Graph and Hypergraph Partitioning For Parallel Computing

Graph and Hypergraph Partitioning
for Parallel Computing
Edmond Chow
School of Computational Science and Engineering
Georgia Institute of Technology
June 29, 2016

Graph and hypergraph partitioning
References:
I Graph Partitioning for High Performance Scientific
Simulations, Kirk Schloegel, George Karypis, and Vipin
Kumar, 2000
I Hypergraph-Partitioning-Based Decomposition for Parallel
Sparse-Matrix Vector Multiplication, Umit V. Catalyurek and
Cevdet Aykanat, 1999
Some figures below are from the above references.

Distributed sparse matrix-vector multiplication
    
y1 A11 A12 · · · A1p x1
y2  A21 A22 · · · A2p   x2 
 ..  =  ..
    
.. .. ..   .. 
.  . . . .  . 
yp Ap1 Ap2 · · · App xp
Processor i stores Ai? , xi and computes yi
(block row-wise partitioning).
Distributed sparse matrix-vector multiplication
    
y1 A11 A12 ··· A1p x1
y2  A21 A22 ··· A2p  x2 
 ..  =  ..
    
.. .. ..   .. 
.  . . . .  . 
yp Ap1 Ap2 ··· App xp
On processor i:
1. Send components of xi needed by other processors
2. Compute ylocal = Aii xi
3. Receive components of xj from other processors j needed
by processor i
4. Compute yext = Ai,ext xext
5. Form yi = ylocal + yext
If processor i contains nonzeros in column k , then processor i

needs the k th scalar component of x.
Graph model (for p = 2)
The graph model shows which nonzeros (edges) induce

communication in SpMV.
Graph partitioning problem
Given G = (V , E), and number of parts p,
find the partitioning of the vertices P1 , P2 , . . . , Pp , where Pi
contains a subset of the vertices in V ,
such that the Pi are disjoint and complete
and X
1
|Pi | ≤ |Pj | (1 + )
k
for all i (number of vertices in each partition is approximately
equal) and the number of cut edges is minimized. A cut edge
straddles two partitions.
This problem is NP-hard.

Multilevel graph partitioning
1. Coarsen the graph by collapsing heavy edges. Vertex

weights and edge weights are updated. Vertex and edge
weights are initially 1.
2. Repeat until the graph is small enough, then partition the
coarsest level graph by some method.
3. Uncoarsen. After each uncoarsening step, refine the
partition (using, e.g., Kerninghan-Lin).
Software: METIS, Chaco, SCOTCH.

Graph coarsening
I Edges are collapsed by using a

heavy-edge matching.
I A matching is a set of edges such that
no two edges share a vertex.
Levelized nested dissection
Possibility for partitioning the coarsest level graph.
Kernighan-Lin
Find pairs of nodes on opposites sides of the partition that can
change partitions so that the edge-cut is improved. Repeat until
all nodes have been moved and then restore the best
configuration. Repeat above procedure until the best
configuration found is unchanged.
Fiduccia-Mattheyses
Compute gain for each node. Change partition of node with
largest gain. Update gains. Gains are stored in priority queue
and a node that has moved will not move again until all others
have moved.
Fiduccia-Mattheyses
Deficiency with graph partitioning
The cut size does not measure the communication volume.
This deficiency can be addressed with hypergraph partitioning.

Hypergraph
Hypergraph H = (V , N), where V is the set of vertices and N is
the set of hyperedges (or nets).
A hyperedge can connect more than 2 vertices.
Hypergraph of a sparse matrix
I Rows are vertices.

I Columns are hyperedges. Nonzeros in column j
correspond to the vertices connected by the hyperedge.
Hypergraph
Hypergraph partitioning problem
Group the vertices into approximately equal partitions such that
X
min (λi − 1)
ni is cut
where λi is the number of parts connected by hyperedge i.
Each hyperedge belongs in a partition, so the number of

neighbors is λi − 1.
NP-hard problem.
Hypergraph partitioning
Hypergraph partitioning
Multilevel algorithm: coarsen hyperedges, refinement with
hyperedges.
Software: PaToH, hMETIS.
For FEM meshes, regular graph partitioning is usually good

enough. For very unstructured meshes, e.g., from circuit
simulation, hypergraph partitioning is needed.
Exercise
Consider this graph consisting of 8 vertices and a partition into
two parts (separated by the dotted line):
:
1-----3--:--5-----7
| | : | |
| | : | |
2-----4--:--6-----8
:
What is the hypergraph corresponding to this graph? How

many hyperedges are cut by the given partitioning?

Graph and Hypergraph Partitioning For Parallel Computing

Uploaded by

Copyright:

Available Formats

Graph and Hypergraph Partitioning For Parallel Computing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Graph and Hypergraph Partitioning For Parallel Computing

Uploaded by

Copyright:

Available Formats

Graph and Hypergraph Partitioning

for Parallel Computing

June 29, 2016

Some figures below are from the above references.

If processor i contains nonzeros in column k , then processor i

The graph model shows which nonzeros (edges) induce

This problem is NP-hard.

1. Coarsen the graph by collapsing heavy edges. Vertex

Software: METIS, Chaco, SCOTCH.

I Edges are collapsed by using a

The cut size does not measure the communication volume.

This deficiency can be addressed with hypergraph partitioning.

A hyperedge can connect more than 2 vertices.

Hypergraph of a sparse matrix

I Rows are vertices.

where λi is the number of parts connected by hyperedge i.

Each hyperedge belongs in a partition, so the number of

Software: PaToH, hMETIS.

For FEM meshes, regular graph partitioning is usually good

What is the hypergraph corresponding to this graph? How

You might also like