For a high performance blocked bloom filter, I would like to align data to cache lines. (I know it's easier to do such tricks in C, but I would like to use Java.)
I do have a solution, but I'm not sure if it's correct, or if there is a better way. My solution tries to find the start of the cache line using the following algorithm:
- for each possible offset o (0..63; I assume cache line length of 64)
- start a thread that reads from data[o] and writes that to data[o + 8]
- in the main thread, write '1' to data[o], and wait until that ends up in data[o + 8] (so wait for the other thread)
- repeat that
Then, measure how fast this was, basically how many increments for a loop of 1 million (in each thread). My logic is, it is slower if the data is in a different cache line.
Here my code:
public static void main(String... args) {
for(int i=0; i<20; i++) {
int size = (int) (1000 + Math.random() * 1000);
byte[] data = new byte[size];
int cacheLineOffset = getCacheLineOffset(data);
System.out.println("offset: " + cacheLineOffset);
}
}
private static int getCacheLineOffset(byte[] data) {
for (int i = 0; i < 10; i++) {
int x = tryGetCacheLineOffset(data, i + 3);
if (x != -1) {
return x;
}
}
System.out.println("Cache line start not found");
return 0;
}
private static int tryGetCacheLineOffset(byte[] data, int testCount) {
// assume synchronization between two threads is faster(?)
// if each thread works on the same cache line
int[] counters = new int[64];
int testOffset = 8;
for (int test = 0; test < testCount; test++) {
for (int offset = 0; offset < 64; offset++) {
final int o = offset;
final Semaphore sema = new Semaphore(0);
Thread t = new Thread() {
public void run() {
try {
sema.acquire();
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
for (int i = 0; i < 1000000; i++) {
data[o + testOffset] = data[o];
}
}
};
t.start();
sema.release();
data[o] = 1;
int counter = 0;
byte waitfor = 1;
for (int i = 0; i < 1000000; i++) {
byte x = data[o + testOffset];
if (x == waitfor) {
data[o]++;
counter++;
waitfor++;
}
}
try {
t.join();
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
counters[offset] += counter;
}
}
Arrays.fill(data, 0, testOffset + 64, (byte) 0);
int low = Integer.MAX_VALUE, high = Integer.MIN_VALUE;
for (int i = 0; i < 64; i++) {
// average of 3
int avg3 = (counters[(i - 1 + 64) % 64] + counters[i] + counters[(i + 1) % 64]) / 3;
low = Math.min(low, avg3);
high = Math.max(high, avg3);
}
if (low * 1.1 > high) {
// no significant difference between low and high
return -1;
}
int lowCount = 0;
boolean[] isLow = new boolean[64];
for (int i = 0; i < 64; i++) {
if (counters[i] < (low + high) / 2) {
isLow[i] = true;
lowCount++;
}
}
if (lowCount != 8) {
// unclear
return -1;
}
for (int i = 0; i < 64; i++) {
if (isLow[(i - 1 + 64) % 64] && !isLow[i]) {
return i;
}
}
return -1;
}
It prints (example):
offset: 16
offset: 24
offset: 0
offset: 40
offset: 40
offset: 8
offset: 24
offset: 40
...
So arrays in Java seems to be aligned to 8 bytes.
You know that the GC can move objects... so your perfectly aligned array may get misaligned later.
I'd try ByteBuffer
; I guess, a direct one gets aligned a lot (to a page boundary).
Unsafe can give you the address and with JNI, you can get an array pinned.