Basic Implementation with Memory Mapped File, SIMD Search

By gamlerhart
This commit is contained in:
Roman Stoffel 2024-01-10 22:25:45 +01:00 committed by Gunnar Morling
parent 80328e4898
commit 64a78c3880
2 changed files with 289 additions and 0 deletions

19
calculate_average_gamlerhart.sh Executable file
View File

@ -0,0 +1,19 @@
#!/bin/sh
#
# Copyright 2023 The original authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
JAVA_OPTS="--enable-preview --add-modules=jdk.incubator.vector"
java $JAVA_OPTS --class-path target/average-1.0.0-SNAPSHOT.jar dev.morling.onebrc.CalculateAverage_gamlerhart

View File

@ -0,0 +1,270 @@
/*
* Copyright 2023 The original authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package dev.morling.onebrc;
import jdk.incubator.vector.*;
import java.lang.foreign.Arena;
import java.lang.foreign.MemorySegment;
import java.lang.foreign.ValueLayout;
import java.nio.ByteOrder;
import java.nio.channels.FileChannel;
import java.nio.file.Path;
import java.util.TreeMap;
import static java.lang.Double.doubleToRawLongBits;
import static java.lang.Double.longBitsToDouble;
import static java.lang.foreign.ValueLayout.*;
/**
* Broad experiments in this implementation:
* - Memory-Map the file with new MemorySegments
* - Use SIMD/vectorized search for the semicolon and new line feeds
* - Use SIMD/vectorized comparison for the 'key'
* <p>
* Absolute stupid things / performance left on the table
* - Single Threaded! Multi threading planned.
* - The hash map/table is super basic.
* - Hash table implementation / hashing has no resizing and is quite basic
* - Zero time spend on profiling =)
* <p>
* <p>
* Cheats used:
* - Only works with Unix line feed \n
* - double parsing is only accepting XX.X and X.X
* - HashMap has no resizing, check, horrible hash etc.
* - Used the double parsing from yemreinci
*/
public class CalculateAverage_gamlerhart {
private static final String FILE = "./measurements.txt";
final static VectorSpecies<Byte> byteVec = ByteVector.SPECIES_PREFERRED;
final static Vector<Byte> zero = byteVec.zero();
final static int vecLen = byteVec.length();
final static Vector<Byte> semiColon = byteVec.broadcast(';');
final static VectorMask<Byte> allTrue = byteVec.maskAll(true);
final static ValueLayout.OfInt INT_UNALIGNED_BIG_ENDIAN = ValueLayout.JAVA_INT_UNALIGNED.withOrder(ByteOrder.BIG_ENDIAN);
public static void main(String[] args) throws Exception {
try (var arena = Arena.ofConfined();
FileChannel fc = FileChannel.open(Path.of(FILE))) {
var map = new PrivateHashMap();
long fileSize = fc.size();
MemorySegment fileContent = fc.map(FileChannel.MapMode.READ_ONLY, 0, fileSize, arena);
var loopBound = byteVec.loopBound(fileSize) - vecLen;
for (long i = 0; i < fileSize;) {
long nameStart = i;
int simdSearchEnd = 0;
int nameLen = 0;
// Vectorized Search
if (i < loopBound) {
do {
var vec = byteVec.fromMemorySegment(fileContent, i, ByteOrder.BIG_ENDIAN);
var hasSemi = vec.eq(semiColon);
simdSearchEnd = hasSemi.firstTrue();
i += simdSearchEnd;
nameLen += simdSearchEnd;
} while (simdSearchEnd == vecLen && i < loopBound);
}
// Left-over search
while (loopBound <= i && fileContent.get(JAVA_BYTE, i) != ';') {
nameLen++;
i++;
}
i++; // Consume ;
// Copied from yemreinci. I mostly wanted to experiment the vector math, not with parsing =)
double val;
{
boolean negative = false;
if ((fileContent.get(JAVA_BYTE, i)) == '-') {
negative = true;
i++;
}
byte b;
double temp;
if ((b = fileContent.get(JAVA_BYTE, i + 1)) == '.') { // temperature is in either XX.X or X.X form
temp = (fileContent.get(JAVA_BYTE, i) - '0') + (fileContent.get(JAVA_BYTE, i + 2) - '0') / 10.0;
i += 3;
}
else {
temp = (fileContent.get(JAVA_BYTE, i) - '0') * 10 + (b - '0')
+ (fileContent.get(JAVA_BYTE, i + 3) - '0') / 10.0;
i += 4;
}
val = (negative ? -temp : temp);
}
i++; // Consume \n
map.add(fileContent, nameStart, nameLen, val);
}
// System.out.println(map.debug_reprobeMax);
var measurements = new TreeMap<String, ResultRow>();
map.fill(fileContent, measurements);
System.out.println(measurements);
}
}
private static class PrivateHashMap {
private static final int SIZE_SHIFT = 14;
public static final int SIZE = 1 << SIZE_SHIFT;
public static int MASK = 0xFFFFFFFF >>> (32 - SIZE_SHIFT);
public static long SHIFT_POS = 16;
public static long MASK_POS = 0xFFFFFFFFFFFF0000L;
public static long MASK_LEN = 0x000000000000FFFFL;
// Encoding:
// - Key: long
// - 48 bits index, 16 bits length
// - min: double
// - max: double
// - sum: double
// - double: double
final long[] keyValues = new long[SIZE * 5];
// int debug_size = 0;
// int debug_reprobeMax = 0;
public PrivateHashMap() {
}
public void add(MemorySegment file, long pos, int len, double val) {
int hashCode = 1;
int i = 0;
int intBound = (len / 4) * 4;
for (; i < intBound; i += 4) {
int v = file.get(INT_UNALIGNED_BIG_ENDIAN, pos + i);
hashCode = 31 * hashCode + v;
}
for (; i < len; i++) {
int v = file.get(JAVA_BYTE, pos + i);
hashCode = 31 * hashCode + v;
}
doAdd(file, hashCode, pos, len, val);
}
private void doAdd(MemorySegment file, int hash, long pos, int len, double val) {
// var debug = new String(file.asSlice(pos, len).toArray(ValueLayout.JAVA_BYTE));
int slot = hash & MASK;
for (var probe = 0; probe < 20000; probe++) {
var iSl = ((slot + probe) & MASK) * 5;
var slotEntry = keyValues[iSl];
var emtpy = slotEntry == 0;
if (emtpy) {
long keyInfo = pos << SHIFT_POS | len;
long valueBits = doubleToRawLongBits(val);
keyValues[iSl] = keyInfo;
keyValues[iSl + 1] = valueBits;
keyValues[iSl + 2] = valueBits;
keyValues[iSl + 3] = valueBits;
keyValues[iSl + 4] = 1;
// debug_size++;
return;
}
else if (isSameEntry(file, slotEntry, pos, len)) {
keyValues[iSl + 1] = doubleToRawLongBits(Math.min(longBitsToDouble(keyValues[iSl + 1]), val));
keyValues[iSl + 2] = doubleToRawLongBits(Math.max(longBitsToDouble(keyValues[iSl + 2]), val));
keyValues[iSl + 3] = doubleToRawLongBits(longBitsToDouble(keyValues[iSl + 3]) + val);
keyValues[iSl + 4] = keyValues[iSl + 4] + 1;
return;
}
else {
// long keyPos = (slotEntry & MASK_POS) >> SHIFT_POS;
// int keyLen = (int) (slotEntry & MASK_LEN);
// System.out.println("Colliding " + new String(file.asSlice(pos,len).toArray(ValueLayout.JAVA_BYTE)) +
// " with key" + new String(file.asSlice(keyPos,keyLen).toArray(ValueLayout.JAVA_BYTE)) +
// " hash " + hash + " slot " + slot + "+" + probe + " at " + iSl);
// debug_reprobeMax = Math.max(debug_reprobeMax, probe);
}
}
throw new IllegalStateException("More than 20000 reprobes");
// throw new IllegalStateException("More than 100 reprobes: At " + debug_size + "");
}
private boolean isSameEntry(MemorySegment file, long slotEntry, long pos, int len) {
long keyPos = (slotEntry & MASK_POS) >> SHIFT_POS;
int keyLen = (int) (slotEntry & MASK_LEN);
var isSame = isSame(file, keyPos, pos, len);
// System.out.println("Entry:" + new String(file.asSlice(pos, len).toArray(JAVA_BYTE)) +
// ",Keys:" + new String(file.asSlice(keyPos, keyLen).toArray(JAVA_BYTE)) +
// ",match " + isSame);
return isSame;
}
private static boolean isSame(MemorySegment file, long i1, long i2, int len) {
int i = 0;
var i1len = i1 + vecLen;
var i2len = i2 + vecLen;
if (len < vecLen && i1len <= file.byteSize() && i2len <= file.byteSize()) {
var v1 = byteVec.fromMemorySegment(file, i1, ByteOrder.BIG_ENDIAN);
var v2 = byteVec.fromMemorySegment(file, i2, ByteOrder.BIG_ENDIAN);
var isTrue = v1.compare(VectorOperators.EQ, v2, allTrue.indexInRange(0, len));
return isTrue.trueCount() == len;
}
while (8 < (len - i)) {
var v1 = file.get(JAVA_LONG_UNALIGNED, i1 + i);
var v2 = file.get(JAVA_LONG_UNALIGNED, i2 + i);
if (v1 != v2) {
return false;
}
i += 8;
}
while (i < len) {
var v1 = file.get(JAVA_BYTE, i1 + i);
var v2 = file.get(JAVA_BYTE, i2 + i);
if (v1 != v2) {
return false;
}
i++;
}
return true;
}
public void fill(MemorySegment file, TreeMap<String, ResultRow> treeMap) {
for (int i = 0; i < keyValues.length / 5; i++) {
var ji = i * 5;
long keyE = keyValues[ji];
if (keyE != 0) {
long keyPos = (keyE & MASK_POS) >> SHIFT_POS;
int keyLen = (int) (keyE & MASK_LEN);
byte[] keyBytes = new byte[keyLen];
MemorySegment.copy(file, JAVA_BYTE, keyPos, keyBytes, 0, keyLen);
var key = new String(keyBytes);
var min = longBitsToDouble(keyValues[ji + 1]);
var max = longBitsToDouble(keyValues[ji + 2]);
var sum = longBitsToDouble(keyValues[ji + 3]);
var count = keyValues[ji + 4];
treeMap.put(key, new ResultRow(min, sum / count, max));
}
}
}
}
private static record ResultRow(double min, double mean, double max) {
public String toString() {
return round(min) + "/" + round(mean) + "/" + round(max);
}
private double round(double value) {
return Math.round(value * 10.0) / 10.0;
}
}
;
}