multi-level edge separator using a hybrid combinatorial ... · function combining fm gains with a...
TRANSCRIPT
![Page 1: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/1.jpg)
Multi-Level Edge Separator Using a Hybrid Combinatorial-Quadratic
Programming Approach
Nuri Yeralan, Timothy A. Davis, and William Hager
University of Florida
![Page 2: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/2.jpg)
Graph Partitioning Problem
• Partition an undirected graph into 2 subgraphs
• Minimize the sum of edge weights in cut set
• Maintain target partition balance
• NP Hard
![Page 3: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/3.jpg)
Multi-Level Partitioning
• Coarsen the input to reduce problem size
• Cut the coarsest graph
• Refine back to the original input
![Page 4: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/4.jpg)
Multi-Level Partitioning
coarsen
coarsen
cut
refine
refine
![Page 5: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/5.jpg)
Multi-Level Partitioning
• Variations
– “V” cycle
• Single sweep, down and up
– “W” cycle
• Refine & alternate between deepest levels
– Randomization
• Bifurcate on way down
• Select best of alternatives on way up
• Parallelism!
![Page 6: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/6.jpg)
Coarsening
coarsen
• Goals
– Reduce problem size
– Remove heavy edges
![Page 7: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/7.jpg)
Matching
• Glue two or more vertices together
– Update connectivity as set union of adjacency lists summing edge weights
– Sum the node weights
![Page 8: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/8.jpg)
Matching Strategies
• Random Matching (RM)
– Match vertices across the graph at random
• Heavy Edge Matching (HEM)
– Traverse adjacency lists of unmatched vertices and match to the heaviest unmatched neighbor.
• Heaviest Edge Matching (SHEM)
– Sort edges in decreasing order by edge weight
– Perform valid matches in order
![Page 9: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/9.jpg)
Matching Strategies
• Heavy Triangle Matching (HTM) – Do a local 2-step BFS on unmatched vertices
– Match the heaviest valid triangle
– (Treat absence of edge as 0-weight edge)
• Heaviest Clique Matching (HCM) – Find and sort 3-cliques in descending order of sum
of edge weights
– Perform valid matches in order
• Many More!
![Page 10: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/10.jpg)
Matching Strategies
• Pitfalls
– Some matching strategies are expensive
• SHEM, HCM, HTM
– No compression guarantees
• High degree vertices, “stars” are notorious
• Villain:
![Page 11: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/11.jpg)
Matching Strategies
• “Brotherly” Matching
– Match topologically close vertices through a common neighbor
– Ex:
![Page 12: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/12.jpg)
Matching Strategies
• “Brotherly” Matching
– Match topologically close vertices through a common neighbor
– Ex:
![Page 13: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/13.jpg)
Matching Strategies
• “Brotherly” Matching
– Match topologically close vertices through a common neighbor
– Ex:
HEM
![Page 14: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/14.jpg)
Matching Strategies
• “Brotherly” Matching
– Match topologically close vertices through a common neighbor
– Ex:
HEM
“Brotherly”
![Page 15: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/15.jpg)
Matching Strategies
• “Adoption” Matching
– Toss another vertex into an existing 2-way match
– Ex:
HEM
“Brotherly”
![Page 16: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/16.jpg)
Matching Strategies
• “Adoption” Matching
– Toss another vertex into an existing 2-way match
– Ex:
HEM
“Brotherly”
“Adoption”
![Page 17: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/17.jpg)
Matching Strategies
• “Adoption” Matching
– Toss another vertex into an existing 2-way match
– Ex:
n=9 n=4
![Page 18: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/18.jpg)
Matching Strategies
• To ensure productive coarsening, we want a bound on the number of coarsening phases.
• Can we guarantee O(log n) coarsening phases?
![Page 19: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/19.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 20: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/20.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 21: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/21.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 22: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/22.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 23: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/23.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 24: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/24.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 25: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/25.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
![Page 26: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/26.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
n=8 n=4
![Page 27: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/27.jpg)
Matching Strategies
• “Community” Matching
– After an adoption, if a second adoption is required at the same match site, match the old adoptee with the new adoptee.
– Ex:
– Equivalent to a brotherly match at the next level.
n=8 n=4
![Page 28: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/28.jpg)
Initial Cut
• Random
• Clustering
• BFS-based
• Brute Force
cut
![Page 29: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/29.jpg)
Refinement
• Pass the partition choices up
• Swap vertex partitions
– Kernighan-Lin (‘70)
– Fiduccia-Mattheyses (‘82)
• Maintain balance refine
![Page 30: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/30.jpg)
Mongoose - Overview
• Nuri Yeralan, Tim Davis, William Hager
• Hybrid graph partitioner
– combinatorial & quadratic programming
• Single-threaded
• No randomization
• Handles floating point weights
• Interoperability with MATLAB, GraphViz, SuiteSparse, QPDelta
![Page 31: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/31.jpg)
Mongoose - Coarsening
• Matching
– “Jumpstart” HEM while building coarse graphs
– Pass over unmatched vertices
• Note: All neighbors of unmatched vertices are matched.
– Select heaviest matched neighbor
– Perform brotherly matching
– Adopt/Community match orphaned vertices
– Guarantees O(log n) coarsening levels
![Page 32: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/32.jpg)
Mongoose - Coarsening
• Matching
– In practice, we initiate brotherly matching if the vertex’s degree exceeds twice the average.
– Still productive (empirically) but no guarantee
![Page 33: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/33.jpg)
Mongoose – Initial Cut
• Single Pseudoperipheral
– Find a Pseudoperipheral Node (George ‘79)
– Collect vertices into one partition using a BFS from a pseudoperipheral node until the partition exceeds the desired size.
– Vertices sitting on the cut are considered “boundary” vertices and placed into a heap corresponding to their partition.
![Page 34: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/34.jpg)
Mongoose - Refinement
• Boundary Fiduccia-Mattheyes – One heap per partition stores boundary vertices.
– An entry’s value is the result of an objective function combining FM gains with a linear balance penalty • Incorporates user-defined balance ratio and tolerance
• Allows otherwise infeasible moves
– Inspect the top 3 elements of each heap and swap the vertex with the highest heuristic value.
– Karypis and Kumar introduced boundary KL in ‘97.
![Page 35: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/35.jpg)
Mongoose - Refinement
• Use optimization to solidify balance constraints and make moves that combinatorial methods miss.
• Quadratic Programming Formulation of the Graph Partitioning Problem (Hager ‘99)
– minimize: 1 − 𝑥 𝑇 𝐴 + 𝐷 𝑥
• D is a diagonal matrix such that 𝑑𝑖𝑖 + 𝑑𝑗𝑗 ≥ 𝑎𝑖𝑗
• Mongoose sets 𝑑𝑖𝑖 to the sum of 𝑖’s edge weights.
– subject to: 0 ≤ 𝑥 ≤ 1, 1𝑇𝑥 = 𝑚
• m: desired node weight for one partition
![Page 36: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/36.jpg)
Mongoose - Refinement
• QP Formulation: Gradient Projection
– Compute a gradient vector at each vertex
• 𝐺𝑟𝑎𝑑 = 0.5 – 𝑥𝑇 𝐷 + 0.5 − 𝑥𝑇 𝐴
– 𝑥𝑖 ∈ 0,1 represents vertex 𝑖’s partition choice
– Move along the gradient
– Project back onto feasible set
![Page 37: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/37.jpg)
Mongoose In Action
• Mongoose cut of 1138_bus from the UF Sparse Matrix Library (Davis ‘10)
– n = 1138
– nz = 4054
– Power network problem (bus power system) submitted by D. Tylavsky (‘87)
• GraphViz is used to render using a force-directed layout (AT&T Research, ‘88)
![Page 38: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/38.jpg)
![Page 39: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/39.jpg)
![Page 40: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/40.jpg)
![Page 41: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/41.jpg)
![Page 42: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/42.jpg)
![Page 43: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/43.jpg)
![Page 44: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/44.jpg)
![Page 45: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/45.jpg)
![Page 46: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/46.jpg)
![Page 47: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/47.jpg)
![Page 48: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/48.jpg)
![Page 49: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/49.jpg)
![Page 50: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/50.jpg)
Boundary FM never finds this
![Page 51: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/51.jpg)
![Page 52: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/52.jpg)
![Page 53: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/53.jpg)
![Page 54: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/54.jpg)
![Page 55: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/55.jpg)
![Page 56: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/56.jpg)
QP makes a subtle adjustment
![Page 57: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/57.jpg)
![Page 58: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/58.jpg)
![Page 59: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/59.jpg)
![Page 60: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/60.jpg)
QP makes a subtle adjustment
![Page 61: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/61.jpg)
Results: Set 1, Hodgepodge
• 1550 square matrices
• Between 15 and 2mil edges
• Come from a variety of contexts
– Optimization, fluid dynamics, quantum mechanics, PDE, circuit simulation, power line networks, 2d & 3d mesh, etc
![Page 62: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/62.jpg)
Profile: Timing 1550 General Problems: FM, QP, Hybrid
FM-Only
Hybrid
QP-Only
![Page 63: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/63.jpg)
Profile: Cut Quality 1550 General Problems: FM, QP, Hybrid
FM-Only
Hybrid
QP-Only
![Page 64: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/64.jpg)
Remarks
• The quadratic programming formulation seems to work well in a multi-level setting.
• Combining the combinatorial method with the quadratic programming formulation is superior to either by itself.
• How well does Mongoose perform against contemporary partitioners?
– METIS 5.0.2
![Page 65: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/65.jpg)
Profile: Timing 1550 General Problems: METIS 5, Mongoose
Mongoose
METIS 5.0.2
![Page 66: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/66.jpg)
Profile: Cut Quality 1550 General Problems: METIS 5, Mongoose
Mongoose
METIS 5.0.2
![Page 67: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/67.jpg)
Remarks
• In the general case, Mongoose seems to perform on par with METIS 5.0.2 in both performance and cut quality.
![Page 68: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/68.jpg)
Results: Set 2, Power Law Graphs
• 25 square matrices
• Between 1mil and 10 mil edges
• Come from social networking, web networks
![Page 69: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/69.jpg)
Profile: Timing 25 Power Law Problems: FM, QP, Hybrid
FM-Only
Hybrid
QP-Only
![Page 70: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/70.jpg)
Profile: Cut Quality 25 Power Law Problems: FM, QP, Hybrid
FM-Only
Hybrid
QP-Only
![Page 71: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/71.jpg)
Remarks
• Providing gradient projection with an good guess via boundary-FM reduces the overall cost of the hybrid approach
– Synergy – FM and QP mutually help each other
• Hybrid approach tends to find better cuts
• How well does Mongoose perform against contemporary partitioners?
– METIS 5.0.2
![Page 72: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/72.jpg)
Profile: Timing 25 Power Law Problems: METIS 5, Mongoose
Mongoose
METIS 5.0.2
![Page 73: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/73.jpg)
Profile: Cut Quality 25 Power Law Problems: METIS 5, Mongoose
Mongoose
METIS 5.0.2
![Page 74: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/74.jpg)
Remarks
• Mongoose tends to find better cuts than METIS 5.0.2 for large graphs.
• Mongoose ties METIS 5.0.2 for time.
– Future Work
• Optimize quadratic programming formulation
• Use randomization and parallelization
![Page 75: Multi-Level Edge Separator Using a Hybrid Combinatorial ... · function combining FM gains with a linear balance penalty •Incorporates user-defined balance ratio and tolerance •Allows](https://reader035.vdocument.in/reader035/viewer/2022070911/5fa6f89a18be5879c2644e79/html5/thumbnails/75.jpg)
Questions?
openclipart.org