|
| 1 | +package org.fastfilter.xor; |
| 2 | + |
| 3 | +import java.util.Locale; |
| 4 | + |
| 5 | +import org.fastfilter.utils.Hash; |
| 6 | +import org.fastfilter.utils.RandomGenerator; |
| 7 | + |
| 8 | +/** |
| 9 | + * Calculate the best segment length for various sizes, and the probability of |
| 10 | + * mapping, for the fuse filter. Specially interesting are "small" set sizes |
| 11 | + * between 100 and 1 million. |
| 12 | + * |
| 13 | + * Unlike the regular fuse filter, which has less density at the beginning and |
| 14 | + * the end of the array, here density is only lower at the end of the array. |
| 15 | + * Index computation is slower, but load should be slightly higher. |
| 16 | + * |
| 17 | + * See also "Peeling Close to the Orientability Threshold - Spatial Coupling in |
| 18 | + * Hashing-Based Data Structures" |
| 19 | + * |
| 20 | + */ |
| 21 | +public class ProbabilityCFuse2 { |
| 22 | + |
| 23 | + private static final int HASHES = 3; |
| 24 | + private static final int BITS_PER_FINGERPRINT = 8; |
| 25 | + |
| 26 | +// size 100 load 0.45 segmentLength 64 bits/key 17.8 p 0.86 |
| 27 | +// size 1000 load 0.70 segmentLength 256 bits/key 11.4 p 0.88 |
| 28 | +// size 10000 load 0.80 segmentLength 1024 bits/key 10.0 p 0.90 |
| 29 | +// size 100000 load 0.85 segmentLength 4096 bits/key 9.4 p 0.87 |
| 30 | + |
| 31 | + private static void testProb() { |
| 32 | + int segmentLength = 16; |
| 33 | + int segmentCount = 100; |
| 34 | + int[] counts = new int[200]; |
| 35 | + for(int i=0; i<1000000; i++) { |
| 36 | + for(int index = 0; index < 3; index++) { |
| 37 | + long seed = 0; |
| 38 | + long key = i; |
| 39 | + long hash = Hash.hash64(key, seed); |
| 40 | + |
| 41 | +// int r0 = (int) Hash.hash64(hash, 1); |
| 42 | +// int x = Hash.reduce(r0, segmentCount); |
| 43 | +// int h0 = x + (int) (Hash.hash64(hash, 2) & (segmentLength - 1)); |
| 44 | +// int h1 = x + (int) (Hash.hash64(hash, 3) & (segmentLength - 1)); |
| 45 | +// int h2 = x + (int) (Hash.hash64(hash, 4) & (segmentLength - 1)); |
| 46 | +// |
| 47 | + int r0 = (int) Hash.hash64(hash, 1); |
| 48 | + int x = Hash.reduce(r0, segmentCount * 2 + segmentLength - 1); |
| 49 | + int h0 = x + (int) (Hash.hash64(hash, 2) & (segmentLength - 1)); |
| 50 | + int h1 = x + (int) (Hash.hash64(hash, 3) & (segmentLength - 1)); |
| 51 | + int h2 = x + (int) (Hash.hash64(hash, 4) & (segmentLength - 1)); |
| 52 | + h0 = Math.abs(h0 - segmentCount - segmentLength + 1); |
| 53 | + h1 = Math.abs(h1 - segmentCount - segmentLength + 1); |
| 54 | + h2 = Math.abs(h2 - segmentCount - segmentLength + 1); |
| 55 | + int idx = index == 0 ? h0 : index == 1 ? h1 : h2; |
| 56 | + counts[idx]++; |
| 57 | + } |
| 58 | + } |
| 59 | + for(int i=0; i<counts.length; i++) { |
| 60 | + System.out.println(i + " " + counts[i]); |
| 61 | + } |
| 62 | + |
| 63 | + } |
| 64 | + |
| 65 | + public static void main(String... args) { |
| 66 | +//testProb(); |
| 67 | +//if(true)return; |
| 68 | + |
| 69 | + |
| 70 | +// for(int size = 100_000; size < 1_000_000; size *= 10) { |
| 71 | + for(int size = 1; size < 1_000_000; size *= 10) { |
| 72 | + // for(int size = 1; size < 1_000_000; size = (size < 100) ? (size + 1) : (int) (size * 1.1)) { |
| 73 | + Data best = null; |
| 74 | + for (int segmentLengthBits = 3; segmentLengthBits <= 12; segmentLengthBits++) { |
| 75 | +// for (int segmentLengthBits = 3; segmentLengthBits < 14; segmentLengthBits++) { |
| 76 | + int segmentLength = 1 << segmentLengthBits; |
| 77 | + if (segmentLength > size) { |
| 78 | + break; |
| 79 | + } |
| 80 | + for(double load = 0.85; load > 0.3; load-= 0.05) { |
| 81 | + Data d = getProbability(size, segmentLengthBits, load, best); |
| 82 | + if (d != null && d.p > 0.85) { |
| 83 | + if (best == null || d.bitsPerKey < best.bitsPerKey) { |
| 84 | + best = d; |
| 85 | + } |
| 86 | + break; |
| 87 | + } |
| 88 | + } |
| 89 | + } |
| 90 | + if (best != null) { |
| 91 | + System.out.println(best); |
| 92 | +// for(int i=0; i<100; i++) { |
| 93 | +// System.out.println(i + ": " + best.data[i]); |
| 94 | +// } |
| 95 | + } |
| 96 | + } |
| 97 | + } |
| 98 | + |
| 99 | + static Data getProbability(int size, int segmentLengthBits, double load, Data best) { |
| 100 | + int segmentLength = 1 << segmentLengthBits; |
| 101 | + int arrayLength = (int) (size / load); |
| 102 | + if (arrayLength <= 0) { |
| 103 | + return null; |
| 104 | + } |
| 105 | + int segmentCount = arrayLength - 1 * segmentLength; |
| 106 | + if (segmentCount <= 0) { |
| 107 | + return null; |
| 108 | + } |
| 109 | + Data d = new Data(); |
| 110 | + d.size = size; |
| 111 | + d.load = load; |
| 112 | + d.segmentLength = segmentLength; |
| 113 | + d.bitsPerKey = (double) arrayLength * 8 / size; |
| 114 | + if (best != null && d.bitsPerKey > best.bitsPerKey) { |
| 115 | + return null; |
| 116 | + } |
| 117 | + // System.out.println(" test " + d); |
| 118 | + int successCount = 0; |
| 119 | + int testCount = Math.max(10, 10_000_000 / size); |
| 120 | + for(int seed = 0; seed < testCount; seed++) { |
| 121 | + long[] keys = new long[size]; |
| 122 | + RandomGenerator.createRandomUniqueListFast(keys, seed); |
| 123 | + int[] success = testMapping(keys, segmentLengthBits, segmentCount, arrayLength, seed); |
| 124 | + if (success != null) { |
| 125 | + d.data = success; |
| 126 | + successCount++; |
| 127 | + } |
| 128 | + } |
| 129 | + double p = 1.0 * successCount / testCount; |
| 130 | + d.p = p; |
| 131 | + return d; |
| 132 | + } |
| 133 | + |
| 134 | + public static int[] testMapping(long[] keys, int segmentLengthBits, int segmentCount, int arrayLength, long seed) { |
| 135 | + int segmentLength = 1 << segmentLengthBits; |
| 136 | + int size = keys.length; |
| 137 | + int m = arrayLength; |
| 138 | + long[] reverseOrder = new long[size]; |
| 139 | + byte[] reverseH = new byte[size]; |
| 140 | + int reverseOrderPos; |
| 141 | + seed = Hash.randomSeed(); |
| 142 | + byte[] t2count = new byte[m]; |
| 143 | + long[] t2 = new long[m]; |
| 144 | + for (long k : keys) { |
| 145 | + for (int hi = 0; hi < HASHES; hi++) { |
| 146 | + int h = getHash(segmentLengthBits, segmentLength, segmentCount, k, seed, hi); |
| 147 | + t2[h] ^= k; |
| 148 | + if (t2count[h] > 120) { |
| 149 | + // probably something wrong with the hash function |
| 150 | + throw new IllegalArgumentException(); |
| 151 | + } |
| 152 | + t2count[h]++; |
| 153 | + } |
| 154 | + } |
| 155 | + reverseOrderPos = 0; |
| 156 | + int[] alone = new int[arrayLength]; |
| 157 | + int alonePos = 0; |
| 158 | + for (int i = 0; i < arrayLength; i++) { |
| 159 | + if (t2count[ i] == 1) { |
| 160 | + alone[alonePos++] = i; |
| 161 | + } |
| 162 | + } |
| 163 | + int found = -1; |
| 164 | + while (alonePos > 0) { |
| 165 | + int i = alone[--alonePos]; |
| 166 | + if (t2count[i] <= 0) { |
| 167 | + continue; |
| 168 | + } |
| 169 | + if (t2count[i] != 1) { |
| 170 | + throw new AssertionError(); |
| 171 | + } |
| 172 | + --t2count[i]; |
| 173 | + long k = t2[i]; |
| 174 | + for (int hi = 0; hi < HASHES; hi++) { |
| 175 | + int h = getHash(segmentLengthBits, segmentLength, segmentCount, k, seed, hi); |
| 176 | + int newCount = --t2count[h]; |
| 177 | + if (h == i) { |
| 178 | + found = hi; |
| 179 | + } else { |
| 180 | + if (newCount == 1) { |
| 181 | + alone[alonePos++] = h; |
| 182 | + } |
| 183 | + t2[h] ^= k; |
| 184 | + } |
| 185 | + } |
| 186 | + reverseOrder[reverseOrderPos] = k; |
| 187 | + reverseH[reverseOrderPos] = (byte) found; |
| 188 | + reverseOrderPos++; |
| 189 | + } |
| 190 | + if (reverseOrderPos != size) { |
| 191 | + return null; |
| 192 | + } |
| 193 | + byte[] fp = new byte[m]; |
| 194 | + for (int i = reverseOrderPos - 1; i >= 0; i--) { |
| 195 | + long k = reverseOrder[i]; |
| 196 | + found = reverseH[i]; |
| 197 | + int change = -1; |
| 198 | + long hash = Hash.hash64(k, seed); |
| 199 | + int xor = fingerprint(hash); |
| 200 | + for (int hi = 0; hi < HASHES; hi++) { |
| 201 | + int h = getHash(segmentLengthBits, segmentLength, segmentCount, k, seed, hi); |
| 202 | + if (found == hi) { |
| 203 | + change = h; |
| 204 | + } else { |
| 205 | + xor ^= fp[h]; |
| 206 | + } |
| 207 | + } |
| 208 | + fp[change] = (byte) xor; |
| 209 | + } |
| 210 | + int[] nonZero = new int[100]; |
| 211 | + for(int i=0; i<fp.length; i++) { |
| 212 | + if (fp[i] != 0) { |
| 213 | + nonZero[i * 100 / fp.length]++; |
| 214 | + } |
| 215 | + } |
| 216 | + return nonZero; |
| 217 | + } |
| 218 | + |
| 219 | + private static int fingerprint(long hash) { |
| 220 | + return (int) (hash & ((1 << BITS_PER_FINGERPRINT) - 1)); |
| 221 | + } |
| 222 | + |
| 223 | + private static int getHash(int segmentLengthBits, int segmentLength, int segmentCount, long key, long seed, int index) { |
| 224 | + long hash = Hash.hash64(key, seed); |
| 225 | + int r0 = (int) Hash.hash64(hash, 1); |
| 226 | + /* |
| 227 | + int x = Hash.reduce(r0, segmentCount * 2 - 1); |
| 228 | + int h0 = x + (int) (Hash.hash64(hash, 2) & (2 * segmentLength - 1)); |
| 229 | + int h1 = x + (int) (Hash.hash64(hash, 3) & (2 * segmentLength - 1)); |
| 230 | + int h2 = x + (int) (Hash.hash64(hash, 4) & (2 * segmentLength - 1)); |
| 231 | + h0 = Math.abs(h0 - segmentCount - segmentLength + 1); |
| 232 | + h1 = Math.abs(h1 - segmentCount - segmentLength + 1); |
| 233 | + h2 = Math.abs(h2 - segmentCount - segmentLength + 1); |
| 234 | + */ |
| 235 | + |
| 236 | + int x = Hash.reduce(r0, segmentCount * 2 + segmentLength - 1); |
| 237 | + int h0 = x + (int) (Hash.hash64(hash, 2) & (segmentLength - 1)); |
| 238 | + int h1 = x + (int) (Hash.hash64(hash, 3) & (segmentLength - 1)); |
| 239 | + int h2 = x + (int) (Hash.hash64(hash, 4) & (segmentLength - 1)); |
| 240 | + h0 = Math.abs(h0 - segmentCount - segmentLength + 1); |
| 241 | + h1 = Math.abs(h1 - segmentCount - segmentLength + 1); |
| 242 | + h2 = Math.abs(h2 - segmentCount - segmentLength + 1); |
| 243 | + |
| 244 | + return index == 0 ? h0 : index == 1 ? h1 : h2; |
| 245 | + } |
| 246 | + |
| 247 | + static class Data { |
| 248 | + int size; |
| 249 | + double load; |
| 250 | + int segmentLength; |
| 251 | + double bitsPerKey; |
| 252 | + double p; |
| 253 | + int[] data; |
| 254 | + |
| 255 | + public String toString() { |
| 256 | + return String.format(Locale.ENGLISH, "size %d load %.2f " + |
| 257 | + "segmentLength %d bits/key %.1f p %.2f" |
| 258 | + , size, load, segmentLength, bitsPerKey, p); |
| 259 | + } |
| 260 | + |
| 261 | + } |
| 262 | + |
| 263 | +} |
0 commit comments