## Local Search for Final Placement in VLSI Design (2001)

### Cached

### Download Links

- [www.sigda.org]
- [www.cecs.uci.edu]
- [www.cs.york.ac.uk]
- [www.diku.dk]
- DBLP

### Other Repositories/Bibliography

Citations: | 6 - 0 self |

### BibTeX

@MISC{Faroe01localsearch,

author = {Oluf Faroe and David Pisinger and Martin Zachariasen},

title = {Local Search for Final Placement in VLSI Design},

year = {2001}

}

### OpenURL

### Abstract

A new heuristic is presented for the general cell placement problem where the objective is to minimize total bounding box netlength. The heuristic is based on the Guided Local Search (GLS) metaheuristic. GLS modifies the objective function in a constructive way to escape local minima. Previous attempts to use local search on final (or detailed) placement problems have often failed as the neighborhood quickly becomes too excessive for large circuits. Nevertheless, by combining GLS with Fast Local Search it is possible to focus the search on appropriate sub-neighborhoods, thus reducing the time complexity considerably.

### Citations

1105 |
An efficient heuristic procedure for partitioning graphs
- Kernighan, S
- 1970
(Show Context)
Citation Context ... is to a large extent ignored. A classical approach is to use hypergraph partitioning in which the modules are recursively divided into subsets for which the need for inter-communication is minimized =-=[1, 14]-=-. A more recent approach is to optimize total netlength directly using the clique or star netmodel, both of which result in a quadratic optimization problem that can be solved quickly. Combined with h... |

109 | The TimberWolf placement and routing package - Sechen, Sangiovanni-Vincentelli |

108 | Efficient and effective placement for very large circuits - Sun, Sechen - 1995 |

87 |
VLSI Placement and Global Routing Using Simulated Annealing
- Sechen
- 1988
(Show Context)
Citation Context ...ristic [28, 29] for controlling the search. The neighbourhood structure is simple: Flipping and/or moving a single module along one of the coordinate axes. This neighbourhood has previously been used =-=[12, 18, 21, 23, 32]-=-, in particular in conjunction with simulated annealing. The weakness of all these algorithms is the slow convergence towards good solutions | which is an inherent feature of simulated annealing. By c... |

84 | Multilevel circuit partitioning
- Alpert, Huang, et al.
- 1997
(Show Context)
Citation Context ... is to a large extent ignored. A classical approach is to use hypergraph partitioning in which the modules are recursively divided into subsets for which the need for inter-communication is minimized =-=[1, 14]-=-. A more recent approach is to optimize total netlength directly using the clique or star netmodel, both of which result in a quadratic optimization problem that can be solved quickly. Combined with h... |

74 | Algorithms for Large-Scale Flat Placementâ€ť, to appear - Vygen - 1997 |

72 | An O-tree representation of non-slicing floorplan and its application - Guo, Chen, et al. - 1999 |

60 |
Simulated Annealing for VLSI Design
- Wong, Leong, et al.
- 1988
(Show Context)
Citation Context ...ristic [28, 29] for controlling the search. The neighbourhood structure is simple: Flipping and/or moving a single module along one of the coordinate axes. This neighbourhood has previously been used =-=[12, 18, 21, 23, 32]-=-, in particular in conjunction with simulated annealing. The weakness of all these algorithms is the slow convergence towards good solutions | which is an inherent feature of simulated annealing. By c... |

53 | Tsang: Guided Local Search and its application to the Travelling Salesman Problem
- Voudouris
- 1999
(Show Context)
Citation Context ...ent problem. The algorithm both takes the packing problem, i.e. placing the modules disjointly, and the total BB netlength into account. The algorithm uses the Guided Local Search (GLS) metaheuristic =-=[28, 29]-=- for controlling the search. The neighbourhood structure is simple: Flipping and/or moving a single module along one of the coordinate axes. This neighbourhood has previously been used [12, 18, 21, 23... |

49 | Fast Local Search and Guided Local Search and Their Application to British Telecomâ€™s Workforce Scheduling Problem
- Tsang, Voudouris
- 1997
(Show Context)
Citation Context ...ent problem. The algorithm both takes the packing problem, i.e. placing the modules disjointly, and the total BB netlength into account. The algorithm uses the Guided Local Search (GLS) metaheuristic =-=[28, 29]-=- for controlling the search. The neighbourhood structure is simple: Flipping and/or moving a single module along one of the coordinate axes. This neighbourhood has previously been used [12, 18, 21, 23... |

46 | Guided Local Search for the Vehicle Routing Problems With Time Windows
- Kilby, Prosser, et al.
- 1999
(Show Context)
Citation Context ...l objective of minimizing total netlength, it is natural to consider an extension of the algorithm for the packing problem. The GLS metaheuristic has proven to be eective on a wide range of problems [=-=15, 27, 28, 2-=-9]. GLS can be applied to any combinatorial optimization problem given by a solution space X for which an objective function value f(x) and neighborhood N (x) X is dened for every solution x 2 X . In... |

42 |
Guided local search for combinatorial optimization problems
- VOUDOURIS
- 1997
(Show Context)
Citation Context ... costs, it is necessary to retract penalties dynamically; penalized features may become less unattractive as the search progresses. Several retract strategies were proposed and evaluated by Voudouris =-=[26]-=-. One possibility is to reset all penalties at certain intervals during the search, denoted the Reset strategy. Another option is to use a short term memory: record the last t features which are penal... |

33 | Tsang: Partial Constraint Satisfaction Problems and Guided Local Search
- Voudouris
- 1996
(Show Context)
Citation Context ...l objective of minimizing total netlength, it is natural to consider an extension of the algorithm for the packing problem. The GLS metaheuristic has proven to be eective on a wide range of problems [=-=15, 27, 28, 2-=-9]. GLS can be applied to any combinatorial optimization problem given by a solution space X for which an objective function value f(x) and neighborhood N (x) X is dened for every solution x 2 X . In... |

22 |
New Algorithms for the Placement and Routing of Macro Cells
- Swartz, Sechen
- 1990
(Show Context)
Citation Context ...ristic [28, 29] for controlling the search. The neighbourhood structure is simple: Flipping and/or moving a single module along one of the coordinate axes. This neighbourhood has previously been used =-=[12, 18, 21, 23, 32]-=-, in particular in conjunction with simulated annealing. The weakness of all these algorithms is the slow convergence towards good solutions | which is an inherent feature of simulated annealing. By c... |

15 |
Some experiments with simulated annealing techniques for packing problems
- Dowsland
- 1993
(Show Context)
Citation Context ...constraint that no modules may overlap, and instead penalize overlap in the objective function. A similar solution space was introduced by Jepsen et al. [12] for the placement problem and by Dowsland =-=[5]-=- and Faroe et al. [7] in the context of packing. 7 Placement area 1 2 3 3 2 Overlap Figure 2: Illustration of possible moves of a single module. The moves are (1) change of orientation combined with e... |

13 |
Clustering based simulated annealing for standard cell placement
- Mallela, Grover
- 1988
(Show Context)
Citation Context |

8 | Guided Local Search for the ThreeDimensional Bin Packing Problem
- Faroe, Pisinger, et al.
- 1999
(Show Context)
Citation Context ...ce the metaheuristic Guided Local Search (GLS) and the concept of Fast Local Search (FLS). The use of GLS for placement was motivated by recent results on packing problems in two and three dimensions =-=[7]-=-. Since a feasible placement is a packing of the modules with the additional objective of minimizing total netlength, it is natural to consider an extension of the algorithm for the packing problem. T... |

8 |
Macro placement by monte carlo annealing
- Jepsen, Jr
- 1983
(Show Context)
Citation Context |

8 |
Integrated placement for mixed macro cell and standard cell designs
- Upton, Samii, et al.
- 1990
(Show Context)
Citation Context ...ctive function. A reduction in the overlap term will nearly always increase the bounding box value and vice versa. As the bounding box term is linear while the overlap term is quadratic, Upton et al. =-=[25]-=- note that this makes it dicult to assign a suitable value tos. A linear overlap function might therefore have been more suitable (e.g., the perimeter of the rectangular intersection), but experiments... |

7 | Worst-case ratios of networks in the rectilinear plane
- Brenner, Vygen
- 2000
(Show Context)
Citation Context ...ing. A drawback with the RSMT netmodel is that the RSMT problem is NP-hard and thus inappropriate to use in a local search heuristic. Based on the theoretical and empirical study by Brenner and Vygen =-=[2, 3-=-] we chose to use the Bounding Box netmodel. This netmodel is dened as the length of the half perimeter of the smallest box which contains all the points in N P . Due to its simplicity and good approx... |

6 |
Plazierung im VLSI-Design und ein zweidimensionales Zerlegungsproblem
- Vygen
- 1997
(Show Context)
Citation Context ...zation problem that can be solved quickly. Combined with hierarchical partitioning that takes the density of the regions into account, placements with a a limited amount of overlap can be constructed =-=[30, 31]-=-. The task of thesnal placement problem is to turn an infeasible placement (e.g., a global placement) into a feasible placement in which all modules are disjoint. One objective is for example to minim... |

4 |
Verdrahtung im VLSI-Design: Spezielle Teilprobleme und ein sequentielles Losungsverfahren
- Hetzel
- 1995
(Show Context)
Citation Context ...ve dierent netmodels for asnite net N P with 5 points. The RSMT , RMST and BB are illustrated with the L 1 metric, while CL and ST are illustrated with the L 2 metric. Experimental results by Hetzel [=-=10]-=- on industrial VLSI circuits show that the RSMT netmodel very closely estimates the actual netlength after routing. A drawback with the RSMT netmodel is that the RSMT problem is NP-hard and thus inapp... |

4 |
Ecient and eective placement for very large circuits
- Sun, Sechen
- 1995
(Show Context)
Citation Context ...used as benchmarks: O-Tree [9], TimberWolf v.1.2 commercial edition (TW) [24] and XQ [30]. The core in TW is a simulated annealing placement heuristic which is specialized for standard cell placement =-=[22]-=-. The XQ placement program uses quadratic optimization and sophisticated partitioning. In Section 5.1 we describe the computational environment and the benchmarks used in our computational study. The ... |

3 |
Algorithms for large-scale at placement
- Vygen
- 1997
(Show Context)
Citation Context ...zation problem that can be solved quickly. Combined with hierarchical partitioning that takes the density of the regions into account, placements with a a limited amount of overlap can be constructed =-=[30, 31]-=-. The task of thesnal placement problem is to turn an infeasible placement (e.g., a global placement) into a feasible placement in which all modules are disjoint. One objective is for example to minim... |

1 |
Plazierung im VLSI-design
- Brenner
- 2000
(Show Context)
Citation Context ...ing. A drawback with the RSMT netmodel is that the RSMT problem is NP-hard and thus inappropriate to use in a local search heuristic. Based on the theoretical and empirical study by Brenner and Vygen =-=[2, 3-=-] we chose to use the Bounding Box netmodel. This netmodel is dened as the length of the half perimeter of the smallest box which contains all the points in N P . Due to its simplicity and good approx... |

1 |
On Steiner trees for bounded point sets. Geom. Dedicata
- Chung, Graham
- 1981
(Show Context)
Citation Context ...viation from RSMT is bounded by 4 BB(N P ) RSMT(N P ) and RSMT(N P ) (d p jN P j e + 3 2 )=2 BB(N P ). For jN P j 3 we have RSMT(N P ) = BB(N P ). These bounds were proven by Chung and Graham [4] with additional notes by Brenner and Vygen [3]. To further correct the BB netmodel a multiplier function w BB (N) may be applied. Brenner [2] empirically show that using the multiplier function w BB ... |

1 |
Placement of modules in VLSI layout
- Faroe
- 2000
(Show Context)
Citation Context ...7 10 jN j 1 4 means that the Bounding Box netmodel gets very close to the RSMT. In our implementation, however, no multiplier function is used since this only makes a very small dierence in practice [=-=6]-=-. Also, using no multiplier function makes it easier to compare our solution values with those from the literature. 3 Guided Local Search and Fast Local Search In this section we introduce the metaheu... |

1 |
An O-Tree representation of non-slicing and its applications
- Guo, Cheng, et al.
- 1999
(Show Context)
Citation Context ...y better solutions than existing algorithms from the literature. For some instances the total netlength is reduced by more than 20 percent when compared to the recent results for the O-Tree algorithm =-=[9-=-]. The paper is organized as follows. In Section 2 we dene thesnal placement problem in details. Section 3 introduces GLS and FLS, and in Section 4 we present the details of applying GLS and FLS to th... |

1 |
Cell Placement Using Constructive and Iterative Improvement Methods
- Kennings
- 1997
(Show Context)
Citation Context ...ion. As the regions overlap, modules are allowed to traverse between the regions and move over the whole placement area. The technique has similarities to the tiles of overlapping windows by Kennings =-=[13]-=-. The overlapping regions are constructed such that they contain at least M modules, where 13 M is an experimentally determined constant (see Section 5.2). Initially the whole placement area A is spli... |