220 likes | 366 Views
Custom Computing Machines for the Set Covering Problem. Paper Written By: Christian Plessl and Marco Platzner Swiss Federal Institute of Technology, 2002. Presentation By Greg Jackson. Overview. Introduction to the Problem Current Solutions Proposed Reconfigurable Architecture
E N D
Custom Computing Machines for the Set Covering Problem Paper Written By: Christian Plessl and Marco Platzner Swiss Federal Institute of Technology, 2002 Presentation By Greg Jackson
Overview • Introduction to the Problem • Current Solutions • Proposed Reconfigurable Architecture • Implementation • Experimentation • Results
The Problem • Boolean Satisfiability (SAT) • A discrete decision problem whereby variables are found such that a boolean expression in Conjunctive Normal Form(CNF) is evaluated true. • CNF? • Given Expression – (A and B) or C • CNF Equivalent – (A or C) and (A or B)
The Problem • The Set Covering Problem is a Minimum Cost SAT Problem • “Given a set U of elements and a set S of subsets ti of U, i = 1... |S|, find the smallest subset T, that contains all elements of U”
The Problem • Often in Expressed in Matrix Form • The Matrix A is a subset of columns with at least one entry for all the variables (A, B, C, D)
The Problem • The Set Covering problem is often associated with the Synthesis of Reconfigurable Designs • Logic Optimization that occurs during synthesis is a Set Covering Problem • e.g. Two-level logic minimization, state minimization
Current Solutions • Software - Mini & Espresso • Utilizes a technique called Branch and Bound • Heuristic Methods • Variable Order, Order of Value Assignments etc. • Branch and Bound • The software creates search trees and iterates over all values, trimming the trees as possible solutions are ruled out
Branch and Bound • A – Matrix Defining the Set Covering Problem • V – Current Variable Assignment • B – Current Low Cost Solution • Heuristics can be applied to Algorithm • Variable Order • Value Assignments • Algorithm can be Bounded by Cost
Why Did They Do It? • Faster • Speedups in the order of several magnitudes to solve more problems • Evaluation of Heuristics • Need fast solvers to evaluate the performance of heuristics and approximations • Development of New Heuristics • New methods = New Heuristics
Proposed Architecture • Basic Accelerator - Four Blocks • An Array of finite state machines (FSMs) • Checker Modules • Cost Counter • Controller • Architecture utilizes 3-valued Logic • {1,0,X}
Proposed Architecture • FSM • Implementation of the Branch and Bound in Hardware
Search Heuristics • Two Things Affecting Efficiency • Variable Order • Index Order – Variables are placed in order of their index • Frequency Order – Variables that appear frequently are placed at the top of the array • Assignment Order • 0 – 1 – FSM assigns a 0 first to the CNF • 1 – 0 – FSM assigns a 1 first to the CNF • Random
Cost Bound • Limit on the amount of assignments • This limits the length of the search
Hardware Reductions • 3 Ways • Don’s Cares • Essentials • Dominated Columns • This Creates a New State Machine
Implementation • Implemented 4 Architectures • CE – Basic • CEDC – Don’t Cares • CEDCES – Essentials • CEDCESDCOL – All 3 forms of Hardware reduction • Used Xilinx Virtex XCV1000 and simulated the system using ModelSim
Experiments • Used two sources for test problems • Two-level logic minimization • Steiner triple system • Selected 21 problems from Espresso distribution for testing • Test Problems Have 4-62 Variables and 4-70 clauses (i.e. A or B) • Would then compare the accelerator to the software implementation
Results – Raw Speedup • Compared the runtimes to that of the Software • Did not take into account delays i.e. Code generation, synthesis, etc. • Using the CEDCESDCOL accelerator, achieved orders of magnitude speed up • Average Speedup – 4 orders of magnitude over the Software
Results – Relative Speedup • Compared the architectures to each other in terms of speed up • For the graph, used best performance heuristics • The CEDCESDCOL architecture was the fastest
Results – Heuristic Performance • Used the four different architectures and determined which heuristics worked best for each • Example: CEDCES • Variable Order – Frequency Variable Order • Variable Assignment – Random
Conclusions • Speedups were promising but... • The tested problems were too small and can be solved in seconds in software (5ms – 1.3s) • The synthesis time of the designs were in the order of minutes • Reducing the hardware made the accelerators slower with larger problems when compared to the software. • Suggest a Hardware/Software Accelerator • Speed up of Hardware + Sophistication of Software
Final Thoughts • Extensive Paper • It was a Continuation of a Previous Paper • Useful application • Hardware Reductions not Explained Well