula for computing the total communication volume.
Examples Figure 2 illustrates the format by providing some examples. The simplest format for a graph G is when
the size and weight of all vertices and the weight of all the edges is the same. This format is illustrated in Figure 2(a).
Note, the optional fmt parameter is skipped in this case. However, there are cases in which the edges in G have
different weights. This is accommodated as shown in Figure 2(b). Now, the adjacency list of each vertex contains the
weight of the edges in addition to the vertices that is connected with. If v has k vertices adjacent to it, then the line
for v in the graph file contains 2 ∗ k numbers, each pair of numbers stores the vertex that v is connected to, and the
weight of the edge. Note that the fmt parameter is equal to 001, indicating the fact that G has weights on the edges.
In addition to having weights on the edges, weights on the vertices are also allowed, as illustrated in Figure 2(c). In
this case, the value of fmt is equal to 011, and each line of the graph file first stores the weight of the vertex, and then
the weighted adjacency list. Finally, Figure 2(d) illustrates the format of the input file when the vertices of the graph
contain multiple weights (3 in this example). In this case, the value of fmt is equal to 010, and the value of ncon
is equal to 3 (since we have three sets of vertex-weights). Each line of the graph file stores the three weights of the
vertices followed by the adjacency list.
4.1.2
Mesh file
The primary input of the mesh partitioning programs in M ETIS is the mesh to be partitioned. This mesh is stored in a
file in the form of the element node array. A mesh with n elements is stored in a plain text file that contains n + 1 lines.
The first line (i.e., the header line) contains information about the size and the type of the mesh, while the remaining
n lines contain the nodes that make up each element.
The first line contains two integer parameters. The first parameter is the number of elements n in the mesh. The
second parameter, which is optional, is the number of weights associated with each element. This is equivalent to the
10
5
5
2
1
6
1
1
6
3
2
3
6
2
3
1
7
2
2
5
4
2
1
4
11 001
1 3 2 2
1 3 2 4
3 4 2 2
1 3 2 6
1 3 3 6
2 4 2 7
6 4 5
1
1
2 1 2
2 7 5
2
6
7
2
Graph File:
7
5
1
5
2
1
5
6
Graph File:
11
3 2
3 4
4 2 1
3 6 7
3 6
4 7
4
(a)
7
5
1
5
2
1
5
6
(b)
Unweighted Graph
[1]
[1, 1, 1]
5
2
1
[4]
1
6
[5]
[4, 1, 1]
2
2
2
5
3
7 [2]
2
[1, 2, 1]
7
4
[3]
1
[0, 2, 2]
4
[2, 2, 3]
2
Graph File:
(c)
6
1
3
7
4
2
5
3
1
6
2
[2, 2, 1]
[1, 2, 0]
6
2
[2]
5
[6]
3
1
Weighted Graph
Weights on edges
Graph File:
11 011
5 1 3 2
1 1 3 2
5 3 4 2
2 1 3 2
1 1 3 3
5 2 4 2
6 6 4 5
2
4
2
6
6
7
7
1
0
4
2
1
2
1
1
1
2 1 2
2 7 5
2
6
Weighted Graph
Weights both on vertices and edges
11 010 3
2 0 5 3 2
2 2 1 3 4
1 1 5 4 2 1
2 3 2 3 6 7
1 1 1 3 6
2 1 5 4 7
2 1 6 4
(d)
Multi−Constraint Graph
Figure 2: Storage format for various type of graphs.
ncon parameter of the graph file and is used to specify the weights of the vertices in the dual graph. If this parameter
is omitted, the weights of the vertices in the dual graph are assumed to be 11 .
After the first line, the remaining n lines store the element node array. In particular for element i, line i + 1 stores
the ncon integer weights associated with the ith element (if the optional ncon parameter has been specified) followed
by the nodes that this element is made off. The weights and nodes are separated by spaces. The numbering of the
nodes starts from 1. The nodes of each element can be stored in any order. The mesh file allows for mixed elements
meshes, and as such the number of nodes that are supplied in each line can vary.
1 The current mesh format does not allow for the specification of weights with the vertices in the nodal graph. If you need to partition the nodal
graph of a mesh whose vertices can have different weights, then the m2gmetis routine should be used to first create the nodal graph, which should
be subsequently edited to include the required weights, prior to using gpmetis to partition it.
11
4.1.3
Target partition weights file
The graph and mesh-partitioning routines take as input an optional file that specifies the target weights of the various
partitions (using the -tpwgts option. This file contains a sequence of lines, whose format are of the form:
frompid [- topid] [: fromcnum [- tocnum]] = twgt
where frompid and topid specify partition numbers (numbering starting from 0), fromcnum and tocnum specify
constraint numbers (numbering starting from 0), and twgt is a floating point number specifying a fraction of the
total weight (i.e., better be ≤ 1.0). The parts in square-brackets indicate optional parts. The meaning of the above
specification is as follows: For each of the constraints from fromcnum up to and including tocnum for the partitions
starting from frompid up to and including topid will be assigned a target weight of twgt. If [- topid]
is not supplied, then topid = frompid. If [- tocnum] is not supplied, then tocnum = fromcnum. If [:
fromcnum [- tocnum]], then fromcnum=0 and tocnum=ncon-1.
If the file does not contain a twgt specification for all the target partition weight/constraint combinations, then the
twgt for the unspecified ones is determined by equally distributing the left-over portion of the total weight for each
constraint. It is important that for each constraint, the sum of the specified twgts values is less than or equal to 1.0.
For example, assuming that ncon=1, nparts=5, then the following tpwgts file
0-1 = .2
4
= .3
specifies the following target partition weights:
part[0]=.2; part[1]=.2; part[2]=.15; part[3]=.15; and part[4]=0.3
Note that the .15 fractions for part[2] and part[3] are due to the equal distribution of the left-over weight (1.0
- .7).
4.2
Output file formats
The output of M ETIS is either a partition or an ordering file, depending on whether M ETIS is used for graph/mesh
partitioning or for sparse matrix ordering. The format of these files are described in the following sections.
4.2.1
Partition file
The partition file of a graph with n vertices consists of n lines with a single number per line. The ith line of the
file contains the partition number that the ith vertex belongs to. Partition numbers start from 0 up to the number of
partitions minus one.
4.2.2
Ordering file
The ordering file of a graph with n vertices consists of n lines with a single number per line. The ith line of the
ordering file contains the new order of the ith vertex of the graph. The numbering in the ordering file starts from 0.
Note that the ordering file stores what is referred to as the the inverse permutation vector iperm of the ordering. Let
A be a matrix and let A0 be the reordered matrix. The inverse permutation vector maps the ith row (column) of A into
the iperm[i] row (column) of A0 .
4.3
Programs
12
gpmetis [options] graphfile nparts
Description
Partitions a graph into a specified number of parts. The computed partitioning is stored in a file named
graphfile.part.nparts, where graphfile and nparts correspond to the specified parameters.
Parameters
graphfile The name of the file that stores the graph to be partitioned (Section 4.1.1).
nparts
The number of parts that the graph will be partitioned into. It should be greater than 1.
Options
-ptype=string
Specifies the scheme used for computing the partitioning. Possible values:
rb
Multilevel recursive bisectioning.
kway Multilevel k-way partitioning [default].
-ctype=string
Specifies the scheme used to match the vertices of the graph during the coarsening. Possible values:
rm
Random matching.
shem Sorted heavy-edge matching [default].
-iptype=string (applies only when -ptype=rb)
Specifies the scheme used to compute the initial partitioning of the graph. Possible values:
grow
Grows a bisection using a greedy strategy [default].
random Computes a bisection at random followed by a refinement.
-objtype=string (applies only when -ptype=kway)
Specifies the partitioning’s objective function. Possible values:
cut Edgecut minimization [default].
vol Total communication volume minimization.
-contig (applies only when -ptype=kway)
Specifies that the partitioning routines should try to produce partitions that are contiguous. Note that
if the input graph is not connected this option is ignored.
-minconn (applies only when -ptype=kway)
Specifies that the partitioning routines should try to minimize the maximum degree of the subdomain
graph, i.e., the graph in which each partition is a node, and edges connect subdomains with a shared
interface.
-tpwgts=filename
Specifies the name of the file that stores for each constraint the target weights for each partition
(Section 4.1.3). By default, for each constraint, all partitions are assumed to be of the same size.
-ufactor=int
Specifies the maximum allowed load imbalance among the partitions. It is this described under
METIS OPTION UFACTOR in Section 5.4.
Note that in the case of multiple constraints, the same load imbalance tolerance is applied to all the
constraints. Use -ubvec to provide per-constraint load imbalance tolerances.
-ubvec=string
Specifies the per-constraint allowed load imbalance among partitions. The string must contain a
space separated list of floating point numbers, one for each of the constraints. For example, for three
constraints, the string can be "1.02 1.2 1.35" indicating a desired maximum load imbalance
of 2%, 20%, and 35%, respectively. The load imbalance is defined in a way similar to -ufactor. If
supplied, this parameter takes priority over ufactor.
13
-niter=int
Specifies the number of iterations for the refinement algorithms at each stage of the uncoarsening
process. Default is 10.
-ncuts=int
Specifies the number of different partitionings that it will compute. The final partitioning is the one
that achieves the best edgecut or communication volume. Default is 1.
-nooutput
Specifies that no partitioning file should be generated.
-seed=int Selects the seed of the random number generator.
-dbglvl=int
Specifies the type of progress/debugging information that will be printed to stdout. The supplied
value corresponds to the addition (bitwise OR) of the various values described in Section 5.4 for
METIS OPTION DBGLVL. The default value is 0 (no progress/debugging information).
-help
Displays the command-line options along with a description.
14
mpmetis [options] meshfile nparts
Description
Partitions a mesh into a specified number of parts. The computed partitioning is stored in two files named:
meshfile.npart.nparts that stores the partitioning of the nodes, and meshfile.epart.nparts
that stores the partitioning of the elements. The meshfile and nparts components of those files correspond
to the specified parameters. The format of the partitioning files is described in Section 4.2.1.
Parameters
meshfile
nparts
The name of the file that stores the mesh to be partitioned (Section 4.1.2).
The number of parts that the mesh will be partitioned into. It should be greater than 1.
Options
-gtype=string
Specifies the graph to be used for computing the partitioning The possible values are:
dual
Partition the dual graph of the mesh [default].
nodal Partition the nodal graph of the mesh.
-ncommon=int (applies only when -gtype=dual)
Specifies the number of common nodes that two elements must have in order to put an edge between
them in the dual graph. Given two elements e1 and e2 , containing n1 and n2 nodes, respectively,
then an edge will connect the vertices in the dual graph corresponding to e1 and e2 if the number of
common nodes between them is greater than or equal to min(ncommon, n1 − 1, n2 − 1).
The default value is 1, indicating that two elements will be connected via an edge as long as they
share one node. However, this will tend to create too many edges (increasing the memory and time
requirements of the partitioning). The user should select higher values that are better suited for the
element types of the mesh that wants to partition. For example, for tetrahedron meshes, ncommon
should be 3, which creates an edge between two tets when they share a triangular face (i.e., 3 nodes).
-ptype, -ctype, -iptype, -objtype, -contig, -minconn, -tpwgts, -ufactor, -niter, -ncuts, -nooutput, -seed,
-dbglvl, -help
Similar to the corresponding options of gpmetis.
Notes
The current version of mpmetis supports only single constraint partitioning.
15
ndmetis [options] graphfile
Description
Computes a fill-reducing ordering of the vertices of the graph using multilevel nested dissection. The computed
ordering is stored in a file named graphfile.iperm, whose format is described in Section 4.2.2.
Parameters
graphfile The name of the file that stores the graph to be re-ordered (Section 4.1.1).
Options
-ctype=string
Specifies the scheme to be used to match the vertices of the graph during the coarsening. The possible
values are:
rm
Random matching [default].
shem Sorted heavy-edge matching.
-iptype=string (applies only when -ptype=rb)
Specifies the scheme to be used to compute the initial vertex separator of the graph. The possible
values are:
edge Derive the separator from an edge cut [default].
node Grow a bisection using a greedy node-based strategy.
-rtype=string
Specifies the scheme to be used for refinement. The possible values are:
1sided One-sided node-based refinement [default].
2sided Two-sided node-based refinement.
-ufactor=int
Specifies the maximum allowed load imbalance between the left and right partitions during each
bisection. It is this described under METIS OPTION UFACTOR in Section 5.4 when the number of
partitions . Default is 30, indicating a load imbalance of 1.03.
-pfactor=int
Specifies the minimum degree of the vertices that will be ordered last. It is this described under
METIS OPTION PFACTOR in Section 5.4. Default value is 0, indicating that no vertices are removed.
-nocompress
Specifies that the graph should not be compressed by combining together vertices that have identical
adjacency lists.
-ccorder Specifies if the connected components of the graph should first be identified and ordered separately.
-niter=int
Specifies the number of iterations for the refinement algorithms at each stage of the uncoarsening
process. Default is 10.
-nseps=int
Specifies the number of different separators that it will compute at each level of nested dissection.
The final separator that is used is the smallest one. Default is 1.
-nooutput
Specifies that no ordering file should be generated.
-seed=int Selects the seed of the random number generator.
-dbglvl=int
Similar to the corresponding option of gpmetis.
-help
Displays the command-line options along with a description.
16
m2gmetis [options] meshfile graphfile
Description
Converts a mesh into a graph that is compatible with M ETIS.
Parameters
meshfile
The name of the file that stores the mesh to be converted (Section 4.1.2).
graphfile The name of the file that will store the generated graph.
Options
-gtype=string
Specifies the type of the graph to be generated. The possible values are:
dual
Generates the dual graph of the mesh [default].
nodal the nodal graph of the mesh.
-ncommon=int (applies only when -gtype=dual)
Similar to the corresponding option of mpmetis.
-dbglvl=int
Similar to the corresponding option of gpmetis.
-help
Displays the command-line options along with a description.
17
graphchk graphfile [fixedfile]
Description
Checks the graph for format correctness and consistency.
Parameters
graphfile The name of the file that stores the graph to be checked (Section 4.1.2).
fixedfile
This is an optional parameter that specifies the name of the file to store the fixed input graph. This
file will only be generated if there were errors in the input file.
18
5
M ETIS’ API
The various routines implemented in M ETIS’ stand-alone programs can be directly accessed from a C, C++, or Fortran
program by using the supplied library. In the rest of this section we describe M ETIS’ API by first describing various
calling and usage conventions, the various data structures used to pass information into and get information out of the
routines, followed by a detailed description of the calling sequence of the various routines.
5.1
Header files
Any program using M ETIS’ API needs to include the metis.h header file. This file provides function prototypes for
the various API routines and defines the various data types and constants used by these routines.
During M ETIS’ installation time, the metis.h defines two important data types and their widths. These are the
idx t data type for storing integer quantities and the real t data type for storing floating point quantities. The
idx t data type can be defined to be either a 32 or 64 bit signed integer, whereas the real t data type can be defined
to be either a single or double precision float point number. All of M ETIS’ API routines take as input arrays and/or
scalars that are of these two data types. In addition, metis.h defines various enum data types for specifying various
options and for returning status codes.
5.2
Use of NULL parameters
M ETIS’ API routines take a large number of parameters, allowing the user to model complex graphs and specify
complex partitioning/ordering requirements. However, for most uses of M ETIS, this level of complexity may not
be required. For that purpose and to also simplify the complexity associated with using its API, M ETIS allows the
application to specify a NULL value to many of these optional/advanced parameters. The API routines that will be
described in subsequent sections will mark these parameters by following them with a (NULL).
5.3
C/C++ and Fortran Support
The various routines in M ETIS’ API can be called from either C/C++ or Fortran programs. Using C/C++ with M ETIS’
API is quite straightforward (as M ETIS is written entirely in C). However, M ETIS’ API fully supports Fortran as well.
This support comes in three forms.
1. All the scalar arguments in the routines are passed by reference to facilitate Fortran programs.
2. All the routines take a parameter called numflag or an options parameter called METIS OPTION NUMBERING
indicating whether or not the numbering of the graph or mesh starts from 0 or 1. In C programs numbering usually starts from 0, whereas in Fortran programs numbering starts from 1.
3. M ETIS’ API incorporates alternative names for each of the routines to facilitate linking the library with Fortran
programs. In particular, for every function, M ETIS’ API provides four additional names, one all capital, one all
lower case, one all lower case with ‘ ’ appended to it, and one with ‘ ’ appended to it. For example, for
METIS PartGraphKway, M ETIS’ API provides METIS PARTGRAPHKWAY, metis partgraphkway,
metis partgraphkway , and metis partgraphkway . These extra names allow the library to be directly linked into Fortran programs on a wide range of architectures including Cray, SGI, and HP. If you still
encounter problems linking with the library let us know so we can include appropriate support.
5.4
Options array
Most of the API routines take as a parameter an array called options, which allow the application to fine-tune and
modify various aspects of the internal algorithms used by M ETIS. The application must define this array as
idx_t options[METIS_NOPTIONS];
and the meaning of its various entries are as follows:
19
options[METIS OPTION PTYPE]
Specifies the partitioning method. Possible values are:
METIS PTYPE RB
Multilevel recursive bisectioning.
METIS PTYPE KWAY
Multilevel k-way partitioning.
options[METIS OPTION OBJTYPE]
Specifies the type of objective. Possible values are:
METIS OBJTYPE CUT Edge-cut minimization.
METIS OBJTYPE VOL Total communication volume minimization.
options[METIS OPTION CTYPE]
Specifies the matching scheme to be used during coarsening. Possible values are:
METIS CTYPE RM
Random matching.
METIS CTYPE SHEM
Sorted heavy-edge matching.
options[METIS OPTION IPTYPE]
Determines the algorithm used during initial partitioning. Possible values are:
METIS IPTYPE GROW
Grows a bisection using a greedy strategy.
METIS IPTYPE RANDOM
Computes a bisection at random followed by a refinement.
METIS IPTYPE EDGE
Derives a separator from an edge cut.
METIS IPTYPE NODE
Grow a bisection using a greedy node-based strategy.
options[METIS OPTION RTYPE]
Determines the algorithm used for refinement. Possible values are:
METIS RTYPE FM
FM-based cut refinement.
METIS RTYPE GREEDY
Greedy-based cut and volume refinement.
METIS RTYPE SEP2SIDED
Two-sided node FM refinement.
METIS RTYPE SEP1SIDED
One-sided node FM refinement.
options[METIS OPTION NCUTS]
Specifies the number of different partitionings that it will compute. The final partitioning is the one that
achieves the best edgecut or communication volume. Default is 1.
options[METIS OPTION NSEPS]
Specifies the number of different separators that it will compute at each level of nested dissection. The final
separator that is used is the smallest one. Default is 1.
options[METIS OPTION NUMBERING]
Used to indicate which numbering scheme is used for the adjacency structure of a graph or the elementnode structure of a mesh. The possible values are:
0
C-style numbering is assumed that starts from 0.
1
Fortran-style numbering is assumed that starts from 1.
options[METIS OPTION NITER]
Specifies the number of iterations for the refinement algorithms at each stage of the uncoarsening process.
Default is 10.
options[METIS OPTION SEED]
Specifies the seed for the random number generator.
20
options[METIS OPTION MINCONN]
Specifies that the partitioning routines should try to minimize the maximum degree of the subdomain graph,
i.e., the graph in which each partition is a node, and edges connect subdomains with a shared interface.
options[METIS OPTION CONTIG]
Specifies that the partitioning routines should try to produce partitions that are contiguous. Note that if the
input graph is not connected this option is ignored.
options[METIS OPTION COMPRESS]
Specifies that the graph should be compressed by combining together vertices that have identical adjacency
lists.
options[METIS OPTION CCORDER]
Specifies if the connected components of the graph should first be identified and ordered separately.
options[METIS OPTION PFACTOR]
Specifies the minimum degree of the vertices that will be ordered last. If the specified value is x > 0, then
any vertices with a degree greater than 0.1*x*(average degree) are removed from the graph, an ordering of
the rest of the vertices is computed, and an overall ordering is computed by ordering the removed vertices
at the end of the overall ordering. For example if x = 40, and the average degree is 5, then the algorithm
will remove all vertices with degree greater than 20. The vertices that are removed are ordered last (i.e.,
they are automatically placed in the top-level separator). Good values are often in the range of 60 to 200
(i.e., 6 to 20 times more than the average). Default value is 0, indicating that no vertices are removed.
Used to control whether or not the ordering algorithm should remove any vertices with high degree (i.e.,
dense columns). This is particularly helpful for certain classes of LP matrices, in which there a few vertices
that are connected to many other vertices. By removing these vertices prior to ordering, the quality and the
amount of time required to do the ordering improves.
options[METIS OPTION UFACTOR]
Specifies the maximum allowed load imbalance among the partitions. A value of x indicates that the
allowed load imbalance is (1 + x)/1000. The load imbalance for the jth constraint is defined to be
maxi (w[j, i])/t[j, i]), where w[j, i] is the fraction of the overall weight of the jth constraint that is assigned to the ith partition and t[j, i] is the desired target weight of the jth constraint for the ith partition
(i.e., that specified via -tpwgts). For -ptype=rb, the default value is 1 (i.e., load imbalance of 1.001) and for
-ptype=kway, the default value is 30 (i.e., load imbalance of 1.03).
options[METIS OPTION DBGLVL]
Specifies the amount of progress/debugging information will be printed during the execution of the algorithms. The default value is 0 (no debugging/progress information). A non-zero value can be supplied that
is obtained by a bit-wise OR of the following values.
METIS DBG INFO (1)
Prints various diagnostic messages.
METIS DBG TIME (2)
Performs timing analysis.
METIS DBG COARSEN (4)
Displays various statistics during coarsening.
METIS DBG REFINE (8)
Displays various statistics during refinement.
METIS DBG IPART (16)
Displays various statistics during initial partitioning.
METIS DBG MOVEINFO (32)
Displays detailed information about vertex moves during refinement.
METIS DBG SEPINFO (64)
Displays information about vertex separators.
METIS DBG CONNINFO (128)
Displays information related to the minimization of subdomain
connectivity.
21
Displays information related to the elimination of connected components.
METIS DBG CONTIGINFO (256)
Note that the numeric values are provided for use with the -dbglvl option of M ETIS’ stand-alone programs. For the API routines it is sufficient to OR the above constants.
If an application does not want to take advantage of this capability, then it can just supply a NULL as the value for
that parameter. For those applications that will like to modify certain elements of the algorithms, M ETIS provide the
METIS SetDefaultOptions routine to set the options to their default values. After that, the application can just
modify the the options that is interested in modifying. This is illustrated as follows:
idx_t options[METIS_NOPTIONS];
METIS_SetDefaultOptions(options);
options[METIS_OPTION_NSEPS] = 10;
options[METIS_OPTION_UFACTOR] = 100;
METIS_NodeND(..., options, ...)
...
5.5
Graph data structure
All of the graph partitioning and sparse matrix ordering routines in M ETIS take as input the adjacency structure of the
graph and the weights of the vertices and edges (if any).
The adjacency structure of the graph is stored using the compressed storage format (CSR). The CSR format is a
widely used scheme for storing sparse graphs. In this format the adjacency structure of a graph with n vertices and
m edges is represented using two arrays xadj and adjncy. The xadj array is of size n + 1 whereas the adjncy
array is of size 2m (this is because for each edge between vertices v and u we actually store both (v, u) and (u, v)).
The adjacency structure of the graph is stored as follows. Assuming that vertex numbering starts from 0 (C style),
then the adjacency list of vertex i is stored in array adjncy starting at index xadj[i] and ending at (but not
including) index xadj[i + 1] (i.e., adjncy[xadj[i]] through and including adjncy[xadj[i + 1]-1]). That
is, for each vertex i, its adjacency list is stored in consecutive locations in the array adjncy, and the array xadj
is used to point to where it begins and where it ends. Figure 3(b) illustrates the CSR format for the 15-vertex graph
shown in Figure 3(a).
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14
(a) A sample graph
xadj
adjncy
0 2 5 8 11 13 16 20 24 28 31 33 36 39 42 44
1 5 0 2 6 1 3 7 2 4 8 3 9 0 6 10 1 5 7 11 2 6 8 12 3 7 9 13 4 8 14 5 11 6 10 12 7 11 13 8 12 14 9 13
(b CSR format
Figure 3: An example of the CSR format for storing sparse graphs.
The weights of the vertices (if any) are stored in an additional array called vwgt. If ncon is the number of weights
associated with each vertex, the array vwgt contains n ∗ ncon elements (recall that n is the number of vertices). The
weights of the ith vertex are stored in ncon consecutive entries starting at location vwgt[i ∗ ncon]. Note that if
each vertex has only a single weight, then vwgt will contain n elements, and vwgt[i] will store the weight of the
22
ith vertex. The vertex-weights must be integers greater or equal to zero. If all the vertices of the graph have the same
weight (i.e., the graph is unweighted), then the vwgt can be set to NULL.
The weights of the edges (if any) are stored in an additional array called adjwgt. This array contains 2m elements,
and the weight of edge adjncy[j] is stored at location adjwgt[j]. The edge-weights must be integers greater
than zero. If all the edges of the graph have the same weight (i.e., the graph is unweighted), then the adjwgt can be
set to NULL.
5.6
Mesh data structure
All of the mesh partitioning and mesh conversion routines in M ETIS take as input the element node array of a mesh.
This element node array is stored using a pair of arrays called eptr and eind, which are similar to the xadj and
adjncy arrays used for storing the adjacency structure of a graph. The size of the eptr array is n + 1, where n is the
number of elements in the mesh. The size of the eind array is of size equal to the sum of the number of nodes in all
the elements of the mesh. The list of nodes belonging to the ith element of the mesh are stored in consecutive locations
of eind starting at position eptr[i] up to (but not including) position eptr[i+1]. This format makes it easy
to specify meshes of any type of elements, including meshes with mixed element types that have different number of
nodes per element. As it was the case with the format of the mesh file described in Section 4.1.2, the ordering of the
nodes in each element is not important.
5.7
Partitioning objectives
The partitioning algorithms in M ETIS can be used to compute a balanced k-way partitioning that minimizes either
the number of edges that straddle partitions (edgecut) or the total communication volume (totalv). In the rest of this
section we briefly describe these two objectives and provide some suggestions on when they should be used.
Minimizing the edgecut Consider a graph G = (V, E), and let P be a vector of size |V | such that P [i] stores the
number of the partition that vertex i belongs to. The edgecut of this partitioning is defined as the number of edges that
straddle partitions. That is, the number of edges (v, u) for which P [v] 6= P [u]. If the graph has weights associated
with the edges, then the edgecut is defined as the sum of the weight of these straddling edges.
Minimizing the total communication volume Consider a graph G = (V, E), and let P be a vector of size |V |
such that P [i] stores the number of the partition that vertex i belongs to. Let Vb ⊂ V be the subset of interface (or
boarder) vertices. That is, each vertex v ∈ Vb is connected to at least one vertex that belongs to a different partition.
For each vertex v ∈ Vb let N adj[v] be the number of domains other than P [v] that the vertices adjacent to v belong to.
The totalv of this partitioning is defined as:
X
totalv =
N adj[v].
(1)
v∈Vb
Equation 1 corresponds to the total communication volume incurred by the partitioning because each interface vertex
v needs to be sent to all of its N adj[v] partitions.
The above model can be extended to instances in which the amount of data that needs to be sent for each node is
different. In particular, if sv is the amount of data that needs to be sent for vertex v, referred to as the vertex size, then
Equation 1 can be re-written as:
X
totalv =
sv N adj[v].
(2)
v∈Vb
M ETIS’ API supports this weighted totalv model by using an array called vsize such that the amount of data that
needs to be sent due to the ith vertex is stored in vsize[i]. Note that the amount of data that needs to be sent is
different from the weight of the vertex. The former corresponds to communication cost whereas the later corresponds
to the computational cost.
Note that for partitioning algorithms to correctly minimize the totalv, the graph should reflect the true information
exchange requirements of the underlying computations. For instance, the dual graph of a finite element mesh does not
23
correctly model the underlying communication, whereas the nodal graph does.
Which one is better? When partitioning is used to distribute a graph or a mesh among the processors of a parallel
computer, the edgecut is only an approximation of the true communication cost resulting from the partitioning. On
the other hand, by minimizing the totalv we can directly minimize the overall communication cost. Despite of that,
for many graphs the solutions obtained by minimizing the edgecut or minimizing the totalv, are comparable. This
is especially true for graphs corresponding to well-shaped finite element meshes. This is because for these graphs,
the degrees of the various vertices are similar and the objectives of minimizing the edgecut or the totalv behave the
same. On the other hand, if the vertex degrees vary significantly (e.g., graphs corresponding to linear programming
matrices), then by minimizing the totalv we can obtain a significant reduction in the total communication volume.
In terms of the amount of time required by these two partitioning objectives, minimizing the edgecut is faster than
minimizing the totalv. For this reason, the totalv objective should be used only for problems in which it actually
reduces the overall communication volume.
24
5.8
Graph partitioning routines
int METIS PartGraphRecursive(idx t *nvtxs, idx t *ncon, idx t *xadj, idx t *adjncy,
idx t *vwgt, idx t *vsize, idx t *adjwgt, idx t *nparts, real t *tpwgts,
real t ubvec, idx t *options, idx t *objval, idx t *part)
int METIS PartGraphKway(idx t *nvtxs, idx t *ncon, idx t *xadj, idx t *adjncy,
idx t *vwgt, idx t *vsize, idx t *adjwgt, idx t *nparts, real t *tpwgts,
real t ubvec, idx t *options, idx t *objval, idx t *part)
Description
Is used to partition a graph into k parts using either multilevel recursive bisection or multilevel k-way partitioning.
Parameters
nvtxs
The number of vertices in the graph.
ncon
The number of balancing constraints. It should be at least 1.
xadj, adjncy
The adjacency structure of the graph as described in Section 5.5.
vwgt (NULL)
The weights of the vertices as described in Section 5.5.
vsize (NULL)
The size of the vertices for computing the total communication volume as described in Section 5.7.
adjwgt (NULL)
The weights of the edges as described in Section 5.5.
nparts
The number of parts to partition the graph.
tpwgts (NULL)
This is an array of size nparts×ncon that specifies the desired weight for each partition and constraint.
The target partition weight for the ith partition and jth constraint is specified at tpwgts[i*ncon+j]
(the numbering for both partitions and constraints starts from 0). For each constraint, the sum of the
P
tpwgts[] entries must be 1.0 (i.e., i tpwgts[i ∗ ncon + j] = 1.0).
A NULL value can be passed to indicate that the graph should be equally divided among the partitions.
ubvec (NULL)
This is an array of size ncon that specifies the allowed load imbalance tolerance for each constraint.
For the ith partition and jth constraint the allowed weight is the ubvec[j]*tpwgts[i*ncon+j] fraction
of the jth’s constraint total weight. The load imbalances must be greater than 1.0.
A NULL value can be passed indicating that the load imbalance tolerance for each constraint should
be 1.001 (for ncon=1) or 1.01 (for ncon¿1).
options (NULL)
This is the array of options as described in Section 5.4.
The following options are valid for METIS PartGraphRecursive:
METIS_OPTION_CTYPE, METIS_OPTION_IPTYPE, METIS_OPTION_RTYPE,
METIS_OPTION_NCUTS, METIS_OPTION_NITER, METIS_OPTION_SEED,
METIS_OPTION_UFACTOR, METIS_OPTION_NUMBERING, METIS_OPTION_DBGLVL
The following options are valid for METIS PartGraphKway:
25
METIS_OPTION_OBJTYPE, METIS_OPTION_CTYPE, METIS_OPTION_IPTYPE,
METIS_OPTION_RTYPE, METIS_OPTION_NCUTS, METIS_OPTION_NITER,
METIS_OPTION_UFACTOR, METIS_OPTION_MINCONN, METIS_OPTION_CONTIG,
METIS_OPTION_SEED, METIS_OPTION_NUMBERING, METIS_OPTION_DBGLVL
objval
Upon successful completion, this variable stores the edge-cut or the total communication volume of
the partitioning solution. The value returned depends on the partitioning’s objective function.
part
This is a vector of size nvtxs that upon successful completion stores the partition vector of the graph.
The numbering of this vector starts from either 0 or 1, depending on the value of
options[METIS OPTION NUMBERING].
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
26
5.9
Mesh partitioning routines
int METIS PartMeshDual(idx t *ne, idx t *nn, idx t *eptr, idx t *eind, idx t *vwgt, idx t *vsize,
idx t *ncommon, idx t *nparts, real t *tpwgts, idx t *options, idx t *objval,
idx t *epart, idx t *npart)
Description
This function is used to partition a mesh into k parts based on a partitioning of the mesh’s dual graph.
Parameters
ne
The number of elements in the mesh.
nn
The number of nodes in the mesh.
eptr, eind
The pair of arrays storing the mesh as described in Section 5.6.
vwgt (NULL)
An array of size ne specifying the weights of the elements. A NULL value can be passed to indicate
that all elements have an equal weight.
vsize (NULL)
An array of size ne specifying the size of the elements that is used for computing the total communication volume as described in Section 5.7. A NULL value can be passed when the objective is cut or
when all elements have an equal size.
ncommon
Specifies the number of common nodes that two elements must have in order to put an edge between
them in the dual graph. Given two elements e1 and e2 , containing n1 and n2 nodes, respectively,
then an edge will connect the vertices in the dual graph corresponding to e1 and e2 if the number of
common nodes between them is greater than or equal to min(ncommon, n1 − 1, n2 − 1).
The default value is 1, indicating that two elements will be connected via an edge as long as they
share one node. However, this will tend to create too many edges (increasing the memory and time
requirements of the partitioning). The user should select higher values that are better suited for the
element types of the mesh that wants to partition. For example, for tetrahedron meshes, ncommon
should be 3, which creates an edge between two tets when they share a triangular face (i.e., 3 nodes).
nparts
The number of parts to partition the mesh.
tpwgts (NULL)
This is an array of size nparts that specifies the desired weight for each partition. The target partition
weight for the ith partition is specified at tpwgts[i] (the numbering for the partitions starts from
0). The sum of the tpwgts[] entries must be 1.0.
A NULL value can be passed to indicate that the graph should be equally divided among the partitions.
options (NULL)
This is the array of options as described in Section 5.4. The following options are valid:
METIS_OPTION_PTYPE, METIS_OPTION_OBJTYPE, METIS_OPTION_CTYPE,
METIS_OPTION_IPTYPE, METIS_OPTION_RTYPE, METIS_OPTION_NCUTS,
METIS_OPTION_NITER, METIS_OPTION_SEED, METIS_OPTION_UFACTOR,
METIS_OPTION_NUMBERING, METIS_OPTION_DBGLVL
objval
Upon successful completion, this variable stores either the edgecut or the total communication volume of the dual graph’s partitioning.
27
epart
This is a vector of size ne that upon successful completion stores the partition vector for the elements
of the mesh. The numbering of this vector starts from either 0 or 1, depending on the value of
options[METIS OPTION NUMBERING].
npart
This is a vector of size nn that upon successful completion stores the partition vector for the nodes
of the mesh. The numbering of this vector starts from either 0 or 1, depending on the value of
options[METIS OPTION NUMBERING].
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
28
int METIS PartMeshNodal( idx t *ne, idx t *nn, idx t *eptr, idx t *eind, idx t *vwgt, idx t *vsize,
idx t *nparts, real t *tpwgts, idx t *options, idx t *objval, idx t *epart, idx t *npart)
Description
This function is used to partition a mesh into k parts based on a partitioning of the mesh’s nodal graph.
Parameters
ne
The number of elements in the mesh.
nn
The number of nodes in the mesh.
eptr, eind
The pair of arrays storing the mesh as described in Section 5.6.
vwgt (NULL)
An array of size nn specifying the weights of the nodes. A NULL value can be passed to indicate that
all nodes have an equal weight.
vsize (NULL)
An array of size nn specifying the size of the nodes that is used for computing the total communication volume as described in Section 5.7. A NULL value can be passed when the objective is cut or
when all nodes have an equal size.
nparts
The number of parts to partition the mesh.
tpwgts (NULL)
This is an array of size nparts that specifies the desired weight for each partition. The target partition
weight for the ith partition is specified at tpwgts[i] (the numbering for the partitions starts from
0). The sum of the tpwgts[] entries must be 1.0.
A NULL value can be passed to indicate that the graph should be equally divided among the partitions.
options (NULL)
This is the array of options as described in Section 5.4. The following options are valid:
METIS_OPTION_PTYPE, METIS_OPTION_OBJTYPE, METIS_OPTION_CTYPE,
METIS_OPTION_IPTYPE, METIS_OPTION_RTYPE, METIS_OPTION_NCUTS,
METIS_OPTION_NITER, METIS_OPTION_SEED, METIS_OPTION_UFACTOR,
METIS_OPTION_NUMBERING, METIS_OPTION_DBGLVL
objval
Upon successful completion, this variable stores either the edgecut or the total communication volume of the nodal graph’s partitioning.
epart
This is a vector of size ne that upon successful completion stores the partition vector for the elements
of the mesh. The numbering of this vector starts from either 0 or 1, depending on the value of
options[METIS OPTION NUMBERING].
npart
This is a vector of size nn that upon successful completion stores the partition vector for the nodes
of the mesh. The numbering of this vector starts from either 0 or 1, depending on the value of
options[METIS OPTION NUMBERING].
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
29
5.10
Sparse Matrix Reordering Routines
int METIS NodeND(idx t *nvtxs, idx t *xadj, idx t *adjncy, idx t *vwgt, idx t *options,
idx t *perm, idx t *iperm)
Description
This function computes fill reducing orderings of sparse matrices using the multilevel nested dissection algorithm.
Parameters
nvtxs
The number of vertices in the graph.
xadj, adjncy
The adjacency structure of the graph as described in Section 5.5.
vwgt (NULL)
An array of size nvtxs specifying the weights of the vertices. If the graph is weighted, the nested
dissection ordering computes vertex separators that minimize the sum of the weights of the vertices
on the separators.
A NULL can be passed to indicate a graph with equal weight vertices (or unweighted).
options (NULL)
This is the array of options as described in Section 5.4. The following options are valid:
METIS_OPTION_CTYPE, METIS_OPTION_RTYPE, METIS_OPTION_NSEPS,
METIS_OPTION_NITER, METIS_OPTION_UFACTOR, METIS_OPTION_COMPRESS,
METIS_OPTION_CCORDER, METIS_OPTION_SEED, METIS_OPTION_PFACTOR,
METIS_OPTION_NUMBERING, METIS_OPTION_DBGLVL
perm, iperm
These are vectors, each of size nvtxs. Upon successful completion, they store the fill-reducing permutation and inverse-permutation. Let A be the original matrix and A0 be the permuted matrix. The
arrays perm and iperm are defined as follows. Row (column) i of A0 is the perm[i] row (column) of
A, and row (column) i of A is the iperm[i] row (column) of A0 . The numbering of this vector starts
from either 0 or 1, depending on the value of options[METIS OPTION NUMBERING].
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
30
5.11
Mesh-to-graph conversion routines
int METIS MeshToDual(idx t *ne, idx t *nn, idx t *eptr, idx t *eind, idx t *ncommon,
idx t *numflag, idx t **xadj, idx t **adjncy)
Description
This function is used to generate the dual graph of a mesh.
Parameters
ne
The number of elements in the mesh.
nn
The number of nodes in the mesh.
eptr, eind
The pair of arrays storing the mesh as described in Section 5.6.
ncommon
Specifies the number of common nodes that two elements must have in order to put an edge between
them in the dual graph. Given two elements e1 and e2 , containing n1 and n2 nodes, respectively,
then an edge will connect the vertices in the dual graph corresponding to e1 and e2 if the number of
common nodes between them is greater than or equal to min(ncommon, n1 − 1, n2 − 1).
The default value is 1, indicating that two elements will be connected via an edge as long as they
share one node. However, this will tend to create too many edges (increasing the memory and time
requirements of the partitioning). The user should select higher values that are better suited for the
element types of the mesh that wants to partition. For example, for tetrahedron meshes, ncommon
should be 3, which creates an edge between two tets when they share a triangular face (i.e., 3 nodes).
numflag Used to indicate which numbering scheme is used for eptr and eind. The possible values are:
0
C-style numbering is assumed that starts from 0
1 Fortran-style numbering is assumed that starts from 1
xadj, adjncy
These arrays store the adjacency structure of the generated dual graph. The format is of the adjacency
structure is described in Section 5.5. Memory for these arrays is allocated by M ETIS’ API using the
standard malloc function. It is the responsibility of the application to free this memory by calling
free. M ETIS provides the METIS Free that is a wrapper to C’s free function.
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
31
int METIS MeshToNodal(idx t *ne, idx t *nn, idx t *eptr, idx t *eind, idx t *numflag,
idx t **xadj, idx t **adjncy)
Description
This function is used to generate the nodal graph of a mesh.
Parameters
ne
The number of elements in the mesh.
nn
The number of nodes in the mesh.
eptr, eind
The pair of arrays storing the mesh as described in Section 5.6.
numflag Used to indicate which numbering scheme is used for eptr and eind. The possible values are:
0
C-style numbering is assumed that starts from 0
1 Fortran-style numbering is assumed that starts from 1
xadj, adjncy
These arrays store the adjacency structure of the generated graph. The format is of the adjacency
structure is described in Section 5.5. Memory for these arrays is allocated by M ETIS’ API using the
standard malloc function. It is the responsibility of the application to free this memory by calling
free. M ETIS provides the METIS Free that is a wrapper to C’s free function.
Returns
METIS OK
Indicates that the function returned normally.
METIS ERROR INPUT
Indicates an input error.
METIS ERROR MEMORY
Indicates that it could not allocate the required memory.
METIS ERROR
Indicates some other type of error.
32
5.12
Utility routines
int METIS SetDefaultOptions(idx t options[METIS NOPTIONS])
Description
Initializes the options array into its default values.
Parameters
options
The array of options that will be initialized. It’s size should be at least METIS NOPTIONS.
Returns
METIS OK
Indicates that the function returned normally.
int METIS Free(idx t *ptr)
Description
Frees the memory that was allocated by either the METIS MeshToDual or the METIS MeshToNodal routines for returning the dual or nodal graph of a mesh.
Parameters
ptr
The pointer to be freed. This pointer should be one of the xadj or adjncy returned by M ETIS’
API routines.
Returns
METIS OK
Indicates that the function returned normally.
33
6
System requirements and contact information
M ETIS is written entirely in ANSI C, and is portable on most Unix systems that have an ANSI C compiler (the GNU C
compiler will do). It has been tested on Linux, SunOS, and OSX. Instructions on how to build and install M ETIS can
be found in the file Install.txt of the distribution.
M ETIS have been extensively tested on a number of different architectures. However, even though M ETIS contains
no known bugs, this does not mean that all of its bugs have been found and fixed. If you have any problems, please
send email to [email protected] with a brief description of the problem. Also, any future updates to M ETIS will be
made available on WWW at http://www.cs.umn.edu/˜metis.
7
Copyright & license notice
M ETIS is copyrighted by the Regents of the University of Minnesota. It can be freely used for educational and research
purposes by non-profit institutions and US government agencies only. Other organizations are allowed to use M ETIS
only for evaluation purposes, and any further uses will require prior approval. The software may not be sold or
redistributed without prior approval. One may make copies of the software for their use provided that the copies, are
not sold or distributed, are used under the same terms and conditions.
As unestablished research software, this code is provided on an “as is” basis without warranty of any kind, either
expressed or implied. The downloading, or executing any part of this software constitutes an implicit agreement to
these terms. These terms and conditions are subject to change at any time without prior notice.
References
[1] Bruce Hendrickson. Graph partitioning and parallel solvers: Has the emperor no clothes? In Proc. of Irregular 1998, 1998.
[2] G. Karypis and V. Kumar. Multilevel algorithms for multi-constraint graph partitioning. In Proceedings of Supercomputing, 1998.
[3] G. Karypis and V. Kumar. Multilevel k-way partitioning scheme for irregular graphs. Journal of Parallel and
Distributed Computing, 48(1):96–129, 1998.
[4] G. Karypis and V. Kumar. A fast and high quality multilevel scheme for partitioning irregular graphs. SIAM
Journal on Scientific Computing, 20(1):359–392, 1999.
[5] B. W. Kernighan and S. Lin. An efficient heuristic procedure for partitioning graphs. The Bell System Technical
Journal, 49(2):291–307, 1970.
34
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement