This package contains the multi-way hypergraph partitioning algorithms: FMS (Fiduccia-Mattheyses-Sanchis), PLM (Partitioning by Locked Moves), PFM (Partitioning by Free Moves), SA (Simulated Annealing - 2 versions), and RSA (Simulated Annealing with ratio cut model - 2-way partitioning only), as detailed in [DaAy97].
The hypergraph partitioning problem is defined as follows: Given an input hypergraph, partition it into a given number of almost equal-sized parts in such a way that the cutsize, i.e., the sum of the net weights whose end cells are in different parts, is minimized. This problem has many variations as well as has important applications in many areas. Unfortunately, the problem is NP-hard so the algorithms in this package are heuristics (but they work very very well).
If you do not know what a hypergraph is, remember this: In a graph, you have vertices and edges, where every edge connects two vertices; in a hypergraph, you have vertices and hyperedges, where every hyperedge connects two or more vertices. Since hypergraphs can model electronic circuits well, a hypergraph is often said to have cells and nets instead of vertices and hyperedges. In a circuit cell where a net is connected is called a pin. You may see the cell, net, and pin terminology in my code.
The basis for these algorithms go back to the Kernighan-Lin (KL) algorithm for hypergraph partitioning. The KL algorithm produces very good partitions but it is slow. The Fiduccia-Mattheyses (FM) algorithm is not only a faster version of the KL algorithm but it also generalizes the KL algorithm to run on hypergraphs. Sanchis generalized the FM algorithm from 2-way partitioning to multi-way partitioning.
All these KL-based algorithms work in multiple passes over the number of vertices; in each pass, these algorithms find the best destination part for a cell and lock this cell from moving again in the rest of the current pass. Realizing the limitations of this "locking" mechanism, I devised ways to relax this mechanism, resulting in the Partitioning by Locked Moves (PLM) algorithm and the Partitioning by Free Moves (PFM) algorithm. The PLM algorithm still uses locking but it goes through multiple phases of locking and unlocking within a pass over the input hypergraph. The PFM algorithm does not use locking at all; it uses a way to penalize the moves to march towards a local minimum. For more details, please refer to [DaAy97].
I originally developed this package in C during my MSc study (around 1991-1993). Before putting this package on github, I converted my code to ANSI C (c99) and cleaned it quite a bit.
In my code, you may notice some biologically inspired names such as 'population', 'chromosome', 'allele', etc. These names actually come from genetic algorithms. During my MSc years, my eventual goal was to implement graph and hypergraph partitioning using genetic algorithms on a hypercube connected parallel computer (from Intel). Once I discovered the limitations of the locking mechanism, I changed my research direction towards what would become PLM and PFM. By the way, I did still work on genetic algorithms but in a different setting (applying genetic algorithms to synthesize unsupervised learning algorithms).
This package is available on an "as is" basis. I do not say or imply that it will be useful for whatever you want to do with it. It may also contain bugs, and I assume no responsibility for any potential problems associated with its use. You can use this package free of charge in academic research and teaching. For any commercial use, contact Ali Dasdan at [email protected]. See the COPYRIGHT section below.
At the top level, you see two directories: 'src' and 'input'. 'src' contains all the source files and 'input' contains the input hypergraphs used for testing the executables. Under 'src', there is one directory, called 'share', for the shared source files as well as one directory for each algorithm: 'fms', 'plm', 'pfm', and 'sa'. Note that there will be three executables under the 'sa' directory, one for each version of simulated annealing.
Under 'src', type 'make' (or 'gmake') to build each executable under its own directory. The executables are all have .x extension: ad_fms.x, ad_plm.x, ad_pfm.x, ad_sa1.x, ad_sa2.x, and ad_rsa.x.
Go to the respective directory for an algorithm and type the name of one of the executables in your command line to get the usage information. At minimum, each executable requires the input hypergraph and the number of parts to partition the hypergraph. PLM and PFM require additional parameters to create different versions of them, which trade off runtime for partition quality.
For example, a run of FMS (in 'ad_fms.x') to partition the input hypergraph 'p9' into two parts will produce this output:
> ad_fms.x input/hp9 2 123456
SEED = 123456 fname = input/p9
pass_no = 7 Final cutsize = 85 Check cutsize = 85
This output shows that FMS took 7 passes over the cells of the input hypergraph 'hp9' when started with a seed of 123456 (which is needed to make the results repeatable). FMS found a cutsize of 85, which is correct as FMS and the other programs will check every cutsize they report for correctness. That is, each of the executables are self validating.
Under 'src', type 'make test' to test each executable on the input hypergraphs under the 'input' directory. The result will be a 'pass' or a 'fail'.
Under 'src', type 'make clean' to clean all temporary files including the object files and the executables.
The input file format is explained below using a very simple hypergraph.
> cat input/hp1
6
7
14
1 2 0 1
1 2 1 2
1 2 2 0
1 2 3 4
1 2 4 5
1 2 5 3
1 2 0 3
1
1
1
1
1
1
The first three lines give the number of cells, the number of nets, and the number of pins (or endpoints), respectively. Thus, 'p1' has 6 vertices, 7 edges, and 14 pins.
The following 7 lines describe the nets, one net per line. The first number is the net weight; the second number is the number of pins on this net (always equal to 2 for graphs)); the third number is the source cell; and the fourth number is the target cell. For example, the first net from cell 0 to cell 1 has a weight of 1.
The last 6 lines describe the cell weights.
Please cite this reference if you use my programs in your research work.
@article{DaAy97,
author = {Ali Dasdan and C. Aykanat},
title = {Two Novel Circuit Partitioning Algorithms Using Relaxed Locking},
journal = {IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD)},
volume = {16},
number = {2},
year = {1997},
pages = {169-178},
url = {http://yoksis.bilkent.edu.tr/doi_getpdf/articles/10.1109-43.573831.pdf},
}
COPYRIGHT C 1991 - Ali Dasdan
Permission to use for non-commercial purposes is granted provided that proper acknowledgments are given. For a commercial licence, contact Ali Dasdan at [email protected].
This software is provided on an "as is" basis, without warranties or conditions of any kind, either express or implied including, without limitation, any warranties or conditions of title, non-infringement, merchantability or fitness for a particular purpose.