1brc/src/main/java/dev/morling/onebrc/CalculateAverage_royvanrijn.java
Roy van Rijn 7c983f3d66
Added dedicated reader (#493)
Started running perf, perhaps this helps. No idea how to use it yet
2024-01-19 22:15:49 +01:00

575 lines
22 KiB
Java

/*
* 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 java.lang.foreign.Arena;
import java.lang.reflect.Field;
import java.nio.channels.FileChannel;
import java.nio.charset.StandardCharsets;
import java.nio.file.Path;
import java.nio.file.StandardOpenOption;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
import java.util.stream.Collectors;
import sun.misc.Unsafe;
/**
* Changelog:
*
* Initial submission: 62000 ms
* Chunked reader: 16000 ms
* Optimized parser: 13000 ms
* Branchless methods: 11000 ms
* Adding memory mapped files: 6500 ms (based on bjhara's submission)
* Skipping string creation: 4700 ms
* Custom hashmap... 4200 ms
* Added SWAR token checks: 3900 ms
* Skipped String creation: 3500 ms (idea from kgonia)
* Improved String skip: 3250 ms
* Segmenting files: 3150 ms (based on spullara's code)
* Not using SWAR for EOL: 2850 ms
* Inlining hash calculation: 2450 ms
* Replacing branchless code: 2200 ms (sometimes we need to kill the things we love)
* Added unsafe memory access: 1900 ms (keeping the long[] small and local)
* Fixed bug, UNSAFE bytes String: 1850 ms
* Separate hash from entries: 1550 ms
* Various tweaks for Linux/cache 1550 ms (should/could make a difference on target machine)
* Improved layout/predictability: 1400 ms
* Delayed String creation again: 1350 ms
* Remove writing to buffer: 1335 ms
* Optimized collecting at the end: 1310 ms
* Adding a lot of comments: priceless
* Changed to flyweight byte[]: 1290 ms (adds even more Unsafe, was initially slower, now faster)
* More LOC now parallel: 1260 ms (moved more to processMemoryArea, recombining in ConcurrentHashMap)
* Storing only the address: 1240 ms (this is now faster, tried before, was slower)
* Unrolling scan-loop: 1200 ms (seems to help, perhaps even more on target machine)
* Adding more readable reader: 1300 ms (scores got worse on target machine anyway)
*
* I've ditched my M2 for an older x86-64 MacBook, this allows me to run `perf` and I'm trying to get lower numbers by trail and error.
*
* Big thanks to Francesco Nigro, Thomas Wuerthinger, Quan Anh Mai and many others for ideas.
*
* Follow me at: @royvanrijn
*/
public class CalculateAverage_royvanrijn {
private static final String FILE = "./measurements.txt";
// private static final String FILE = "src/test/resources/samples/measurements-1.txt";
private static final Unsafe UNSAFE = initUnsafe();
// Twice the processors, smoothens things out.
private static final int PROCESSORS = Runtime.getRuntime().availableProcessors();
/**
* Flyweight entry in a byte[], max 128 bytes.
*
* long: sum
* int: min
* int: max
* int: count
* byte: length
* byte[]: cityname
*/
// ------------------------------------------------------------------------
private static final int ENTRY_LENGTH = (Unsafe.ARRAY_BYTE_BASE_OFFSET);
private static final int ENTRY_SUM = (ENTRY_LENGTH + Byte.BYTES);
private static final int ENTRY_MIN = (ENTRY_SUM + Long.BYTES);
private static final int ENTRY_MAX = (ENTRY_MIN + Integer.BYTES);
private static final int ENTRY_COUNT = (ENTRY_MAX + Integer.BYTES);
private static final int ENTRY_NAME = (ENTRY_COUNT + Integer.BYTES);
private static final int ENTRY_NAME_8 = ENTRY_NAME + 8;
private static final int ENTRY_NAME_16 = ENTRY_NAME + 16;
private static final int ENTRY_BASESIZE_WHITESPACE = ENTRY_NAME + 7; // with enough empty bytes to fill a long
// ------------------------------------------------------------------------
private static final int PREMADE_MAX_SIZE = 1 << 5; // pre-initialize some entries in memory, keep them close
private static final int PREMADE_ENTRIES = 512; // amount of pre-created entries we should use
private static final int TABLE_SIZE = 1 << 19; // large enough for the contest.
private static final int TABLE_MASK = (TABLE_SIZE - 1);
public static void main(String[] args) throws Exception {
// Calculate input segments.
final FileChannel fileChannel = FileChannel.open(Path.of(FILE), StandardOpenOption.READ);
final long fileSize = fileChannel.size();
final long segmentSize = (fileSize + PROCESSORS - 1) / PROCESSORS;
final long mapAddress = fileChannel.map(FileChannel.MapMode.READ_ONLY, 0, fileSize, Arena.global()).address();
final Thread[] parallelThreads = new Thread[PROCESSORS - 1];
// This is where the entries will land:
final ConcurrentHashMap<String, byte[]> measurements = new ConcurrentHashMap(1 << 10);
// We create separate threads for twice the amount of processors.
long lastAddress = mapAddress;
final long endOfFile = mapAddress + fileSize;
for (int i = 0; i < PROCESSORS - 1; ++i) {
final long fromAddress = lastAddress;
final long toAddress = Math.min(endOfFile, fromAddress + segmentSize);
final Thread thread = new Thread(() -> {
// The actual work is done here:
final byte[][] table = processMemoryArea(fromAddress, toAddress, fromAddress == mapAddress);
for (byte[] entry : table) {
if (entry != null) {
measurements.merge(entryToName(entry), entry, CalculateAverage_royvanrijn::mergeEntry);
}
}
});
thread.start(); // start a.s.a.p.
parallelThreads[i] = thread;
lastAddress = toAddress;
}
// Use the current thread for the part of memory:
final byte[][] table = processMemoryArea(lastAddress, mapAddress + fileSize, false);
for (byte[] entry : table) {
if (entry != null) {
measurements.merge(entryToName(entry), entry, CalculateAverage_royvanrijn::mergeEntry);
}
}
// Wait for all threads to finish:
for (Thread thread : parallelThreads) {
// Can we implement work-stealing? Not sure how...
thread.join();
}
// If we don't reach start of file,
System.out.print("{" +
measurements.entrySet().stream().sorted(Map.Entry.comparingByKey())
.map(entry -> entry.getKey() + '=' + entryValuesToString(entry.getValue()))
.collect(Collectors.joining(", ")));
System.out.println("}");
// System.out.println(measurements.entrySet().stream().mapToLong(e -> UNSAFE.getInt(e.getValue(), ENTRY_COUNT + Unsafe.ARRAY_BYTE_BASE_OFFSET)).sum());
}
private static byte[] fillEntry(final byte[] entry, final long fromAddress, final int length, final int temp) {
UNSAFE.putLong(entry, ENTRY_SUM, temp);
UNSAFE.putInt(entry, ENTRY_MIN, temp);
UNSAFE.putInt(entry, ENTRY_MAX, temp);
UNSAFE.putInt(entry, ENTRY_COUNT, 1);
UNSAFE.putByte(entry, ENTRY_LENGTH, (byte) length);
UNSAFE.copyMemory(null, fromAddress, entry, ENTRY_NAME, length);
return entry;
}
public static void updateEntry(final byte[] entry, final int temp) {
int entryMin = UNSAFE.getInt(entry, ENTRY_MIN);
int entryMax = UNSAFE.getInt(entry, ENTRY_MAX);
entryMin = Math.min(temp, entryMin);
entryMax = Math.max(temp, entryMax);
long entrySum = UNSAFE.getLong(entry, ENTRY_SUM) + temp;
int entryCount = UNSAFE.getInt(entry, ENTRY_COUNT) + 1;
UNSAFE.putInt(entry, ENTRY_MIN, entryMin);
UNSAFE.putInt(entry, ENTRY_MAX, entryMax);
UNSAFE.putInt(entry, ENTRY_COUNT, entryCount);
UNSAFE.putLong(entry, ENTRY_SUM, entrySum);
}
public static byte[] mergeEntry(final byte[] entry, final byte[] merge) {
long sum = UNSAFE.getLong(merge, ENTRY_SUM);
final int mergeMin = UNSAFE.getInt(merge, ENTRY_MIN);
final int mergeMax = UNSAFE.getInt(merge, ENTRY_MAX);
int count = UNSAFE.getInt(merge, ENTRY_COUNT);
sum += UNSAFE.getLong(entry, ENTRY_SUM);
int entryMin = UNSAFE.getInt(entry, ENTRY_MIN);
int entryMax = UNSAFE.getInt(entry, ENTRY_MAX);
count += UNSAFE.getInt(entry, ENTRY_COUNT);
entryMin = Math.min(entryMin, mergeMin);
entryMax = Math.max(entryMax, mergeMax);
UNSAFE.putLong(entry, ENTRY_SUM, sum);
UNSAFE.putInt(entry, ENTRY_MIN, entryMin);
UNSAFE.putInt(entry, ENTRY_MAX, entryMax);
UNSAFE.putInt(entry, ENTRY_COUNT, count);
return entry;
}
private static String entryToName(final byte[] entry) {
// Get the length from memory:
int length = UNSAFE.getByte(entry, ENTRY_LENGTH);
byte[] name = new byte[length];
UNSAFE.copyMemory(entry, ENTRY_NAME, name, Unsafe.ARRAY_BYTE_BASE_OFFSET, length);
// Create a new String with the existing byte[]:
return new String(name, StandardCharsets.UTF_8);
}
private static String entryValuesToString(final byte[] entry) {
return round(UNSAFE.getInt(entry, ENTRY_MIN))
+ "/" +
round((1.0 * UNSAFE.getLong(entry, ENTRY_SUM)) /
UNSAFE.getInt(entry, ENTRY_COUNT))
+ "/" +
round(UNSAFE.getInt(entry, ENTRY_MAX));
}
// Print a piece of memory:
// For debug.
private static String printMemory(final Object target, final long address, int length) {
String result = "";
for (int i = 0; i < length; i++) {
result += (char) UNSAFE.getByte(target, address + i);
}
return result;
}
// Print a piece of memory:
// For debug.
private static String printMemory(final long value, int length) {
String result = "";
for (int i = 0; i < length; i++) {
result += (char) ((value >> (i << 3)) & 0xFF);
}
return result;
}
private static double round(final double value) {
return Math.round(value) / 10.0;
}
private static final class Reader {
private long ptr;
private long delimiterMask;
private long lastRead;
private long lastReadMinOne;
private long hash;
private long entryStart;
private long entryDelimiter;
private final long endAddress;
Reader(final long startAddress, final long endAddress, final boolean isFileStart) {
this.ptr = startAddress;
this.endAddress = endAddress;
// Adjust start to next delimiter:
if (!isFileStart) {
ptr--;
while (ptr < endAddress) {
if (UNSAFE.getByte(ptr++) == '\n') {
break;
}
}
}
}
private void processStart() {
hash = 0;
entryStart = ptr;
}
private boolean hasNext() {
return (ptr < endAddress);
}
private static final long DELIMITER_MASK = 0x3B3B3B3B3B3B3B3BL;
private boolean readFirst() {
lastRead = UNSAFE.getLong(ptr);
final long match = lastRead ^ DELIMITER_MASK;
delimiterMask = (match - 0x0101010101010101L) & (~match & 0x8080808080808080L);
return delimiterMask == 0;
}
private boolean readNext() {
lastReadMinOne = lastRead;
return readFirst();
}
private void processName() {
hash ^= lastRead;
ptr += 8;
}
private int processEndAndGetTemperature() {
processFinalBytes();
finalizeHash();
finalizeDelimiter();
return readTemperature();
}
private void processFinalBytes() {
// Shift and read the last bytes:
lastRead &= ((delimiterMask >>> 7) - 1);
}
private void finalizeHash() {
// Finalize hash:
hash ^= lastRead;
hash ^= hash >> 32;
hash ^= hash >> 17; // extra entropy
}
private void finalizeDelimiter() {
// Found delimiter:
entryDelimiter = ptr + (Long.numberOfTrailingZeros(delimiterMask) >> 3);
}
private static final long DOT_BITS = 0x10101000;
private static final long MAGIC_MULTIPLIER = (100 * 0x1000000 + 10 * 0x10000 + 1);
// Awesome idea of merykitty:
private int readTemperature() {
// This is the number part: X.X, -X.X, XX.x or -XX.X
long numberBytes = UNSAFE.getLong(entryDelimiter + 1);
long invNumberBytes = ~numberBytes;
int dotPosition = Long.numberOfTrailingZeros(invNumberBytes & DOT_BITS);
// Update the pointer here, bit awkward, but we have all the data
ptr = entryDelimiter + (dotPosition >> 3) + 4;
int min28 = (28 - dotPosition);
// Calculates the sign
final long signed = (invNumberBytes << 59) >> 63;
final long minusFilter = ~(signed & 0xFF);
// Use the pre-calculated decimal position to adjust the values
long digits = ((numberBytes & minusFilter) << min28) & 0x0F000F0F00L;
// Multiply by a magic (100 * 0x1000000 + 10 * 0x10000 + 1), to get the result
final long absValue = ((digits * MAGIC_MULTIPLIER) >>> 32) & 0x3FF;
// And perform abs()
return (int) ((absValue + signed) ^ signed); // non-patented method of doing the same trick
}
private boolean matchesEntryFull(final byte[] entry) {
int longs = (int) (entryDelimiter - entryStart) >> 3;
int step = 0;
for (int i = 0; i < longs - 2; i++) {
if (UNSAFE.getLong(entryStart + step) != UNSAFE.getLong(entry, ENTRY_NAME + step)) {
return false;
}
step += 8;
}
if (lastReadMinOne != UNSAFE.getLong(entry, (ENTRY_NAME_8) + step)) {
return false;
}
if (lastRead != UNSAFE.getLong(entry, (ENTRY_NAME_16) + step)) {
return false;
}
return true;
}
private boolean matchesEntryMedium(final byte[] entry) {
if (UNSAFE.getLong(entryStart) != UNSAFE.getLong(entry, ENTRY_NAME)) {
return false;
}
if (lastReadMinOne != UNSAFE.getLong(entry, ENTRY_NAME_8)) {
return false;
}
if (lastRead != UNSAFE.getLong(entry, ENTRY_NAME_16)) {
return false;
}
return true;
}
private boolean matchesEntryShort(final byte[] entry) {
if (lastReadMinOne != UNSAFE.getLong(entry, ENTRY_NAME)) {
return false;
}
if (lastRead != UNSAFE.getLong(entry, ENTRY_NAME_8)) {
return false;
}
return true;
}
private boolean matchesEnding(final byte[] entry) {
return lastRead == UNSAFE.getLong(entry, ENTRY_NAME);
}
private int length() {
return (int) (entryDelimiter - entryStart);
}
}
private static byte[][] processMemoryArea(final long startAddress, final long endAddress, boolean isFileStart) {
final byte[][] table = new byte[TABLE_SIZE][];
final byte[][] preConstructedEntries = new byte[PREMADE_ENTRIES][ENTRY_BASESIZE_WHITESPACE + PREMADE_MAX_SIZE];
final Reader reader = new Reader(startAddress, endAddress, isFileStart);
byte[] entry;
int entryCount = 0;
// Find the correct starting position
while (reader.hasNext()) {
reader.processStart();
if (!reader.readFirst()) {
int temperature = reader.processEndAndGetTemperature();
// Find or insert the entry:
int index = (int) (reader.hash & TABLE_MASK);
while (true) {
entry = table[index];
if (entry == null) {
int length = reader.length();
byte[] entryBytes = (entryCount < PREMADE_ENTRIES) ? preConstructedEntries[entryCount++]
: new byte[ENTRY_BASESIZE_WHITESPACE + length];
table[index] = fillEntry(entryBytes, reader.entryStart, length, temperature);
break;
}
else if (reader.matchesEnding(entry)) {
updateEntry(entry, temperature);
break;
}
else {
// Move to the next index
index = (index + 1) & TABLE_MASK;
}
}
}
else {
reader.processName();
if (!reader.readNext()) {
int temperature = reader.processEndAndGetTemperature();
// Find or insert the entry:
int index = (int) (reader.hash & TABLE_MASK);
while (true) {
entry = table[index];
if (entry == null) {
int length = reader.length();
byte[] entryBytes = (entryCount < PREMADE_ENTRIES) ? preConstructedEntries[entryCount++]
: new byte[ENTRY_BASESIZE_WHITESPACE + length];
table[index] = fillEntry(entryBytes, reader.entryStart, length, temperature);
break;
}
else if (reader.matchesEntryShort(entry)) {
updateEntry(entry, temperature);
break;
}
else {
// Move to the next index
index = (index + 1) & TABLE_MASK;
}
}
}
else {
reader.processName();
if (!reader.readNext()) {
int temperature = reader.processEndAndGetTemperature();
// Find or insert the entry:
int index = (int) (reader.hash & TABLE_MASK);
while (true) {
entry = table[index];
if (entry == null) {
int length = reader.length();
byte[] entryBytes = (entryCount < PREMADE_ENTRIES) ? preConstructedEntries[entryCount++]
: new byte[ENTRY_BASESIZE_WHITESPACE + length];
table[index] = fillEntry(entryBytes, reader.entryStart, length, temperature);
break;
}
else if (reader.matchesEntryMedium(entry)) {
updateEntry(entry, temperature);
break;
}
else {
// Move to the next index
index = (index + 1) & TABLE_MASK;
}
}
}
else {
reader.processName();
while (reader.readNext()) {
reader.processName();
}
int temperature = reader.processEndAndGetTemperature();
// Find or insert the entry:
int index = (int) (reader.hash & TABLE_MASK);
while (true) {
entry = table[index];
if (entry == null) {
int length = reader.length();
byte[] entryBytes = (length < PREMADE_MAX_SIZE && entryCount < PREMADE_ENTRIES) ? preConstructedEntries[entryCount++]
: new byte[ENTRY_BASESIZE_WHITESPACE + length]; // with enough room
table[index] = fillEntry(entryBytes, reader.entryStart, length, temperature);
break;
}
else if (reader.matchesEntryFull(entry)) {
updateEntry(entry, temperature);
break;
}
else {
// Move to the next index
index = (index + 1) & TABLE_MASK;
}
}
}
}
}
}
return table;
}
/*
* `___` ___ ___ _ ___` ` ___ ` _ ` _ ` _` ___
* / ` \| _ \ __| \| \ \ / /_\ | | | | | | __|
* | () | _ / __|| . |\ V / _ \| |_| |_| | ._|
* \___/|_| |___|_|\_| \_/_/ \_\___|\___/|___|
* ---------------- BETTER SOFTWARE, FASTER --
*
* https://www.openvalue.eu/
*/
private static Unsafe initUnsafe() {
try {
final Field theUnsafe = Unsafe.class.getDeclaredField("theUnsafe");
theUnsafe.setAccessible(true);
return (Unsafe) theUnsafe.get(Unsafe.class);
}
catch (NoSuchFieldException | IllegalAccessException e) {
throw new RuntimeException(e);
}
}
}