Playground/1brc
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Some fine tuning for thomaswue (#606)

Browse Source 
* Some fine tuning.

* Process 2MB segments to make all threads finish at the same time.
Process with 3 scanners in parallel in the same thread.
This commit is contained in:
Thomas Wuerthinger 2024-01-28 17:59:57 +01:00 committed by GitHub
parent c8dd691a27
commit 7e525c5992
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 247 additions and 160 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
prepare_thomaswue.sh
View File

@ -20,7 +20,7 @@ sdk use java 21.0.2-graal 1>&2
# ./mvnw clean verify removes target/ and will re-trigger native image creation. # ./mvnw clean verify removes target/ and will re-trigger native image creation.
if [ ! -f target/CalculateAverage_thomaswue_image ]; then if [ ! -f target/CalculateAverage_thomaswue_image ]; then
NATIVE_IMAGE_OPTS="--gc=epsilon -O3 -march=native --enable-preview -H:InlineAllBonus=10 -H:-ParseRuntimeOptions --initialize-at-build-time=dev.morling.onebrc.CalculateAverage_thomaswue\$Scanner" NATIVE_IMAGE_OPTS="--gc=epsilon -O3 -H:-GenLoopSafepoints -march=native --enable-preview -H:InlineAllBonus=10 -H:-ParseRuntimeOptions --initialize-at-build-time=dev.morling.onebrc.CalculateAverage_thomaswue\$Scanner"
# Use -H:MethodFilter=CalculateAverage_thomaswue.* -H:Dump=:2 -H:PrintGraph=Network for IdealGraphVisualizer graph dumping. # Use -H:MethodFilter=CalculateAverage_thomaswue.* -H:Dump=:2 -H:PrintGraph=Network for IdealGraphVisualizer graph dumping.
native-image $NATIVE_IMAGE_OPTS -cp target/average-1.0.0-SNAPSHOT.jar -o target/CalculateAverage_thomaswue_image dev.morling.onebrc.CalculateAverage_thomaswue native-image $NATIVE_IMAGE_OPTS -cp target/average-1.0.0-SNAPSHOT.jar -o target/CalculateAverage_thomaswue_image dev.morling.onebrc.CalculateAverage_thomaswue
fi fi

259
src/main/java/dev/morling/onebrc/CalculateAverage_thomaswue.java
View File

@ -23,16 +23,17 @@ import java.nio.charset.StandardCharsets;
import java.nio.file.Path; import java.nio.file.Path;
import java.nio.file.StandardOpenOption; import java.nio.file.StandardOpenOption;
import java.util.*; import java.util.*;
import java.util.concurrent.atomic.AtomicLong;
import java.util.stream.IntStream; import java.util.stream.IntStream;
/** /**
* Simple solution that memory maps the input file, then splits it into one segment per available core and uses * Simple solution that memory maps the input file, then splits it into one segment per available core and uses
* sun.misc.Unsafe to directly access the mapped memory. Uses a long at a time when checking for collision. * sun.misc.Unsafe to directly access the mapped memory. Uses a long at a time when checking for collision.
* <p> * <p>
* Runs in 0.60s on my Intel i9-13900K * Runs in 0.41s on my Intel i9-13900K
* Perf stats: * Perf stats:
* 34,716,719,245 cpu_core/cycles/ * 25,286,227,376 cpu_core/cycles/
* 40,776,530,892 cpu_atom/cycles/ * 26,833,723,225 cpu_atom/cycles/
*/ */
public class CalculateAverage_thomaswue { public class CalculateAverage_thomaswue {
private static final String FILE = "./measurements.txt"; private static final String FILE = "./measurements.txt";
@ -42,10 +43,11 @@ public class CalculateAverage_thomaswue {
// Holding the current result for a single city. // Holding the current result for a single city.
private static class Result { private static class Result {
long lastNameLong, secondLastNameLong; long lastNameLong, secondLastNameLong;
long[] name; long min, max;
int count;
short min, max;
long sum; long sum;
int count;
long[] name;
String nameAsString;
private Result() { private Result() {
this.min = MAX_TEMP; this.min = MAX_TEMP;
@ -73,36 +75,59 @@ public class CalculateAverage_thomaswue {
} }
public String calcName() { public String calcName() {
if (nameAsString == null) {
ByteBuffer bb = ByteBuffer.allocate(name.length * Long.BYTES).order(ByteOrder.nativeOrder()); ByteBuffer bb = ByteBuffer.allocate(name.length * Long.BYTES).order(ByteOrder.nativeOrder());
bb.asLongBuffer().put(name); bb.asLongBuffer().put(name);
byte[] array = bb.array(); byte[] array = bb.array();
int i = 0; int i = 0;
while (array[i++] != ';') while (array[i++] != ';')
; ;
return new String(array, 0, i - 1, StandardCharsets.UTF_8); nameAsString = new String(array, 0, i - 1, StandardCharsets.UTF_8);
}
return nameAsString;
} }
} }
public static void main(String[] args) throws IOException { public static void main(String[] args) throws IOException, InterruptedException {
if (args.length == 0 || !("--worker".equals(args[0]))) { if (args.length == 0 || !("--worker".equals(args[0]))) {
spawnWorker(); spawnWorker();
return; return;
} }
// Calculate input segments. // Calculate input segments.
int numberOfChunks = Runtime.getRuntime().availableProcessors(); int numberOfWorkers = Runtime.getRuntime().availableProcessors();
long[] chunks = getSegments(numberOfChunks); final AtomicLong cursor = new AtomicLong();
final long fileEnd;
final long fileStart;
try (var fileChannel = FileChannel.open(Path.of(FILE), StandardOpenOption.READ)) {
long fileSize = fileChannel.size();
fileStart = fileChannel.map(FileChannel.MapMode.READ_ONLY, 0, fileSize, java.lang.foreign.Arena.global()).address();
cursor.set(fileStart);
fileEnd = fileStart + fileSize;
}
// Parallel processing of segments. // Parallel processing of segments.
List<List<Result>> allResults = IntStream.range(0, chunks.length - 1).mapToObj(chunkIndex -> parseLoop(chunks[chunkIndex], chunks[chunkIndex + 1])) Thread[] threads = new Thread[numberOfWorkers];
.map(resultArray -> { List<Result>[] allResults = new List[numberOfWorkers];
List<Result> results = new ArrayList<>(); for (int i = 0; i < threads.length; ++i) {
final int index = i;
threads[i] = new Thread(() -> {
Result[] resultArray = parseLoop(cursor, fileEnd, fileStart);
List<Result> results = new ArrayList<>(500);
for (Result r : resultArray) { for (Result r : resultArray) {
if (r != null) { if (r != null) {
r.calcName();
results.add(r); results.add(r);
} }
} }
return results; allResults[index] = results;
}).parallel().toList(); });
threads[i].start();
}
for (Thread thread : threads) {
thread.join();
}
// Final output. // Final output.
System.out.println(accumulateResults(allResults)); System.out.println(accumulateResults(allResults));
@ -115,17 +140,12 @@ public class CalculateAverage_thomaswue {
info.command().ifPresent(workerCommand::add); info.command().ifPresent(workerCommand::add);
info.arguments().ifPresent(args -> workerCommand.addAll(Arrays.asList(args))); info.arguments().ifPresent(args -> workerCommand.addAll(Arrays.asList(args)));
workerCommand.add("--worker"); workerCommand.add("--worker");
new ProcessBuilder() new ProcessBuilder().command(workerCommand).inheritIO().redirectOutput(ProcessBuilder.Redirect.PIPE)
.command(workerCommand) .start().getInputStream().transferTo(System.out);
.inheritIO()
.redirectOutput(ProcessBuilder.Redirect.PIPE)
.start()
.getInputStream()
.transferTo(System.out);
} }
// Accumulate results sequentially for simplicity. // Accumulate results sequentially for simplicity.
private static TreeMap<String, Result> accumulateResults(List<List<Result>> allResults) { private static TreeMap<String, Result> accumulateResults(List<Result>[] allResults) {
TreeMap<String, Result> result = new TreeMap<>(); TreeMap<String, Result> result = new TreeMap<>();
for (List<Result> resultArr : allResults) { for (List<Result> resultArr : allResults) {
for (Result r : resultArr) { for (Result r : resultArr) {
@ -139,15 +159,13 @@ public class CalculateAverage_thomaswue {
return result; return result;
} }
// Main parse loop. private static Result findResult(long initialWord, long initialPos, Scanner scanner, Result[] results) {
private static Result[] parseLoop(long chunkStart, long chunkEnd) {
Result[] results = new Result[1 << 17]; Result existingResult;
Scanner scanner = new Scanner(chunkStart, chunkEnd); long word = initialWord;
long word = scanner.getLong(); long pos = initialPos;
long pos = findDelimiter(word); long hash;
while (scanner.hasNext()) {
long nameAddress = scanner.pos(); long nameAddress = scanner.pos();
long hash = 0;
// Search for ';', one long at a time. // Search for ';', one long at a time.
if (pos != 0) { if (pos != 0) {
@ -156,20 +174,16 @@ public class CalculateAverage_thomaswue {
word = mask(word, pos); word = mask(word, pos);
hash = word; hash = word;
int number = scanNumber(scanner); int index = hashToIndex(hash, results);
long nextWord = scanner.getLong(); existingResult = results[index];
long nextPos = findDelimiter(nextWord);
Result existingResult = results[hashToIndex(hash, results)];
if (existingResult != null && existingResult.lastNameLong == word) { if (existingResult != null && existingResult.lastNameLong == word) {
word = nextWord; return existingResult;
pos = nextPos;
record(existingResult, number);
continue;
} }
else {
scanner.setPos(nameAddress + pos); scanner.setPos(nameAddress + pos);
} }
}
else { else {
scanner.add(8); scanner.add(8);
hash = word; hash = word;
@ -181,14 +195,14 @@ public class CalculateAverage_thomaswue {
scanner.add(pos); scanner.add(pos);
word = mask(word, pos); word = mask(word, pos);
hash ^= word; hash ^= word;
int index = hashToIndex(hash, results);
existingResult = results[index];
Result existingResult = results[hashToIndex(hash, results)];
if (existingResult != null && existingResult.lastNameLong == word && existingResult.secondLastNameLong == prevWord) { if (existingResult != null && existingResult.lastNameLong == word && existingResult.secondLastNameLong == prevWord) {
int number = scanNumber(scanner); return existingResult;
word = scanner.getLong(); }
pos = findDelimiter(word); else {
record(existingResult, number); scanner.setPos(nameAddress + pos + 8);
continue;
} }
} }
else { else {
@ -214,20 +228,19 @@ public class CalculateAverage_thomaswue {
// Save length of name for later. // Save length of name for later.
int nameLength = (int) (scanner.pos() - nameAddress); int nameLength = (int) (scanner.pos() - nameAddress);
int number = scanNumber(scanner);
// Final calculation for index into hash table. // Final calculation for index into hash table.
int tableIndex = hashToIndex(hash, results); int tableIndex = hashToIndex(hash, results);
outer: while (true) { outer: while (true) {
Result existingResult = results[tableIndex]; existingResult = results[tableIndex];
if (existingResult == null) { if (existingResult == null) {
existingResult = newEntry(results, nameAddress, tableIndex, nameLength, scanner); existingResult = newEntry(results, nameAddress, tableIndex, nameLength, scanner);
} }
// Check for collision. // Check for collision.
int i = 0; int i = 0;
int namePos = 0; long[] name = existingResult.name;
for (; i < nameLength + 1 - 8; i += 8) { for (; i < nameLength + 1 - 8; i += 8) {
if (namePos >= existingResult.name.length || existingResult.name[namePos++] != scanner.getLongAt(nameAddress + i)) { if (scanner.getLongAt(i, name) != scanner.getLongAt(nameAddress + i)) {
tableIndex = (tableIndex + 31) & (results.length - 1); tableIndex = (tableIndex + 31) & (results.length - 1);
continue outer; continue outer;
} }
@ -235,7 +248,6 @@ public class CalculateAverage_thomaswue {
int remainingShift = (64 - (nameLength + 1 - i) << 3); int remainingShift = (64 - (nameLength + 1 - i) << 3);
if (((existingResult.lastNameLong ^ (scanner.getLongAt(nameAddress + i) << remainingShift)) == 0)) { if (((existingResult.lastNameLong ^ (scanner.getLongAt(nameAddress + i) << remainingShift)) == 0)) {
record(existingResult, number);
break; break;
} }
else { else {
@ -243,37 +255,124 @@ public class CalculateAverage_thomaswue {
tableIndex = (tableIndex + 31) & (results.length - 1); tableIndex = (tableIndex + 31) & (results.length - 1);
} }
} }
return existingResult;
word = scanner.getLong();
pos = findDelimiter(word);
} }
private static long nextNL(long prev) {
while (true) {
long currentWord = Scanner.UNSAFE.getLong(prev);
long pos = findNewLine(currentWord);
if (pos != 0) {
prev += Long.numberOfTrailingZeros(pos) >>> 3;
break;
}
else {
prev += 8;
}
}
return prev;
}
private static final int SEGMENT_SIZE = 1024 * 1024 * 2;
// Main parse loop.
private static Result[] parseLoop(AtomicLong counter, long fileEnd, long fileStart) {
Result[] results = new Result[1 << 17];
while (true) {
long current = counter.addAndGet(SEGMENT_SIZE) - SEGMENT_SIZE;
if (current >= fileEnd) {
return results; return results;
} }
private static int scanNumber(Scanner scanPtr) { long segmentEnd = nextNL(Math.min(fileEnd - 1, current + SEGMENT_SIZE));
long segmentStart;
if (current == fileStart) {
segmentStart = current;
}
else {
segmentStart = nextNL(current) + 1;
}
long dist = (segmentEnd - segmentStart) / 3;
long midPoint1 = nextNL(segmentStart + dist);
long midPoint2 = nextNL(segmentStart + dist + dist);
Scanner scanner1 = new Scanner(segmentStart, midPoint1);
Scanner scanner2 = new Scanner(midPoint1 + 1, midPoint2);
Scanner scanner3 = new Scanner(midPoint2 + 1, segmentEnd);
while (true) {
if (!scanner1.hasNext()) {
break;
}
if (!scanner2.hasNext()) {
break;
}
if (!scanner3.hasNext()) {
break;
}
long word1 = scanner1.getLong();
long word2 = scanner2.getLong();
long word3 = scanner3.getLong();
long pos1 = findDelimiter(word1);
long pos2 = findDelimiter(word2);
long pos3 = findDelimiter(word3);
Result existingResult1 = findResult(word1, pos1, scanner1, results);
Result existingResult2 = findResult(word2, pos2, scanner2, results);
Result existingResult3 = findResult(word3, pos3, scanner3, results);
long number1 = scanNumber(scanner1);
long number2 = scanNumber(scanner2);
long number3 = scanNumber(scanner3);
record(existingResult1, number1);
record(existingResult2, number2);
record(existingResult3, number3);
}
while (scanner1.hasNext()) {
long word = scanner1.getLong();
long pos = findDelimiter(word);
record(findResult(word, pos, scanner1, results), scanNumber(scanner1));
}
while (scanner2.hasNext()) {
long word = scanner2.getLong();
long pos = findDelimiter(word);
record(findResult(word, pos, scanner2, results), scanNumber(scanner2));
}
while (scanner3.hasNext()) {
long word = scanner3.getLong();
long pos = findDelimiter(word);
record(findResult(word, pos, scanner3, results), scanNumber(scanner3));
}
}
}
private static long scanNumber(Scanner scanPtr) {
scanPtr.add(1); scanPtr.add(1);
long numberWord = scanPtr.getLong(); long numberWord = scanPtr.getLong();
int decimalSepPos = Long.numberOfTrailingZeros(~numberWord & 0x10101000); int decimalSepPos = Long.numberOfTrailingZeros(~numberWord & 0x10101000);
int number = convertIntoNumber(decimalSepPos, numberWord); long number = convertIntoNumber(decimalSepPos, numberWord);
scanPtr.add((decimalSepPos >>> 3) + 3); scanPtr.add((decimalSepPos >>> 3) + 3);
return number; return number;
} }
private static void record(Result existingResult, int number) { private static void record(Result existingResult, long number) {
if (number < existingResult.min) { if (number < existingResult.min) {
existingResult.min = (short) number; existingResult.min = number;
} }
if (number > existingResult.max) { if (number > existingResult.max) {
existingResult.max = (short) number; existingResult.max = number;
} }
existingResult.sum += number; existingResult.sum += number;
existingResult.count++; existingResult.count++;
} }
private static int hashToIndex(long hash, Result[] results) { private static int hashToIndex(long hash, Result[] results) {
int hashAsInt = (int) (hash ^ (hash >>> 28)); long hashAsInt = hash ^ (hash >>> 37) ^ (hash >>> 17);
int finalHash = (hashAsInt ^ (hashAsInt >>> 17)); return (int) (hashAsInt & (results.length - 1));
return (finalHash & (results.length - 1));
} }
private static long mask(long word, long pos) { private static long mask(long word, long pos) {
@ -281,7 +380,7 @@ public class CalculateAverage_thomaswue {
} }
// Special method to convert a number in the ascii number into an int without branches created by Quan Anh Mai. // Special method to convert a number in the ascii number into an int without branches created by Quan Anh Mai.
private static int convertIntoNumber(int decimalSepPos, long numberWord) { private static long convertIntoNumber(int decimalSepPos, long numberWord) {
int shift = 28 - decimalSepPos; int shift = 28 - decimalSepPos;
// signed is -1 if negative, 0 otherwise // signed is -1 if negative, 0 otherwise
long signed = (~numberWord << 59) >> 63; long signed = (~numberWord << 59) >> 63;
@ -292,8 +391,7 @@ public class CalculateAverage_thomaswue {
// 0xUU00TTHH00 * (100 * 0x1000000 + 10 * 0x10000 + 1) = // 0xUU00TTHH00 * (100 * 0x1000000 + 10 * 0x10000 + 1) =
// 0x000000UU00TTHH00 + 0x00UU00TTHH000000 * 10 + 0xUU00TTHH00000000 * 100 // 0x000000UU00TTHH00 + 0x00UU00TTHH000000 * 10 + 0xUU00TTHH00000000 * 100
long absValue = ((digits * 0x640a0001) >>> 32) & 0x3FF; long absValue = ((digits * 0x640a0001) >>> 32) & 0x3FF;
long value = (absValue ^ signed) - signed; return (absValue ^ signed) - signed;
return (int) value;
} }
private static long findDelimiter(long word) { private static long findDelimiter(long word) {
@ -302,6 +400,12 @@ public class CalculateAverage_thomaswue {
return tmp; return tmp;
} }
private static long findNewLine(long word) {
long input = word ^ 0x0A0A0A0A0A0A0A0AL;
long tmp = (input - 0x0101010101010101L) & ~input & 0x8080808080808080L;
return tmp;
}
private static Result newEntry(Result[] results, long nameAddress, int hash, int nameLength, Scanner scanner) { private static Result newEntry(Result[] results, long nameAddress, int hash, int nameLength, Scanner scanner) {
Result r = new Result(); Result r = new Result();
results[hash] = r; results[hash] = r;
@ -324,27 +428,6 @@ public class CalculateAverage_thomaswue {
return r; return r;
} }
private static long[] getSegments(int numberOfChunks) throws IOException {
try (var fileChannel = FileChannel.open(Path.of(FILE), StandardOpenOption.READ)) {
long fileSize = fileChannel.size();
long segmentSize = (fileSize + numberOfChunks - 1) / numberOfChunks;
long[] chunks = new long[numberOfChunks + 1];
long mappedAddress = fileChannel.map(FileChannel.MapMode.READ_ONLY, 0, fileSize, java.lang.foreign.Arena.global()).address();
chunks[0] = mappedAddress;
long endAddress = mappedAddress + fileSize;
Scanner s = new Scanner(mappedAddress, mappedAddress + fileSize);
for (int i = 1; i < numberOfChunks; ++i) {
long chunkAddress = mappedAddress + i * segmentSize;
// Align to first row start.
while (chunkAddress < endAddress && (s.getLongAt(chunkAddress++) & 0xFF) != '\n')
;
chunks[i] = Math.min(chunkAddress, endAddress);
}
chunks[numberOfChunks] = endAddress;
return chunks;
}
}
private static class Scanner { private static class Scanner {
private static final sun.misc.Unsafe UNSAFE = initUnsafe(); private static final sun.misc.Unsafe UNSAFE = initUnsafe();
@ -387,6 +470,10 @@ public class CalculateAverage_thomaswue {
return UNSAFE.getLong(pos); return UNSAFE.getLong(pos);
} }
long getLongAt(long pos, long[] array) {
return UNSAFE.getLong(array, pos + sun.misc.Unsafe.ARRAY_LONG_BASE_OFFSET);
}
void setPos(long l) { void setPos(long l) {
this.pos = l; this.pos = l;
} }