test/loops/TSPExchangerTest.java

/*
 * Written by Doug Lea and Bill Scherer with assistance from members
 * of JCP JSR-166 Expert Group and released to the public domain, as
 * explained at http://creativecommons.org/licenses/publicdomain
 */

import java.util.*;
import java.util.concurrent.*;
import java.util.concurrent.atomic.*;
import java.util.concurrent.locks.*;

/**
 * A parallel Traveling Salesperson Problem (TSP) program based on a
 * genetic algorithm using an Exchanger.  A population of chromosomes is
 * distributed among "subpops".  Each chromosomes represents a tour,
 * and its fitness is the total tour length.
 *
 * A set of worker threads perform updates on subpops. The basic
 * update step is:
 * <ol>
 *   <li> Select a breeder b from the subpop
 *   <li> Create a strand of its tour with a random starting point and length
 *   <li> Offer the strand to the exchanger, receiving a strand from
 *        another subpop
 *   <li> Combine b and the received strand using crossing function to
 *        create new chromosome c.
 *   <li> Replace a chromosome in the subpop with c.
 * </ol>
 *
 * This continues for a given number of generations per subpop.
 * Because there are normally more subpops than threads, each worker
 * thread performs small (randomly sized) run of updates for one
 * subpop and then selects another. A run continues until there is at
 * most one remaining thread performing updates.
 *
 * See below for more details.
 */
public class TSPExchangerTest {
    static final int NCPUS = Runtime.getRuntime().availableProcessors();

    /** Runs start with two threads, increasing by two through max */
    static final int DEFAULT_MAX_THREADS  = Math.max(4, NCPUS + NCPUS/2);

    /** The number of replication runs per thread value */
    static final int DEFAULT_REPLICATIONS = 3;

    /** If true, print statistics in SNAPSHOT_RATE intervals */
    static boolean verbose = true;
    static final long SNAPSHOT_RATE = 10000; // in milliseconds

    /**
     * The problem size. Each city is a random point. The goal is to
     * find a tour among them with smallest total Euclidean distance.
     */
    static final int DEFAULT_CITIES = 144;

    // Tuning parameters.

    /**
     * The number of chromosomes per subpop. Must be a power of two.
     *
     * Smaller values lead to faster iterations but poorer quality
     * results
     */
    static final int DEFAULT_SUBPOP_SIZE = 32;

    /**
     * The number of iterations per subpop. Convergence appears
     * to be roughly proportional to #cities-squared
     */
    static final int DEFAULT_GENERATIONS = DEFAULT_CITIES * DEFAULT_CITIES;

    /**
     * The number of subpops. The total population is #subpops * subpopSize,
     * which should be roughly on the order of #cities-squared
     *
     * Smaller values lead to faster total runs but poorer quality
     * results
     */
    static final int DEFAULT_NSUBPOPS = DEFAULT_GENERATIONS / DEFAULT_SUBPOP_SIZE;

    /**
     * The minimum length for a random chromosome strand.
     * Must be at least 1.
     */
    static final int MIN_STRAND_LENGTH = 3;

    /**
     * The probability mask value for creating random strands,
     * that have lengths at least MIN_STRAND_LENGTH, and grow
     * with exposnential decay 2^(-(1/(RANDOM_STRAND_MASK + 1)
     * Must be 1 less than a power of two.
     */
    static final int RANDOM_STRAND_MASK = 7;

    /**
     * Probability control for selecting breeders.
     * Breeders are selected starting at the best-fitness chromosome,
     * with exponentially decaying probability
     * 1 / (subpopSize >>> BREEDER_DECAY).
     *
     * Larger values usually cause faster convergence but poorer
     * quality results
     */
    static final int BREEDER_DECAY = 1;

    /**
     * Probability control for selecting dyers.
     * Dyers are selected starting at the worst-fitness chromosome,
     * with exponentially decaying probability
     * 1 / (subpopSize >>> DYER_DECAY)
     *
     * Larger values usually cause faster convergence but poorer
     * quality results
     */
    static final int DYER_DECAY = 1;

    /**
     * The set of cities. Created once per program run, to
     * make it easier to compare solutions across different runs.
     */
    static CitySet cities;

    public static void main(String[] args) throws Exception {
        int maxThreads = DEFAULT_MAX_THREADS;
        int nCities = DEFAULT_CITIES;
        int subpopSize = DEFAULT_SUBPOP_SIZE;
        int nGen = nCities * nCities;
        int nSubpops = nCities * nCities / subpopSize;
        int nReps = DEFAULT_REPLICATIONS;

        try {
            int argc = 0;
            while (argc < args.length) {
                String option = args[argc++];
                if (option.equals("-c")) {
                    nCities = Integer.parseInt(args[argc]);
                    nGen = nCities * nCities;
                    nSubpops = nCities * nCities / subpopSize;
                }
                else if (option.equals("-p"))
                    subpopSize = Integer.parseInt(args[argc]);
                else if (option.equals("-g"))
                    nGen = Integer.parseInt(args[argc]);
                else if (option.equals("-n"))
                    nSubpops = Integer.parseInt(args[argc]);
                else if (option.equals("-q")) {
                    verbose = false;
                    argc--;
                }
                else if (option.equals("-r"))
                    nReps = Integer.parseInt(args[argc]);
                else
                    maxThreads = Integer.parseInt(option);
                argc++;
            }
        }
        catch (Exception e) {
            reportUsageErrorAndDie();
        }

        System.out.print("TSPExchangerTest");
        System.out.print(" -c " + nCities);
        System.out.print(" -g " + nGen);
        System.out.print(" -p " + subpopSize);
        System.out.print(" -n " + nSubpops);
        System.out.print(" -r " + nReps);
        System.out.print(" max threads " + maxThreads);
        System.out.println();

        cities = new CitySet(nCities);

        if (false && NCPUS > 4) {
            int h = NCPUS/2;
            System.out.printf("Threads: %4d Warmup\n", h);
            oneRun(h, nSubpops, subpopSize, nGen);
            Thread.sleep(500);
        }

        int maxt = (maxThreads < nSubpops) ? maxThreads : nSubpops;
        for (int j = 0; j < nReps; ++j) {
            for (int i = 2; i <= maxt; i += 2) {
                System.out.printf("Threads: %4d Replication: %2d\n", i, j);
                oneRun(i, nSubpops, subpopSize, nGen);
                Thread.sleep(500);
            }
        }
    }

    static void reportUsageErrorAndDie() {
        System.out.print("usage: TSPExchangerTest");
        System.out.print(" [-c #cities]");
        System.out.print(" [-p #subpopSize]");
        System.out.print(" [-g #generations]");
        System.out.print(" [-n #subpops]");
        System.out.print(" [-r #replications]");
        System.out.print(" [-q <quiet>]");
        System.out.print(" #threads]");
        System.out.println();
        System.exit(0);
    }

    /**
     * Perform one run with the given parameters.  Each run complete
     * when there are fewer than 2 active threads.  When there is
     * only one remaining thread, it will have no one to exchange
     * with, so it is terminated (via interrupt).
     */
    static void oneRun(int nThreads, int nSubpops, int subpopSize, int nGen)
        throws InterruptedException {
        Population p = new Population(nThreads, nSubpops, subpopSize, nGen);
        ProgressMonitor mon = null;
        if (verbose) {
            p.printSnapshot(0);
            mon = new ProgressMonitor(p);
            mon.start();
        }
        long startTime = System.nanoTime();
        p.start();
        p.awaitDone();
        long stopTime = System.nanoTime();
        if (mon != null)
            mon.interrupt();
        p.shutdown();
        //        Thread.sleep(100);

        long elapsed = stopTime - startTime;
        double secs = (double)elapsed / 1000000000.0;
        p.printSnapshot(secs);
    }


    /**
     * A Population creates the subpops, subpops, and threads for a run
     * and has control methods to start, stop, and report progress.
     */
    static final class Population {
        final Worker[] threads;
        final Subpop[] subpops;
        final Exchanger<Strand> exchanger;
        final CountDownLatch done;
        final int nGen;
        final int subpopSize;
        final int nThreads;

        Population(int nThreads, int nSubpops, int subpopSize, int nGen) {
            this.nThreads = nThreads;
            this.nGen = nGen;
            this.subpopSize = subpopSize;
            this.exchanger = new Exchanger<Strand>();
            this.done = new CountDownLatch(nThreads - 1);

            this.subpops = new Subpop[nSubpops];
            for (int i = 0; i < nSubpops; i++)
                subpops[i] = new Subpop(this);

            this.threads = new Worker[nThreads];
            int maxExchanges = nGen * nSubpops / nThreads;
            for (int i = 0; i < nThreads; ++i) {
                threads[i] = new Worker(this, maxExchanges);
            }

        }

        void start() {
            for (int i = 0; i < nThreads; ++i) {
                threads[i].start();
            }
        }

        /** Stop the tasks */
        void shutdown() {
            for (int i = 0; i < threads.length; ++ i)
                threads[i].interrupt();
        }

        void threadDone() {
            done.countDown();
        }

        /** Wait for tasks to complete */
        void awaitDone() throws InterruptedException {
            done.await();
        }

        int totalExchanges() {
            int xs = 0;
            for (int i = 0; i < threads.length; ++i)
                xs += threads[i].exchanges;
            return xs;
        }

        /**
         * Prints statistics, including best and worst tour lengths
         * for points scaled in [0,1), scaled by the square root of
         * number of points. This simplifies checking results.  The
         * expected optimal TSP for random points is believed to be
         * around 0.76 * sqrt(N). For papers discussing this, see
         * http://www.densis.fee.unicamp.br/~moscato/TSPBIB_home.html
         */
        void printSnapshot(double secs) {
            int xs = totalExchanges();
            long rate = (xs == 0)? 0L : (long)((secs * 1000000000.0) / xs);
            Chromosome bestc = subpops[0].chromosomes[0];
            Chromosome worstc = bestc;
            for (int k = 0; k < subpops.length; ++k) {
                Chromosome[] cs = subpops[k].chromosomes;
                if (cs[0].fitness < bestc.fitness)
                    bestc = cs[0];
                int w = cs[cs.length-1].fitness;
                if (cs[cs.length-1].fitness > worstc.fitness)
                    worstc = cs[cs.length-1];
            }
            double sqrtn = Math.sqrt(cities.length);
            double best = bestc.unitTourLength() / sqrtn;
            double worst = worstc.unitTourLength() / sqrtn;
            System.out.printf("N:%4d T:%8.3f B:%6.3f W:%6.3f X:%9d R:%7d\n",
                              nThreads, secs, best, worst, xs, rate);
            //            exchanger.printStats();
            //            System.out.print(" s: " + exchanger.aveSpins());
            //            System.out.print(" p: " + exchanger.aveParks());
        }
    }

    /**
     * Worker threads perform updates on subpops.
     */
    static final class Worker extends Thread {
        final Population pop;
        final int maxExchanges;
        int exchanges;
        final RNG rng = new RNG();

        Worker(Population pop, int maxExchanges) {
            this.pop = pop;
            this.maxExchanges = maxExchanges;
        }

        /**
         * Repeatedly, find a subpop that is not being updated by
         * another thread, and run a random number of updates on it.
         */
        public void run() {
            try {
                int len = pop.subpops.length;
                int pos = (rng.next() & 0x7FFFFFFF) % len;
                while (exchanges < maxExchanges) {
                    Subpop s = pop.subpops[pos];
                    AtomicBoolean busy = s.busy;
                    if (!busy.get() && busy.compareAndSet(false, true)) {
                        exchanges += s.runUpdates();
                        busy.set(false);
                        pos = (rng.next() & 0x7FFFFFFF) % len;
                    }
                    else if (++pos >= len)
                        pos = 0;
                }
                pop.threadDone();
            } catch (InterruptedException fallthrough) {
            }
        }
    }

    /**
     * A Subpop maintains a set of chromosomes..
     */
    static final class Subpop {
        /** The chromosomes, kept in sorted order */
        final Chromosome[] chromosomes;
        /** The parent population */
        final Population pop;
        /** Reservation bit for worker threads */
        final AtomicBoolean busy;
        /** The common exchanger, same for all subpops */
        final Exchanger<Strand> exchanger;
        /** The current strand being exchanged */
        Strand strand;
        /** Bitset used in cross */
        final int[] inTour;
        final RNG rng;
        final int subpopSize;

        Subpop(Population pop) {
            this.pop = pop;
            this.subpopSize = pop.subpopSize;
            this.exchanger = pop.exchanger;
            this.busy = new AtomicBoolean(false);
            this.rng = new RNG();
            int length = cities.length;
            this.strand = new Strand(length);
            this.inTour = new int[(length >>> 5) + 1];
            this.chromosomes = new Chromosome[subpopSize];
            for (int j = 0; j < subpopSize; ++j)
                chromosomes[j] = new Chromosome(length, rng);
            Arrays.sort(chromosomes);
        }

        /**
         * Run a random number of updates.  The number of updates is
         * at least 1 and no more than subpopSize.  This
         * controls the granularity of multiplexing subpop updates on
         * to threads. It is small enough to balance out updates
         * across tasks, but large enough to avoid having runs
         * dominated by subpop selection. It is randomized to avoid
         * long runs where pairs of subpops exchange only with each
         * other.  It is hardwired because small variations of it
         * don't matter much.
         *
         * @param g the first generation to run.
         */
        int runUpdates() throws InterruptedException {
            int n = 1 + (rng.next() & ((subpopSize << 1) - 1));
            for (int i = 0; i < n; ++i)
                update();
            return n;
        }

        /**
         * Choose a breeder, exchange strand with another subpop, and
         * cross them to create new chromosome to replace a chosen
         * dyer.
         */
        void update() throws InterruptedException {
            int b = chooseBreeder();
            int d = chooseDyer(b);
            Chromosome breeder = chromosomes[b];
            Chromosome child = chromosomes[d];
            chooseStrand(breeder);
            strand = exchanger.exchange(strand);
            cross(breeder, child);
            fixOrder(child, d);
        }

        /**
         * Choose a breeder, with exponentially decreasing probability
         * starting at best.
         * @return index of selected breeder
         */
        int chooseBreeder() {
            int mask = (subpopSize >>> BREEDER_DECAY) - 1;
            int b = 0;
            while ((rng.next() & mask) != mask) {
                if (++b >= subpopSize)
                    b = 0;
            }
            return b;
        }

        /**
         * Choose a chromosome that will be replaced, with
         * exponentially decreasing probability starting at
         * worst, ignoring the excluded index
         * @param exclude index to ignore; use -1 to not exclude any
         * @return index of selected dyer
         */
        int chooseDyer(int exclude) {
            int mask = (subpopSize >>> DYER_DECAY)  - 1;
            int d = subpopSize - 1;
            while (d == exclude || (rng.next() & mask) != mask) {
                if (--d < 0)
                    d = subpopSize - 1;
            }
            return d;
        }

        /**
         * Select a random strand of b's.
         * @param breeder the breeder
         */
        void chooseStrand(Chromosome breeder) {
            int[] bs = breeder.alleles;
            int length = bs.length;
            int strandLength = MIN_STRAND_LENGTH;
            while (strandLength < length &&
                   (rng.next() & RANDOM_STRAND_MASK) != RANDOM_STRAND_MASK)
                strandLength++;
            strand.strandLength = strandLength;
            int[] ss = strand.alleles;
            int k = (rng.next() & 0x7FFFFFFF) % length;
            for (int i = 0; i < strandLength; ++i) {
                ss[i] = bs[k];
                if (++k >= length) k = 0;
            }
        }

        /**
         * Copy current strand to start of c's, and then append all
         * remaining b's that aren't in the strand.
         * @param breeder the breeder
         * @param child the child
         */
        void cross(Chromosome breeder, Chromosome child) {
            for (int k = 0; k < inTour.length; ++k) // clear bitset
                inTour[k] = 0;

            // Copy current strand to c
            int[] cs = child.alleles;
            int ssize = strand.strandLength;
            int[] ss = strand.alleles;
            int i;
            for (i = 0; i < ssize; ++i) {
                int x = ss[i];
                cs[i] = x;
                inTour[x >>> 5] |= 1 << (x & 31); // record in bit set
            }

            // Find index of matching origin in b
            int first = cs[0];
            int j = 0;
            int[] bs = breeder.alleles;
            while (bs[j] != first)
                ++j;

            // Append remaining b's that aren't already in tour
            while (i < cs.length) {
                if (++j >= bs.length) j = 0;
                int x = bs[j];
                if ((inTour[x >>> 5] & (1 << (x & 31))) == 0)
                    cs[i++] = x;
            }

        }

        /**
         * Fix the sort order of a changed Chromosome c at position k
         * @param c the chromosome
         * @param k the index
         */
        void fixOrder(Chromosome c, int k) {
            Chromosome[] cs = chromosomes;
            int oldFitness = c.fitness;
            c.recalcFitness();
            int newFitness = c.fitness;
            if (newFitness < oldFitness) {
                int j = k;
                int p = j - 1;
                while (p >= 0 && cs[p].fitness > newFitness) {
                    cs[j] = cs[p];
                    j = p--;
                }
                cs[j] = c;
            } else if (newFitness > oldFitness) {
                int j = k;
                int n = j + 1;
                while (n < cs.length && cs[n].fitness < newFitness) {
                    cs[j] = cs[n];
                    j = n++;
                }
                cs[j] = c;
            }
        }
    }

    /**
     * A Chromosome is a candidate TSP tour.
     */
    static final class Chromosome implements Comparable {
        /** Index of cities in tour order */
        final int[] alleles;
        /** Total tour length */
        int fitness;

        /**
         * Initialize to random tour
         */
        Chromosome(int length, RNG random) {
            alleles = new int[length];
            for (int i = 0; i < length; i++)
                alleles[i] = i;
            for (int i = length - 1; i > 0; i--) {
                int idx = (random.next() & 0x7FFFFFFF) % alleles.length;
                int tmp = alleles[i];
                alleles[i] = alleles[idx];
                alleles[idx] = tmp;
            }
            recalcFitness();
        }

        public int compareTo(Object x) { // to enable sorting
            int xf = ((Chromosome)x).fitness;
            int f = fitness;
            return ((f == xf)? 0 :((f < xf)? -1 : 1));
        }

        void recalcFitness() {
            int[] a = alleles;
            int len = a.length;
            int p = a[0];
            long f = cities.distanceBetween(a[len-1], p);
            for (int i = 1; i < len; i++) {
                int n = a[i];
                f += cities.distanceBetween(p, n);
                p = n;
            }
            fitness = (int)(f / len);
        }

        /**
         * Return tour length for points scaled in [0, 1).
         */
        double unitTourLength() {
            int[] a = alleles;
            int len = a.length;
            int p = a[0];
            double f = cities.unitDistanceBetween(a[len-1], p);
            for (int i = 1; i < len; i++) {
                int n = a[i];
                f += cities.unitDistanceBetween(p, n);
                p = n;
            }
            return f;
        }

        /**
         * Check that this tour visits each city
         */
        void validate() {
            int len = alleles.length;
            boolean[] used = new boolean[len];
            for (int i = 0; i < len; ++i)
                used[alleles[i]] = true;
            for (int i = 0; i < len; ++i)
                if (!used[i])
                    throw new Error("Bad tour");
        }

    }

    /**
     * A Strand is a random sub-sequence of a Chromosome.  Each subpop
     * creates only one strand, and then trades it with others,
     * refilling it on each iteration.
     */
    static final class Strand {
        final int[] alleles;
        int strandLength;
        Strand(int length) { alleles = new int[length]; }
    }

    /**
     * A collection of (x,y) points that represent cities.
     */
    static final class CitySet {

        final int length;
        final int[] xPts;
        final int[] yPts;
        final int[][] distances;

        CitySet(int n) {
            this.length = n;
            this.xPts = new int[n];
            this.yPts = new int[n];
            this.distances = new int[n][n];

            RNG random = new RNG();
            for (int i = 0; i < n; i++) {
                xPts[i] = (random.next() & 0x7FFFFFFF);
                yPts[i] = (random.next() & 0x7FFFFFFF);
            }

            for (int i = 0; i < n; i++) {
                for (int j = 0; j < n; j++) {
                    double dx = (double)xPts[i] - (double)xPts[j];
                    double dy = (double)yPts[i] - (double)yPts[j];
                    double dd = Math.hypot(dx, dy) / 2.0;
                    long ld = Math.round(dd);
                    distances[i][j] = (ld >= Integer.MAX_VALUE)?
                        Integer.MAX_VALUE : (int)ld;
                }
            }
        }

        /**
         *  Returns the cached distance between a pair of cities
         */
        int distanceBetween(int i, int j) {
            return distances[i][j];
        }

        // Scale ints to doubles in [0,1)
        static final double PSCALE = (double)0x80000000L;

        /**
         * Return distance for points scaled in [0,1). This simplifies
         * checking results.  The expected optimal TSP for random
         * points is believed to be around 0.76 * sqrt(N). For papers
         * discussing this, see
         * http://www.densis.fee.unicamp.br/~moscato/TSPBIB_home.html
         */
        double unitDistanceBetween(int i, int j) {
            double dx = ((double)xPts[i] - (double)xPts[j]) / PSCALE;
            double dy = ((double)yPts[i] - (double)yPts[j]) / PSCALE;
            return Math.hypot(dx, dy);
        }

    }

    /**
     * Cheap XorShift random number generator
     */
    static final class RNG {
        /** Seed generator for XorShift RNGs */
        static final Random seedGenerator = new Random();

        int seed;
        RNG(int seed) { this.seed = seed; }
        RNG()         { this.seed = seedGenerator.nextInt() | 1;  }

        int next() {
            int x = seed;
            x ^= x << 6;
            x ^= x >>> 21;
            x ^= x << 7;
            seed = x;
            return x;
        }
    }

    static final class ProgressMonitor extends Thread {
        final Population pop;
        ProgressMonitor(Population p) { pop = p; }
        public void run() {
            double time = 0;
            try {
                while (!Thread.interrupted()) {
                    sleep(SNAPSHOT_RATE);
                    time += SNAPSHOT_RATE;
                    pop.printSnapshot(time / 1000.0);
                }
            } catch (InterruptedException ie) {}
        }
    }
}
Revision:	1.6
Committed:	Thu Oct 29 23:09:08 2009 UTC (14 years, 6 months ago) by jsr166
Branch:	MAIN
Changes since 1.5:	+23 -23 lines
Log Message:	whitespace
#	Content
1	/*
2	* Written by Doug Lea and Bill Scherer with assistance from members
3	* of JCP JSR-166 Expert Group and released to the public domain, as
4	* explained at http://creativecommons.org/licenses/publicdomain
5	*/
6
7	import java.util.*;
8	import java.util.concurrent.*;
9	import java.util.concurrent.atomic.*;
10	import java.util.concurrent.locks.*;
11
12	/**
13	* A parallel Traveling Salesperson Problem (TSP) program based on a
14	* genetic algorithm using an Exchanger. A population of chromosomes is
15	* distributed among "subpops". Each chromosomes represents a tour,
16	* and its fitness is the total tour length.
17	*
18	* A set of worker threads perform updates on subpops. The basic
19	* update step is:
20	* <ol>
21	* <li> Select a breeder b from the subpop
22	* <li> Create a strand of its tour with a random starting point and length
23	* <li> Offer the strand to the exchanger, receiving a strand from
24	* another subpop
25	* <li> Combine b and the received strand using crossing function to
26	* create new chromosome c.
27	* <li> Replace a chromosome in the subpop with c.
28	* </ol>
29	*
30	* This continues for a given number of generations per subpop.
31	* Because there are normally more subpops than threads, each worker
32	* thread performs small (randomly sized) run of updates for one
33	* subpop and then selects another. A run continues until there is at
34	* most one remaining thread performing updates.
35	*
36	* See below for more details.
37	*/
38	public class TSPExchangerTest {
39	static final int NCPUS = Runtime.getRuntime().availableProcessors();
40
41	/** Runs start with two threads, increasing by two through max */
42	static final int DEFAULT_MAX_THREADS = Math.max(4, NCPUS + NCPUS/2);
43
44	/** The number of replication runs per thread value */
45	static final int DEFAULT_REPLICATIONS = 3;
46
47	/** If true, print statistics in SNAPSHOT_RATE intervals */
48	static boolean verbose = true;
49	static final long SNAPSHOT_RATE = 10000; // in milliseconds
50
51	/**
52	* The problem size. Each city is a random point. The goal is to
53	* find a tour among them with smallest total Euclidean distance.
54	*/
55	static final int DEFAULT_CITIES = 144;
56
57	// Tuning parameters.
58
59	/**
60	* The number of chromosomes per subpop. Must be a power of two.
61	*
62	* Smaller values lead to faster iterations but poorer quality
63	* results
64	*/
65	static final int DEFAULT_SUBPOP_SIZE = 32;
66
67	/**
68	* The number of iterations per subpop. Convergence appears
69	* to be roughly proportional to #cities-squared
70	*/
71	static final int DEFAULT_GENERATIONS = DEFAULT_CITIES * DEFAULT_CITIES;
72
73	/**
74	* The number of subpops. The total population is #subpops * subpopSize,
75	* which should be roughly on the order of #cities-squared
76	*
77	* Smaller values lead to faster total runs but poorer quality
78	* results
79	*/
80	static final int DEFAULT_NSUBPOPS = DEFAULT_GENERATIONS / DEFAULT_SUBPOP_SIZE;
81
82	/**
83	* The minimum length for a random chromosome strand.
84	* Must be at least 1.
85	*/
86	static final int MIN_STRAND_LENGTH = 3;
87
88	/**
89	* The probability mask value for creating random strands,
90	* that have lengths at least MIN_STRAND_LENGTH, and grow
91	* with exposnential decay 2^(-(1/(RANDOM_STRAND_MASK + 1)
92	* Must be 1 less than a power of two.
93	*/
94	static final int RANDOM_STRAND_MASK = 7;
95
96	/**
97	* Probability control for selecting breeders.
98	* Breeders are selected starting at the best-fitness chromosome,
99	* with exponentially decaying probability
100	* 1 / (subpopSize >>> BREEDER_DECAY).
101	*
102	* Larger values usually cause faster convergence but poorer
103	* quality results
104	*/
105	static final int BREEDER_DECAY = 1;
106
107	/**
108	* Probability control for selecting dyers.
109	* Dyers are selected starting at the worst-fitness chromosome,
110	* with exponentially decaying probability
111	* 1 / (subpopSize >>> DYER_DECAY)
112	*
113	* Larger values usually cause faster convergence but poorer
114	* quality results
115	*/
116	static final int DYER_DECAY = 1;
117
118	/**
119	* The set of cities. Created once per program run, to
120	* make it easier to compare solutions across different runs.
121	*/
122	static CitySet cities;
123
124	public static void main(String[] args) throws Exception {
125	int maxThreads = DEFAULT_MAX_THREADS;
126	int nCities = DEFAULT_CITIES;
127	int subpopSize = DEFAULT_SUBPOP_SIZE;
128	int nGen = nCities * nCities;
129	int nSubpops = nCities * nCities / subpopSize;
130	int nReps = DEFAULT_REPLICATIONS;
131
132	try {
133	int argc = 0;
134	while (argc < args.length) {
135	String option = args[argc++];
136	if (option.equals("-c")) {
137	nCities = Integer.parseInt(args[argc]);
138	nGen = nCities * nCities;
139	nSubpops = nCities * nCities / subpopSize;
140	}
141	else if (option.equals("-p"))
142	subpopSize = Integer.parseInt(args[argc]);
143	else if (option.equals("-g"))
144	nGen = Integer.parseInt(args[argc]);
145	else if (option.equals("-n"))
146	nSubpops = Integer.parseInt(args[argc]);
147	else if (option.equals("-q")) {
148	verbose = false;
149	argc--;
150	}
151	else if (option.equals("-r"))
152	nReps = Integer.parseInt(args[argc]);
153	else
154	maxThreads = Integer.parseInt(option);
155	argc++;
156	}
157	}
158	catch (Exception e) {
159	reportUsageErrorAndDie();
160	}
161
162	System.out.print("TSPExchangerTest");
163	System.out.print(" -c " + nCities);
164	System.out.print(" -g " + nGen);
165	System.out.print(" -p " + subpopSize);
166	System.out.print(" -n " + nSubpops);
167	System.out.print(" -r " + nReps);
168	System.out.print(" max threads " + maxThreads);
169	System.out.println();
170
171	cities = new CitySet(nCities);
172
173	if (false && NCPUS > 4) {
174	int h = NCPUS/2;
175	System.out.printf("Threads: %4d Warmup\n", h);
176	oneRun(h, nSubpops, subpopSize, nGen);
177	Thread.sleep(500);
178	}
179
180	int maxt = (maxThreads < nSubpops) ? maxThreads : nSubpops;
181	for (int j = 0; j < nReps; ++j) {
182	for (int i = 2; i <= maxt; i += 2) {
183	System.out.printf("Threads: %4d Replication: %2d\n", i, j);
184	oneRun(i, nSubpops, subpopSize, nGen);
185	Thread.sleep(500);
186	}
187	}
188	}
189
190	static void reportUsageErrorAndDie() {
191	System.out.print("usage: TSPExchangerTest");
192	System.out.print(" [-c #cities]");
193	System.out.print(" [-p #subpopSize]");
194	System.out.print(" [-g #generations]");
195	System.out.print(" [-n #subpops]");
196	System.out.print(" [-r #replications]");
197	System.out.print(" [-q <quiet>]");
198	System.out.print(" #threads]");
199	System.out.println();
200	System.exit(0);
201	}
202
203	/**
204	* Perform one run with the given parameters. Each run complete
205	* when there are fewer than 2 active threads. When there is
206	* only one remaining thread, it will have no one to exchange
207	* with, so it is terminated (via interrupt).
208	*/
209	static void oneRun(int nThreads, int nSubpops, int subpopSize, int nGen)
210	throws InterruptedException {
211	Population p = new Population(nThreads, nSubpops, subpopSize, nGen);
212	ProgressMonitor mon = null;
213	if (verbose) {
214	p.printSnapshot(0);
215	mon = new ProgressMonitor(p);
216	mon.start();
217	}
218	long startTime = System.nanoTime();
219	p.start();
220	p.awaitDone();
221	long stopTime = System.nanoTime();
222	if (mon != null)
223	mon.interrupt();
224	p.shutdown();
225	// Thread.sleep(100);
226
227	long elapsed = stopTime - startTime;
228	double secs = (double)elapsed / 1000000000.0;
229	p.printSnapshot(secs);
230	}
231
232
233	/**
234	* A Population creates the subpops, subpops, and threads for a run
235	* and has control methods to start, stop, and report progress.
236	*/
237	static final class Population {
238	final Worker[] threads;
239	final Subpop[] subpops;
240	final Exchanger<Strand> exchanger;
241	final CountDownLatch done;
242	final int nGen;
243	final int subpopSize;
244	final int nThreads;
245
246	Population(int nThreads, int nSubpops, int subpopSize, int nGen) {
247	this.nThreads = nThreads;
248	this.nGen = nGen;
249	this.subpopSize = subpopSize;
250	this.exchanger = new Exchanger<Strand>();
251	this.done = new CountDownLatch(nThreads - 1);
252
253	this.subpops = new Subpop[nSubpops];
254	for (int i = 0; i < nSubpops; i++)
255	subpops[i] = new Subpop(this);
256
257	this.threads = new Worker[nThreads];
258	int maxExchanges = nGen * nSubpops / nThreads;
259	for (int i = 0; i < nThreads; ++i) {
260	threads[i] = new Worker(this, maxExchanges);
261	}
262
263	}
264
265	void start() {
266	for (int i = 0; i < nThreads; ++i) {
267	threads[i].start();
268	}
269	}
270
271	/** Stop the tasks */
272	void shutdown() {
273	for (int i = 0; i < threads.length; ++ i)
274	threads[i].interrupt();
275	}
276
277	void threadDone() {
278	done.countDown();
279	}
280
281	/** Wait for tasks to complete */
282	void awaitDone() throws InterruptedException {
283	done.await();
284	}
285
286	int totalExchanges() {
287	int xs = 0;
288	for (int i = 0; i < threads.length; ++i)
289	xs += threads[i].exchanges;
290	return xs;
291	}
292
293	/**
294	* Prints statistics, including best and worst tour lengths
295	* for points scaled in [0,1), scaled by the square root of
296	* number of points. This simplifies checking results. The
297	* expected optimal TSP for random points is believed to be
298	* around 0.76 * sqrt(N). For papers discussing this, see
299	* http://www.densis.fee.unicamp.br/~moscato/TSPBIB_home.html
300	*/
301	void printSnapshot(double secs) {
302	int xs = totalExchanges();
303	long rate = (xs == 0)? 0L : (long)((secs * 1000000000.0) / xs);
304	Chromosome bestc = subpops[0].chromosomes[0];
305	Chromosome worstc = bestc;
306	for (int k = 0; k < subpops.length; ++k) {
307	Chromosome[] cs = subpops[k].chromosomes;
308	if (cs[0].fitness < bestc.fitness)
309	bestc = cs[0];
310	int w = cs[cs.length-1].fitness;
311	if (cs[cs.length-1].fitness > worstc.fitness)
312	worstc = cs[cs.length-1];
313	}
314	double sqrtn = Math.sqrt(cities.length);
315	double best = bestc.unitTourLength() / sqrtn;
316	double worst = worstc.unitTourLength() / sqrtn;
317	System.out.printf("N:%4d T:%8.3f B:%6.3f W:%6.3f X:%9d R:%7d\n",
318	nThreads, secs, best, worst, xs, rate);
319	// exchanger.printStats();
320	// System.out.print(" s: " + exchanger.aveSpins());
321	// System.out.print(" p: " + exchanger.aveParks());
322	}
323	}
324
325	/**
326	* Worker threads perform updates on subpops.
327	*/
328	static final class Worker extends Thread {
329	final Population pop;
330	final int maxExchanges;
331	int exchanges;
332	final RNG rng = new RNG();
333
334	Worker(Population pop, int maxExchanges) {
335	this.pop = pop;
336	this.maxExchanges = maxExchanges;
337	}
338
339	/**
340	* Repeatedly, find a subpop that is not being updated by
341	* another thread, and run a random number of updates on it.
342	*/
343	public void run() {
344	try {
345	int len = pop.subpops.length;
346	int pos = (rng.next() & 0x7FFFFFFF) % len;
347	while (exchanges < maxExchanges) {
348	Subpop s = pop.subpops[pos];
349	AtomicBoolean busy = s.busy;
350	if (!busy.get() && busy.compareAndSet(false, true)) {
351	exchanges += s.runUpdates();
352	busy.set(false);
353	pos = (rng.next() & 0x7FFFFFFF) % len;
354	}
355	else if (++pos >= len)
356	pos = 0;
357	}
358	pop.threadDone();
359	} catch (InterruptedException fallthrough) {
360	}
361	}
362	}
363
364	/**
365	* A Subpop maintains a set of chromosomes..
366	*/
367	static final class Subpop {
368	/** The chromosomes, kept in sorted order */
369	final Chromosome[] chromosomes;
370	/** The parent population */
371	final Population pop;
372	/** Reservation bit for worker threads */
373	final AtomicBoolean busy;
374	/** The common exchanger, same for all subpops */
375	final Exchanger<Strand> exchanger;
376	/** The current strand being exchanged */
377	Strand strand;
378	/** Bitset used in cross */
379	final int[] inTour;
380	final RNG rng;
381	final int subpopSize;
382
383	Subpop(Population pop) {
384	this.pop = pop;
385	this.subpopSize = pop.subpopSize;
386	this.exchanger = pop.exchanger;
387	this.busy = new AtomicBoolean(false);
388	this.rng = new RNG();
389	int length = cities.length;
390	this.strand = new Strand(length);
391	this.inTour = new int[(length >>> 5) + 1];
392	this.chromosomes = new Chromosome[subpopSize];
393	for (int j = 0; j < subpopSize; ++j)
394	chromosomes[j] = new Chromosome(length, rng);
395	Arrays.sort(chromosomes);
396	}
397
398	/**
399	* Run a random number of updates. The number of updates is
400	* at least 1 and no more than subpopSize. This
401	* controls the granularity of multiplexing subpop updates on
402	* to threads. It is small enough to balance out updates
403	* across tasks, but large enough to avoid having runs
404	* dominated by subpop selection. It is randomized to avoid
405	* long runs where pairs of subpops exchange only with each
406	* other. It is hardwired because small variations of it
407	* don't matter much.
408	*
409	* @param g the first generation to run.
410	*/
411	int runUpdates() throws InterruptedException {
412	int n = 1 + (rng.next() & ((subpopSize << 1) - 1));
413	for (int i = 0; i < n; ++i)
414	update();
415	return n;
416	}
417
418	/**
419	* Choose a breeder, exchange strand with another subpop, and
420	* cross them to create new chromosome to replace a chosen
421	* dyer.
422	*/
423	void update() throws InterruptedException {
424	int b = chooseBreeder();
425	int d = chooseDyer(b);
426	Chromosome breeder = chromosomes[b];
427	Chromosome child = chromosomes[d];
428	chooseStrand(breeder);
429	strand = exchanger.exchange(strand);
430	cross(breeder, child);
431	fixOrder(child, d);
432	}
433
434	/**
435	* Choose a breeder, with exponentially decreasing probability
436	* starting at best.
437	* @return index of selected breeder
438	*/
439	int chooseBreeder() {
440	int mask = (subpopSize >>> BREEDER_DECAY) - 1;
441	int b = 0;
442	while ((rng.next() & mask) != mask) {
443	if (++b >= subpopSize)
444	b = 0;
445	}
446	return b;
447	}
448
449	/**
450	* Choose a chromosome that will be replaced, with
451	* exponentially decreasing probability starting at
452	* worst, ignoring the excluded index
453	* @param exclude index to ignore; use -1 to not exclude any
454	* @return index of selected dyer
455	*/
456	int chooseDyer(int exclude) {
457	int mask = (subpopSize >>> DYER_DECAY) - 1;
458	int d = subpopSize - 1;
459	while (d == exclude \|\| (rng.next() & mask) != mask) {
460	if (--d < 0)
461	d = subpopSize - 1;
462	}
463	return d;
464	}
465
466	/**
467	* Select a random strand of b's.
468	* @param breeder the breeder
469	*/
470	void chooseStrand(Chromosome breeder) {
471	int[] bs = breeder.alleles;
472	int length = bs.length;
473	int strandLength = MIN_STRAND_LENGTH;
474	while (strandLength < length &&
475	(rng.next() & RANDOM_STRAND_MASK) != RANDOM_STRAND_MASK)
476	strandLength++;
477	strand.strandLength = strandLength;
478	int[] ss = strand.alleles;
479	int k = (rng.next() & 0x7FFFFFFF) % length;
480	for (int i = 0; i < strandLength; ++i) {
481	ss[i] = bs[k];
482	if (++k >= length) k = 0;
483	}
484	}
485
486	/**
487	* Copy current strand to start of c's, and then append all
488	* remaining b's that aren't in the strand.
489	* @param breeder the breeder
490	* @param child the child
491	*/
492	void cross(Chromosome breeder, Chromosome child) {
493	for (int k = 0; k < inTour.length; ++k) // clear bitset
494	inTour[k] = 0;
495
496	// Copy current strand to c
497	int[] cs = child.alleles;
498	int ssize = strand.strandLength;
499	int[] ss = strand.alleles;
500	int i;
501	for (i = 0; i < ssize; ++i) {
502	int x = ss[i];
503	cs[i] = x;
504	inTour[x >>> 5] \|= 1 << (x & 31); // record in bit set
505	}
506
507	// Find index of matching origin in b
508	int first = cs[0];
509	int j = 0;
510	int[] bs = breeder.alleles;
511	while (bs[j] != first)
512	++j;
513
514	// Append remaining b's that aren't already in tour
515	while (i < cs.length) {
516	if (++j >= bs.length) j = 0;
517	int x = bs[j];
518	if ((inTour[x >>> 5] & (1 << (x & 31))) == 0)
519	cs[i++] = x;
520	}
521
522	}
523
524	/**
525	* Fix the sort order of a changed Chromosome c at position k
526	* @param c the chromosome
527	* @param k the index
528	*/
529	void fixOrder(Chromosome c, int k) {
530	Chromosome[] cs = chromosomes;
531	int oldFitness = c.fitness;
532	c.recalcFitness();
533	int newFitness = c.fitness;
534	if (newFitness < oldFitness) {
535	int j = k;
536	int p = j - 1;
537	while (p >= 0 && cs[p].fitness > newFitness) {
538	cs[j] = cs[p];
539	j = p--;
540	}
541	cs[j] = c;
542	} else if (newFitness > oldFitness) {
543	int j = k;
544	int n = j + 1;
545	while (n < cs.length && cs[n].fitness < newFitness) {
546	cs[j] = cs[n];
547	j = n++;
548	}
549	cs[j] = c;
550	}
551	}
552	}
553
554	/**
555	* A Chromosome is a candidate TSP tour.
556	*/
557	static final class Chromosome implements Comparable {
558	/** Index of cities in tour order */
559	final int[] alleles;
560	/** Total tour length */
561	int fitness;
562
563	/**
564	* Initialize to random tour
565	*/
566	Chromosome(int length, RNG random) {
567	alleles = new int[length];
568	for (int i = 0; i < length; i++)
569	alleles[i] = i;
570	for (int i = length - 1; i > 0; i--) {
571	int idx = (random.next() & 0x7FFFFFFF) % alleles.length;
572	int tmp = alleles[i];
573	alleles[i] = alleles[idx];
574	alleles[idx] = tmp;
575	}
576	recalcFitness();
577	}
578
579	public int compareTo(Object x) { // to enable sorting
580	int xf = ((Chromosome)x).fitness;
581	int f = fitness;
582	return ((f == xf)? 0 :((f < xf)? -1 : 1));
583	}
584
585	void recalcFitness() {
586	int[] a = alleles;
587	int len = a.length;
588	int p = a[0];
589	long f = cities.distanceBetween(a[len-1], p);
590	for (int i = 1; i < len; i++) {
591	int n = a[i];
592	f += cities.distanceBetween(p, n);
593	p = n;
594	}
595	fitness = (int)(f / len);
596	}
597
598	/**
599	* Return tour length for points scaled in [0, 1).
600	*/
601	double unitTourLength() {
602	int[] a = alleles;
603	int len = a.length;
604	int p = a[0];
605	double f = cities.unitDistanceBetween(a[len-1], p);
606	for (int i = 1; i < len; i++) {
607	int n = a[i];
608	f += cities.unitDistanceBetween(p, n);
609	p = n;
610	}
611	return f;
612	}
613
614	/**
615	* Check that this tour visits each city
616	*/
617	void validate() {
618	int len = alleles.length;
619	boolean[] used = new boolean[len];
620	for (int i = 0; i < len; ++i)
621	used[alleles[i]] = true;
622	for (int i = 0; i < len; ++i)
623	if (!used[i])
624	throw new Error("Bad tour");
625	}
626
627	}
628
629	/**
630	* A Strand is a random sub-sequence of a Chromosome. Each subpop
631	* creates only one strand, and then trades it with others,
632	* refilling it on each iteration.
633	*/
634	static final class Strand {
635	final int[] alleles;
636	int strandLength;
637	Strand(int length) { alleles = new int[length]; }
638	}
639
640	/**
641	* A collection of (x,y) points that represent cities.
642	*/
643	static final class CitySet {
644
645	final int length;
646	final int[] xPts;
647	final int[] yPts;
648	final int[][] distances;
649
650	CitySet(int n) {
651	this.length = n;
652	this.xPts = new int[n];
653	this.yPts = new int[n];
654	this.distances = new int[n][n];
655
656	RNG random = new RNG();
657	for (int i = 0; i < n; i++) {
658	xPts[i] = (random.next() & 0x7FFFFFFF);
659	yPts[i] = (random.next() & 0x7FFFFFFF);
660	}
661
662	for (int i = 0; i < n; i++) {
663	for (int j = 0; j < n; j++) {
664	double dx = (double)xPts[i] - (double)xPts[j];
665	double dy = (double)yPts[i] - (double)yPts[j];
666	double dd = Math.hypot(dx, dy) / 2.0;
667	long ld = Math.round(dd);
668	distances[i][j] = (ld >= Integer.MAX_VALUE)?
669	Integer.MAX_VALUE : (int)ld;
670	}
671	}
672	}
673
674	/**
675	* Returns the cached distance between a pair of cities
676	*/
677	int distanceBetween(int i, int j) {
678	return distances[i][j];
679	}
680
681	// Scale ints to doubles in [0,1)
682	static final double PSCALE = (double)0x80000000L;
683
684	/**
685	* Return distance for points scaled in [0,1). This simplifies
686	* checking results. The expected optimal TSP for random
687	* points is believed to be around 0.76 * sqrt(N). For papers
688	* discussing this, see
689	* http://www.densis.fee.unicamp.br/~moscato/TSPBIB_home.html
690	*/
691	double unitDistanceBetween(int i, int j) {
692	double dx = ((double)xPts[i] - (double)xPts[j]) / PSCALE;
693	double dy = ((double)yPts[i] - (double)yPts[j]) / PSCALE;
694	return Math.hypot(dx, dy);
695	}
696
697	}
698
699	/**
700	* Cheap XorShift random number generator
701	*/
702	static final class RNG {
703	/** Seed generator for XorShift RNGs */
704	static final Random seedGenerator = new Random();
705
706	int seed;
707	RNG(int seed) { this.seed = seed; }
708	RNG() { this.seed = seedGenerator.nextInt() \| 1; }
709
710	int next() {
711	int x = seed;
712	x ^= x << 6;
713	x ^= x >>> 21;
714	x ^= x << 7;
715	seed = x;
716	return x;
717	}
718	}
719
720	static final class ProgressMonitor extends Thread {
721	final Population pop;
722	ProgressMonitor(Population p) { pop = p; }
723	public void run() {
724	double time = 0;
725	try {
726	while (!Thread.interrupted()) {
727	sleep(SNAPSHOT_RATE);
728	time += SNAPSHOT_RATE;
729	pop.printSnapshot(time / 1000.0);
730	}
731	} catch (InterruptedException ie) {}
732	}
733	}
734	}