## Combining Problem Reduction and Adaptive Multistart: A New Technique for Superior [terative Partitioning (1997)

Venue: | IEEE Transactions on Computer-Aided Design |

Citations: | 19 - 3 self |

### BibTeX

@ARTICLE{Hagen97combiningproblem,

author = {Lars W. Hagen and Andrew B. Kahng},

title = {Combining Problem Reduction and Adaptive Multistart: A New Technique for Superior [terative Partitioning},

journal = {IEEE Transactions on Computer-Aided Design},

year = {1997},

pages = {998}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract —VLSI netlist partitioning has been addressed chiefly by iterative methods (e.g., Kernighan–Lin and Fiduccia–Mattheyses) and spectral methods (e.g., Hagen–Kahng). Iterative methods are the de facto industry standard, but suffer diminished stability and solution quality when instances grow large. Spectral methods have achieved high-quality solutions, particularly for the ratio cut objective, but suffer excessive memory requirements and the inability to capture practical constraints (preplacements, variable module areas, etc.). This work develops a new approach to Fiduccia–Mattheyses (FM)-based iterative partitioning. We combine two concepts: 1) problem reduction using clustering and the two-phase FM methodology and 2) adaptive multistart, i.e., the intelligent selection of starting points for the iterative optimization, based on the results of previous optimizations. The resulting clustered adaptive multistart (CAMS) methodology [18] substantially improves upon previous partitioning results in the literature, for both unit module areas and actual module areas, and for both the min-cut bisection and minimum ratio cut objectives. The CAMS method is surprisingly fast and has very stable solution quality, even for large benchmark instances. It has been applied as the basis of a clustering methodology within an industry placement tool. I.

### Citations

528 |
Optimization by simulated annealing: An experimental evaluation; part i, graph partitioning
- Johnson, Aragon, et al.
- 1989
(Show Context)
Citation Context ...lt [23] [30] in order to attain "stability": we call this the random multi-start approach. The number of runs required to achieve stability via random multi-start grows very rapidly with pro=-=blem size [19]-=- [31]. Despite these shortfalls, iterative algorithms---and FM in particular---still possess many appealing advantages over the spectral approach. These advantages include smaller memory requirements,... |

434 |
Combinatorial Algorithms for Integrated Circuit Layout
- Lengauer
- 1990
(Show Context)
Citation Context ...reedy or a hill-climbing strategy. The iterative algorithm of Fiduccia and Mattheyses (FM) [13] (a variant of [21] that uses linear time per pass) is the method that is most widely used for bisection =-=[23]-=-. Wei and Cheng [30] use an adaptation of [13] to address the ratio cut objective. Spectral methods use eigenvectors of the Laplacian of a netlist-derived graph to deterministically find a partitionin... |

317 |
A Linear Time Heuristic for Improving Network Partitions
- Fiduccia, Mattheyses
- 1982
(Show Context)
Citation Context ...or Superior Iterative Partitioning Lars W. Hagen and Andrew B. Kahng Abstract VLSI netlist partitioning has been addressed chiefly by iterative methods (e.g. KernighanLin [21] and Fiduccia-Mattheyses =-=[13]-=-) and spectral methods (e.g. Hagen-Kahng [14]). Iterative methods are the de facto industry standard, but suffer diminished stability and solution quality when instances grow large. Spectral methods h... |

251 |
A Connectionist Machine For Genetic Hillclimbing
- Ackley
- 1987
(Show Context)
Citation Context ...e likely to reach the center of the big valley. Work in the genetic algorithms literature has discussed similar ideas within the context of "hybrid genetic-local search" [24] or "learn =-=as you search" [1]. By combi-=-ning genetic algorithms 1 This particular instance is from the class of "difficult" bisection inputs proposed by Bui et al. [5]. Specifically, a random graph in the class GBui (n; d; b) has ... |

226 | New spectral methods for ratio cut partitioning and clustering
- Hagen, Kahng
- 1992
(Show Context)
Citation Context ...gen and Andrew B. Kahng Abstract VLSI netlist partitioning has been addressed chiefly by iterative methods (e.g. KernighanLin [21] and Fiduccia-Mattheyses [13]) and spectral methods (e.g. Hagen-Kahng =-=[14]-=-). Iterative methods are the de facto industry standard, but suffer diminished stability and solution quality when instances grow large. Spectral methods have achieved high-quality solutions, particul... |

135 |
Spectral k-way ratio-cut partitioning and clustering
- Chan, Schlag, et al.
- 1993
(Show Context)
Citation Context ...total number of nets cut by the clustering, the sum of net-degrees ("pins") over all clusters, and the scaled cost value of the clustering (a multi-way generalization of ratio cut, due to Ch=-=an et al. [7]-=-). Although we have concentrated on the improved two-way partitionings afforded by CAMS, it 9 The MBC method [5] relies on random matchings, it probabilistically identifies clusters; see also the rece... |

132 |
Parallel genetic algorithms, population genetics and combinatorial optimization
- Mühlenbein
- 1989
(Show Context)
Citation Context ...evious solutions, and thus more likely to reach the center of the big valley. Work in the genetic algorithms literature has discussed similar ideas within the context of "hybrid genetic-local sea=-=rch" [24] or "learn as y-=-ou search" [1]. By combining genetic algorithms 1 This particular instance is from the class of "difficult" bisection inputs proposed by Bui et al. [5]. Specifically, a random graph in ... |

125 |
Multiple-way Network Partitioning
- Sanchis
- 1989
(Show Context)
Citation Context ...eneral clustering applications. It is possible that more "natural" clusterings may be obtained by using a CAMS-like methodology with an iterative multi-way clustering approach such as that o=-=f Sanchis [28], but this is beyond-=- the scope of the current work. We also note that in practice, CAMS finds its "optimum" solution after only a few iterations, seemingly before any effect of "adaptation" would have... |

119 | A new adaptive multistart technique for combinatorial global optimization - Boese, Kahng, et al. - 1994 |

119 |
Graph Bisection Algorithms with Good Average Case Behavior
- Bui, Chaudhuri, et al.
- 1987
(Show Context)
Citation Context ...int subsets, or clusters. Contracting the modules of each cluster into a single node induces a compacted, or condensed, representation of the original problem which may be easier to solve. Bui et al. =-=[5] [6] propo-=-sed the "matchingbased compaction" (MBC) algorithm, where the edges of a maximal random matching in the netlist graph induce a compacted instance of d n 2 e vertices, and the compaction is i... |

118 |
An Improved Min-Cut Algorithm For Partitioning VLSI Networks
- Krishnamurthy
- 1984
(Show Context)
Citation Context ... minor changes to the code but did not produce any noticeable change in solution quality. Also note that our FM implementation does not take advantage of the enhancements proposed by Krishnamurthy in =-=[22]-=-. The AMS results were derived from an implementation which follows the description in [3]. Our AMS implementation generates each new starting point stochastically, based on the k best-known local min... |

90 |
Evolution algorithms in combinatorial optimisation
- Muhlenbein, Gorges-schleuter, et al.
- 1992
(Show Context)
Citation Context ...nima plotted is 2,343. For each solution, we plot its cost against its average distance, in terms of single-vertex moves "shift-moves", to all 2,499 other solutions. with local search strate=-=gies, [4] [25]-=- [24] [29] showed that improved results were possible for the Traveling Salesman Problem (TSP) and partitioning. The basic approach in these works allow an iterative algorithm to improve each individu... |

66 |
An improved two-way partitioning algorithm with stable performance
- Cheng, Wei
- 1991
(Show Context)
Citation Context ...g AGG clusters in the two-phase FM approach yields bisections that are an average of 26.9% better than the results of running "flat" FM 200 times. Another 4 strong result is reported by Chen=-=g and Wei [8]-=- for partitioning with a 1:3 size ratio bound (as opposed to exact bisection); their stable two-way (STW) partitioning algorithm uses recursive FM-based ratio cut partitioning [30] to achieve a circui... |

63 | Genetic local search algorithms for the traveling salesman problem
- Ulder, Aarts, et al.
- 1990
(Show Context)
Citation Context ...ed is 2,343. For each solution, we plot its cost against its average distance, in terms of single-vertex moves "shift-moves", to all 2,499 other solutions. with local search strategies, [4] =-=[25] [24] [29]-=- showed that improved results were possible for the Traveling Salesman Problem (TSP) and partitioning. The basic approach in these works allow an iterative algorithm to improve each individual, either... |

62 | A Parallel Bottom-up Clustering Algorithm with Applications to Circuit Partitioning
- Cong, Smith
- 1993
(Show Context)
Citation Context ...s "natural clusters" via a self-tuning random walk in the circuit netlist (strongly connected regions of the netlist are detected by multiple revisitations of modules within the walk). Cong =-=and Smith [11]-=- generalize Bui et al.'s matching to a clique-finding scheme, and provide a parallel implementation. The strongest two-phase FM results for netlist bisection seem to be those of Alpert and Kahng [2], ... |

52 | Global Min-cuts in RNC, and Other Ramifications of a Simple Min-cut Algorithm
- Karger
- 1993
(Show Context)
Citation Context ...d two-way partitionings afforded by CAMS, it 9 The MBC method [5] relies on random matchings, it probabilistically identifies clusters; see also the recent randomized "contraction" strategy =-=of Karger [20]-=- for determining global minimum cuts in a network. The randomwalk approach of [16] is another example of probabilistic extraction of clusters. 16 Clustering Number of Total Total Scaled 3 Largest Clus... |

46 | A new approach to effective circuit clustering
- Hagen, Kahng
- 1992
(Show Context)
Citation Context ...which attempted to minimize the Rent parameter of the condensed netlist; [12] also used Rentbased clustering to improve the performance of a placement algorithm. Other two-phase FM approaches include =-=[16], whose pr-=-obabilistic RW-ST method finds "natural clusters" via a self-tuning random walk in the circuit netlist (strongly connected regions of the netlist are detected by multiple revisitations of mo... |

45 | Improving the Performance of the Kernighan-Lin and Simulated Annealing Graph Bisection Algorithms - Bui, Heighham, et al. - 1989 |

44 |
Towards Efficient Hierarchical Designs by Ratio Cut Partitioning
- Wei, Cheng
- 1989
(Show Context)
Citation Context ...bing strategy. The iterative algorithm of Fiduccia and Mattheyses (FM) [13] (a variant of [21] that uses linear time per pass) is the method that is most widely used for bisection [23]. Wei and Cheng =-=[30]-=- use an adaptation of [13] to address the ratio cut objective. Spectral methods use eigenvectors of the Laplacian of a netlist-derived graph to deterministically find a partitioning solution. The dete... |

33 | Geometric embeddings for faster and better multi-way netlist partitioning
- Alpert, Kahng
- 1993
(Show Context)
Citation Context ... [11] generalize Bui et al.'s matching to a clique-finding scheme, and provide a parallel implementation. The strongest two-phase FM results for netlist bisection seem to be those of Alpert and Kahng =-=[2]-=-, whose AGG algorithm applies geometric clustering to a multi-dimensional spectral embedding of the netlist. Using AGG clusters in the two-phase FM approach yields bisections that are an average of 26... |

29 |
An Efficient Heuristic Procedure for Partitioning of Electrical Circuits
- Kernighan, Lin
- 1970
(Show Context)
Citation Context ...ulti-Start: A New Technique for Superior Iterative Partitioning Lars W. Hagen and Andrew B. Kahng Abstract VLSI netlist partitioning has been addressed chiefly by iterative methods (e.g. KernighanLin =-=[21]-=- and Fiduccia-Mattheyses [13]) and spectral methods (e.g. Hagen-Kahng [14]). Iterative methods are the de facto industry standard, but suffer diminished stability and solution quality when instances g... |

28 |
Optimization strategies gleaned from biological evolution
- Brady
- 1985
(Show Context)
Citation Context ...l minima plotted is 2,343. For each solution, we plot its cost against its average distance, in terms of single-vertex moves "shift-moves", to all 2,499 other solutions. with local search st=-=rategies, [4]-=- [25] [24] [29] showed that improved results were possible for the Traveling Salesman Problem (TSP) and partitioning. The basic approach in these works allow an iterative algorithm to improve each ind... |

24 | Net Partitions Yield Better Module Partitions
- Cong, Kahng
- 1992
(Show Context)
Citation Context ...Spectral heuristics developed by Hagen and Kahng [15] use eigenvectors to define linear orderings of either modules or nets, and find good partitions by splitting the linear ordering. An extension in =-=[10] produced -=-results for ratio cut partitioning corresponding to an average of 28.8% improvement over the method of [30]. The main weakness of FM is that its solution quality is not very "stable", i.e., ... |

15 |
Improvements of a mincut partition algorithm
- Ng, Oldfield, et al.
- 1987
(Show Context)
Citation Context ...le. A two-phase FM methodology results: the FM algorithm is applied to the compacted netlist, and the result is re-expanded into a flat initial configuration for a second FM application. Ng et al. in =-=[26]-=- used a similar approach with a clustering algorithm which attempted to minimize the Rent parameter of the condensed netlist; [12] also used Rentbased clustering to improve the performance of a placem... |

9 | A two-level two-way partitioning algorithm - Wei, Cheng - 1990 |

8 |
Towards E cient Hierarchical Designs by Ratio Cut Partitioning
- Wei, Cheng
- 1989
(Show Context)
Citation Context ...mbing strategy. The iterative algorithm of Fiduccia and Mattheyses (FM) [13] (avariant of [21] that uses linear time per pass) is the method that is most widely used for bisection [23]. Wei and Cheng =-=[30]-=- use an adaptation of [13] to address the ratio cut objective. Spectral methods use eigenvectors of the Laplacian of a netlist-derived graph to deterministically nd a partitioning solution. The determ... |

6 |
A new optimization driven clustering algorithm for large circuits” (extended abstract
- Ding, Ho, et al.
- 1993
(Show Context)
Citation Context ... flat initial configuration for a second FM application. Ng et al. in [26] used a similar approach with a clustering algorithm which attempted to minimize the Rent parameter of the condensed netlist; =-=[12] also used-=- Rentbased clustering to improve the performance of a placement algorithm. Other two-phase FM approaches include [16], whose probabilistic RW-ST method finds "natural clusters" via a self-tu... |

6 |
A new approach to e ective circuit clustering
- Hagen, Kahng
- 1992
(Show Context)
Citation Context ...which attempted to minimize the Rent parameter of the condensed netlist; [12] also used Rentbased clustering to improve the performance of a placement algorithm. Other two-phase FM approaches include =-=[16]-=-, whose probabilistic RW-ST method nds \natural clusters" via a self-tuning random walk in the circuit netlist (strongly connected regions of the netlist are detected by multiple revisitations of modu... |

5 |
C 2003 private communication
- Chou
(Show Context)
Citation Context ...or the bisection experiments and the modified FM code used for the ratio cut experiments. Professor C.K. Cheng and Dr. Nan-Chi Chou of UC San Diego provided us with copies of their benchmark examples =-=[9]-=-. The anonymous reviewers provided many constructive suggestions and comments on the original draft of this paper. Part of this work (ABK) was performed during a Spring 1993 sabbatical at UC Berkeley;... |

5 |
spectral methods for ratio cut partitioning and clustering
- “New
- 1992
(Show Context)
Citation Context ...g solution. The determinism of spectral methods is appealing, and the need for only one run keeps CPU requirements reasonable as instances grow large. Spectral heuristics developed by Hagen and Kahng =-=[15]-=- use eigenvectors to define linear orderings of either modules or nets, and find good partitions by splitting the linear ordering. An extension in [10] produced results for ratio cut partitioning corr... |

3 |
A Timing-Driven N-Way Chip and Multi-Chip Partitioner
- Roy, Sechen
- 1993
(Show Context)
Citation Context ...tion in problem size which allows most of the optimization effort to address the "difficult part" of the problem. Another assumption in previous work is that clusters should be of uniform si=-=ze (e.g., [27]-=- p. 243, or the original method of [5] which employs iterative matching). However, our work shows clearly that uniform cluster size need not be a dominant concern: Table 4 shows some example profiles ... |

3 |
Global Min-Cuts in RNC, and Other Rami cations of a Simple Min-Cut Algorithm
- Karger
- 1993
(Show Context)
Citation Context ...ved two-way partitionings a orded by CAMS, it 9 The MBC method [5] relies on random matchings, it probabilistically identi es clusters; see also the recent randomized \contraction" strategy of Karger =-=[20]-=- for determining global minimum cuts in a network. The randomwalk approach of [16] is another example of probabilistic extraction of clusters. 16sClustering Number of Total Total Scaled 3 Largest Clus... |

2 |
On the Intrinsic Rent Parameter and New Spectra-Based Methods for Wireability Estimation
- Hagen, Kahng, et al.
- 1992
(Show Context)
Citation Context ...or netlist bisection may make recursive bipartitioning-based placement a more appealing methodology. Moreover, the substantial improvement in solution quality may well put into question previous work =-=[17] which com-=-pared "partitioning-based Rent parameters" of various algorithms and found ratio-cut hierarchies to be superior. Finally, we note that CAMS is also naturally suited for clustering applicatio... |

2 | two-level two-way partitioning algorithm - “A - 1990 |

1 |
Method of Solving Discrete Global Optimization Problems," U.S. Patent Application No. 08/296,808, patent pending (filing date
- Hagen, Kahng
- 1993
(Show Context)
Citation Context ...art, i.e., the intelligent selection of starting points for the iterative optimization, based on the results of previous optimizations. The resulting Clustered Adaptive Multi-Start (CAMS) methodology =-=[18]-=- substantially improves upon previous partitioning results in the literature, for both unit module areas and actual module areas, and for both the min-cut bisection and minimum ratio cut objectives. T... |

1 |
new approach to effective circuit clustering
- “A
- 1992
(Show Context)
Citation Context ...ed to minimize the Rent parameter of the condensed netlist; Ding et al. [12] also used Rent-based clustering to improve the performance of a placement algorithm. Other two-phase FM approaches include =-=[16]-=-, whose probabilistic RW-ST method finds “natural clusters” via a selftuning random walk in the circuit netlist (strongly connected regions of the netlist are detected by multiple revisitations of mod... |

1 | A timing-driven N -way chip and multi-chip partitioner - Roy, Sechen - 1993 |