SHOW:
|
|
- or go back to the newest paste.
| 1 | package frans; | |
| 2 | ||
| 3 | /** | |
| 4 | * Class to test access time (latency) for cache and RAM. | |
| 5 | * Sequential access and random access of values in a long[] array. | |
| 6 | * | |
| 7 | * @author Frans Lundberg | |
| 8 | */ | |
| 9 | public class MemoryAccess {
| |
| 10 | static final int SIZE_CACHE = 100000; // 100003 is prime. Total: 0.8 MB. | |
| 11 | static final long REPS1_CACHE = 10 * 1000; // for seq | |
| 12 | static final long REPS2_CACHE = 100 * 1000000; // for random | |
| 13 | static final int SIZE_RAM = 500 * SIZE_CACHE; // 400 MB of data | |
| 14 | static final long REPS1_RAM = 10; // for seq | |
| 15 | static final long REPS2_RAM = 10 * 1000000; // for random | |
| 16 | ||
| 17 | /** | |
| 18 | * Sequential memory accesses to a long[] array. | |
| 19 | * 'size' must be SIZE_CACHE or SIZE_RAM. | |
| 20 | */ | |
| 21 | Result seq(final int size) {
| |
| 22 | long[] data = new long[size]; | |
| 23 | ||
| 24 | for (int i = 0; i < data.length; i++) {
| |
| 25 | data[i] = 1; | |
| 26 | } | |
| 27 | ||
| 28 | long t0 = 0; | |
| 29 | long sum = 0; | |
| 30 | long count; | |
| 31 | ||
| 32 | t0 = System.nanoTime(); | |
| 33 | ||
| 34 | if (size == SIZE_CACHE) {
| |
| 35 | count = SIZE_CACHE * REPS1_CACHE; | |
| 36 | for (int k = 0; k < REPS1_CACHE; k++) {
| |
| 37 | for (int i = 0; i < SIZE_CACHE; i++) {
| |
| 38 | sum += data[i]; | |
| 39 | } | |
| 40 | } | |
| 41 | } else if (size == SIZE_RAM) {
| |
| 42 | count = SIZE_RAM * REPS1_RAM; | |
| 43 | for (int k = 0; k < REPS1_RAM; k++) {
| |
| 44 | for (int i = 0; i < SIZE_RAM; i++) {
| |
| 45 | sum += data[i]; | |
| 46 | } | |
| 47 | } | |
| 48 | } else {
| |
| 49 | throw new Error("Bad size: " + size);
| |
| 50 | } | |
| 51 | ||
| 52 | double time = (System.nanoTime() - t0) * 1e-9; | |
| 53 | return new Result(count, time, sum); | |
| 54 | } | |
| 55 | ||
| 56 | ||
| 57 | /** | |
| 58 | * Random memory access to a long[] array. | |
| 59 | * 'size' must be SIZE_CACHE or SIZE_RAM. | |
| 60 | */ | |
| 61 | Result random(final int size) {
| |
| 62 | long[] data = new long[size]; | |
| 63 | long t0; | |
| 64 | long sum = 0; | |
| 65 | long index = 0; | |
| 66 | long count; | |
| 67 | ||
| 68 | for (int i = 0; i < size; i++) {
| |
| 69 | data[i] = 1L; | |
| 70 | } | |
| 71 | ||
| 72 | t0 = System.nanoTime(); | |
| 73 | ||
| 74 | if (size == SIZE_CACHE) {
| |
| 75 | count = REPS2_CACHE; | |
| 76 | for (int k = 0; k < REPS2_CACHE; k++) {
| |
| 77 | sum += data[(int)(index % SIZE_CACHE)]; | |
| 78 | index += 836413; | |
| 79 | } | |
| 80 | } else if (size == SIZE_RAM) {
| |
| 81 | count = REPS2_RAM; | |
| 82 | for (int k = 0; k < REPS2_RAM; k++) {
| |
| 83 | sum += data[(int)(index % SIZE_RAM)]; | |
| 84 | index += 836413; | |
| 85 | } | |
| 86 | } else {
| |
| 87 | throw new Error("Bad size: " + size);
| |
| 88 | } | |
| 89 | ||
| 90 | double time = (System.nanoTime() - t0) * 1e-9; | |
| 91 | return new Result(count, time, sum); | |
| 92 | } | |
| 93 | ||
| 94 | void runTest(String name, TestMethod test) {
| |
| 95 | int reps = 8; | |
| 96 | double minTime = Double.MAX_VALUE; | |
| 97 | Result winner = null; | |
| 98 | ||
| 99 | System.out.println("---- " + name + " ----");
| |
| 100 | ||
| 101 | for (int i = 0; i < reps; i++) {
| |
| 102 | Result res = test.run(); | |
| 103 | if (res.time < minTime) {
| |
| 104 | winner = res; | |
| 105 | } | |
| 106 | } | |
| 107 | ||
| 108 | System.out.println(winner.toString()); | |
| 109 | } | |
| 110 | ||
| 111 | /** To store the result. */ | |
| 112 | static class Result {
| |
| 113 | long count; | |
| 114 | long sum; | |
| 115 | double time; | |
| 116 | ||
| 117 | Result(long count, double time, long sum) {
| |
| 118 | this.count = count; | |
| 119 | this.time = time; | |
| 120 | this.sum = sum; | |
| 121 | } | |
| 122 | ||
| 123 | public String toString() {
| |
| 124 | double nanos = (time * 1e9) / count; | |
| 125 | return String.format("Access time %.2f ns, time: %.3f s (sum:%d).\n", nanos, time, sum);
| |
| 126 | } | |
| 127 | } | |
| 128 | ||
| 129 | static interface TestMethod {
| |
| 130 | public Result run(); | |
| 131 | } | |
| 132 | ||
| 133 | void go() {
| |
| 134 | runTest("cacheSeq", new TestMethod() {
| |
| 135 | public Result run() {
| |
| 136 | return seq(SIZE_CACHE); | |
| 137 | } | |
| 138 | }); | |
| 139 | ||
| 140 | runTest("cacheRan", new TestMethod() {
| |
| 141 | public Result run() {
| |
| 142 | return random(SIZE_CACHE); | |
| 143 | } | |
| 144 | }); | |
| 145 | ||
| 146 | runTest("ramSeq", new TestMethod() {
| |
| 147 | public Result run() {
| |
| 148 | return seq(SIZE_RAM); | |
| 149 | } | |
| 150 | }); | |
| 151 | ||
| 152 | runTest("ramRan", new TestMethod() {
| |
| 153 | public Result run() {
| |
| 154 | return random(SIZE_RAM); | |
| 155 | } | |
| 156 | }); | |
| 157 | } | |
| 158 | ||
| 159 | public static void main(String[] args) {
| |
| 160 | new MemoryAccess().go(); | |
| 161 | } | |
| 162 | } | |
| 163 | ||
| 164 | /* | |
| 165 | 131106 | |
| 166 | ====== | |
| 167 | ||
| 168 | - | "sequential" / "random" access are better terms. Now used. |
| 168 | + | |
| 169 | - | Refactored. Result: |
| 169 | + | |
| 170 | cacheRan 3.16 ns, 6.64 cycles | |
| 171 | ramSeq 0.56 ns, 1.37 cycles | |
| 172 | ramRan 20.07 ns, 42.16 cycles | |
| 173 | - | cacheRan 4.39 ns, 9.20 cycles |
| 173 | + | |
| 174 | On my computer "Cissi": | |
| 175 | Ubuntu 12.04, CPU: Intel(R) Core(TM) i7-3687U CPU @ 2.10GHz, Cache: 4096 KiB. | |
| 176 | Java version: | |
| 177 | Oracle JVM 1.7.0_25. | |
| 178 | Java HotSpot(TM) 64-Bit Server VM (build 23.25-b01, mixed mode) | |
| 179 | ||
| 180 | sudo cat /proc/cpuinfo | grep "cache" --> | |
| 181 | cache size : 4096 KB | |
| 182 | sudo cat /proc/cpuinfo | fgrep 'cpu MHz' | |
| 183 | Max freq: cpu MHz : 2101.000 <--> 0.476 ns / cycle | |
| 184 | */ |