METIS for Python

Wrapper for the METIS library for partitioning graphs (and other stuff).

This library is unrelated to PyMetis, except that they wrap the same library. PyMetis is a Boost Python extension, while this library is pure python and will run under PyPy and interpreters with similarly compatible ctypes libraries.

NetworkX is recommended for representing graphs for use with this wrapper, but it isn’t required. Simple adjacency lists are supported as well.

The function of primary interest in this module is part_graph().

Other objects in the module may be of interest to those looking to mangle their graph datastructures into the required format. Examples of this include the networkx_to_metis() and adjlist_to_metis() functions. These are automatically called by part_graph(), so there is little need to call them manually.

See the BitBucket repository for updates and issue reporting.


It’s on PyPI, so installation should be as easy as:

pip install metis
easy_install metis

METIS itself is not included with this wrapper. Get it here.

Note that the shared library is needed, and isn’t enabled by default by the configuration process. Turn it on by issuing:

make config shared=1

Your operating system’s package manager might know about METIS, but this wrapper was designed for use with METIS 5. Packages with METIS 4 will not work.

This wrapper uses a few environment variables:


This wrapper uses Python’s ctypes module to interface with the METIS shared library. If it is unable to automatically locate the library, you may specify the full path to the library file in this environment variable.


The sizes of the idx_t and real_t types are not easily determinable at runtime, so they can be provided with these environment variables. The default value for each of these (at both compile time and in this library) is 32, but they may be set to 64 if desired. If the values do not match what was used to compile the library, Bad Things(TM) will occur.


>>> import networkx as nx
>>> import metis
>>> G = metis.example_networkx()
>>> (edgecuts, parts) = metis.part_graph(G, 3)
>>> colors = ['red','blue','green']
>>> for i, p in enumerate(parts):
...     G.node[i]['color'] = colors[p]
>>> nx.write_dot(G, '') # Requires pydot or pygraphviz
strict graph G {
0 [color=blue];
1 [color=blue];
2 [color=blue];
3 [color=blue];
4 [color=blue];
5 [color=blue];
6 [color=blue];
7 [color=red];
8 [color=red];
9 [color=red];
10 [color=red];
11 [color=red];
12 [color=red];
13 [color=green];
14 [color=green];
15 [color=green];
16 [color=green];
17 [color=green];
18 [color=green];
0 -- 1;
0 -- 2;
0 -- 3;
0 -- 4;
4 -- 5;
5 -- 6;
6 -- 13;
6 -- 7;
7 -- 8;
8 -- 9;
8 -- 10;
8 -- 11;
8 -- 12;
13 -- 14;
14 -- 15;
15 -- 16;
15 -- 17;
15 -- 18;
metis.part_graph(graph, nparts=2, tpwgts=None, ubvec=None, recursive=False, **opts)[source]

Perform graph partitioning using k-way or recursive methods.

Returns a 2-tuple (objval, parts), where parts is a list of partition indices corresponding and objval is the value of the objective function that was minimized (either the edge cuts or the total volume).

  • graph

    may be a NetworkX graph, an adjacency list, or a METIS_Graph named tuple. To use the named tuple approach, you’ll need to read the METIS manual for the meanings of the fields.

    See networkx_to_metis() for help and details on how the graph is converted and how node/edge weights and sizes can be specified.

    See adjlist_to_metis() for information on the use of adjacency lists. The extra nodew and nodesz keyword arguments of that function may be given directly to this function and will be forwarded to the converter. Alternatively, a dictionary can be provided as graph and its items will be passed as keyword arguments.

  • nparts – The target number of partitions. You might get fewer.
  • tpwgts

    Target partition weights. For each partition, there should be one (float) weight for each node constraint. That is, if nparts is 3 and each node of the graph has two weights, then tpwgts might look like this:

    [(0.5, 0.1), (0.25, 0.8), (0.25, 0.1)]

    This list may be provided flattened. The internal tuples are for convenience. The partition weights for each constraint must sum to 1.

  • ubvec – The load imalance tolerance for each node constraint. Should be a list of floating point values each greater than 1.
  • recursive – Determines whether the partitioning should be done by direct k-way cuts or by a series of recursive cuts. These correspond to METIS_PartGraphKway() and METIS_PartGraphRecursive() in METIS’s C API.

Any additional METIS options may be specified as keyword parameters.

For k-way clustering, the appropriate options are:

objtype   = 'cut' or 'vol'
ctype     = 'rm' or 'shem'
iptype    = 'grow', 'random', 'edge', 'node'
rtype     = 'fm', 'greedy', 'sep2sided', 'sep1sided'
ncuts     = integer, number of cut attempts (default = 1)
niter     = integer, number of iterations (default = 10)
ufactor   = integer, maximum load imbalance of (1+x)/1000
minconn   = bool, minimize degree of subdomain graph
contig    = bool, force contiguous partitions
seed      = integer, RNG seed
numbering = 0 (C-style) or 1 (Fortran-style) indices
dbglvl    = Debug flag bitfield

For recursive clustering, the appropraite options are:

ctype     = 'rm' or 'shem'
iptype    = 'grow', 'random', 'edge', 'node'
rtype     = 'fm', 'greedy', 'sep2sided', 'sep1sided'
ncuts     = integer, number of cut attempts (default = 1)
niter     = integer, number of iterations (default = 10)
ufactor   = integer, maximum load imbalance of (1+x)/1000
seed      = integer, RNG seed
numbering = 0 (C-style) or 1 (Fortran-style) indices
dbglvl    = Debug flag bitfield

See the METIS manual for specific meaning of each option.


Convert NetworkX graph into something METIS can consume The graph may specify weights and sizes using the following graph attributes:

  • edge_weight_attr
  • node_weight_attr (multiple names allowed)
  • node_size_attr

For example:

>>> G.adj[0][1]['weight'] = 3
>>> G.node[0]['quality'] = 5
>>> G.node[0]['specialness'] = 8
>>> G.graph['edge_weight_attr'] = 'weight'
>>> G.graph['node_weight_attr'] = ['quality', 'specialness']

If node_weight_attr is a list instead of a string, then multiple node weight labels can be provided.

All weights must be integer values. If an attr label is specified but a node/edge is missing that attribute, it defaults to 1.

If a graph attribute is not provided, no defaut is used. That is, if edge_weight_attr is not set, then 'weight' is not used as the default, and the graph will appear unweighted to METIS.

metis.adjlist_to_metis(adjlist, nodew=None, nodesz=None)[source]

Rudimentary adjacency list converter. Primarily of use if you don’t have or don’t want to use NetworkX.

  • adjlist

    A list of tuples. Each list element represents a node or vertex in the graph. Each item in the tuples represents an edge. These items may be single integers representing neighbor index, or they may be an (index, weight) tuple if you want weighted edges. Default weight is 1 for missing weights.

    The graph must be undirected, and each edge must be represented twice (once for each node). The weights should be identical, if provided.

  • nodew – is a list of node weights, and must be the same size as adjlist if given. If desired, the elements of nodew may be tuples of the same size (>= 1) to provided multiple weights for each node.
  • nodesz – is a list of node sizes. These are relevant when doing volume-based partitioning.

Note that all weights and sizes must be non-negative integers.

Indices and tables