HADOOP-14424. Add CRC32C performance test. Contributed by LiXin Ge.

(cherry picked from commit 75043d3429)
This commit is contained in:
Masatake Iwasaki 2017-06-19 09:16:45 +09:00
parent 3106f256e2
commit 023dcdc73d
1 changed files with 169 additions and 82 deletions

View File

@ -33,7 +33,7 @@ import org.apache.hadoop.fs.ChecksumException;
import org.apache.log4j.Level;
/**
* Performance tests to compare performance of Crc32 implementations
* Performance tests to compare performance of Crc32|Crc32C implementations
* This can be run from the command line with:
*
* java -cp path/to/test/classes:path/to/common/classes \
@ -43,17 +43,21 @@ import org.apache.log4j.Level;
*
* hadoop org.apache.hadoop.util.Crc32PerformanceTest
*
* If any argument is provided, this test will run with non-directly buffer.
*
* The output is in JIRA table format.
*/
public class Crc32PerformanceTest {
static final int MB = 1024 * 1024;
static interface Crc32 {
interface Crc32 {
public void verifyChunked(ByteBuffer data, int bytesPerCrc, ByteBuffer crcs,
void verifyChunked(ByteBuffer data, int bytesPerCrc, ByteBuffer crcs,
String filename, long basePos) throws ChecksumException;
static final class Native implements Crc32 {
DataChecksum.Type crcType();
final class Native implements Crc32 {
@Override
public void verifyChunked(ByteBuffer data, int bytesPerSum,
ByteBuffer sums, String fileName, long basePos)
@ -61,40 +65,91 @@ public class Crc32PerformanceTest {
NativeCrc32.verifyChunkedSums(bytesPerSum, DataChecksum.Type.CRC32.id,
sums, data, fileName, basePos);
}
@Override
public DataChecksum.Type crcType() {
return DataChecksum.Type.CRC32;
}
}
final class NativeC implements Crc32 {
@Override
public void verifyChunked(ByteBuffer data, int bytesPerSum,
ByteBuffer sums, String fileName, long basePos)
throws ChecksumException {
static abstract class AbstractCrc32<T extends Checksum> implements Crc32 {
if (data.isDirect()) {
NativeCrc32.verifyChunkedSums(bytesPerSum,
DataChecksum.Type.CRC32C.id, sums, data, fileName, basePos);
} else {
final int dataOffset = data.arrayOffset() + data.position();
final int crcsOffset = sums.arrayOffset() + sums.position();
NativeCrc32.verifyChunkedSumsByteArray(bytesPerSum,
DataChecksum.Type.CRC32C.id, sums.array(), crcsOffset,
data.array(), dataOffset, data.remaining(), fileName, basePos);
}
}
@Override
public DataChecksum.Type crcType() {
return DataChecksum.Type.CRC32C;
}
}
abstract class AbstractCrc32<T extends Checksum> implements Crc32 {
abstract T newAlgorithm();
@Override
public void verifyChunked(ByteBuffer data, int bytesPerCrc,
ByteBuffer crcs, String filename, long basePos)
ByteBuffer sums, String filename, long basePos)
throws ChecksumException {
final Checksum algorithm = newAlgorithm();
if (data.hasArray() && crcs.hasArray()) {
DataChecksum.verifyChunked(DataChecksum.Type.CRC32, algorithm,
data.array(), data.position(), data.remaining(), bytesPerCrc,
crcs.array(), crcs.position(), filename, basePos);
final DataChecksum.Type type = crcType();
if (data.hasArray() && sums.hasArray()) {
DataChecksum.verifyChunked(type, algorithm, data.array(),
data.position(), data.remaining(), bytesPerCrc, sums.array(),
sums.position(), filename, basePos);
} else {
DataChecksum.verifyChunked(DataChecksum.Type.CRC32, algorithm,
data, bytesPerCrc, crcs, filename, basePos);
DataChecksum.verifyChunked(type, algorithm, data, bytesPerCrc,
sums, filename, basePos);
}
}
}
static final class Zip extends AbstractCrc32<CRC32> {
final class Zip extends AbstractCrc32<CRC32> {
@Override
public CRC32 newAlgorithm() {
return new CRC32();
}
@Override
public DataChecksum.Type crcType() {
return DataChecksum.Type.CRC32;
}
}
static final class PureJava extends AbstractCrc32<PureJavaCrc32> {
final class PureJava extends AbstractCrc32<PureJavaCrc32> {
@Override
public PureJavaCrc32 newAlgorithm() {
return new PureJavaCrc32();
}
@Override
public DataChecksum.Type crcType() {
return DataChecksum.Type.CRC32;
}
}
final class PureJavaC extends AbstractCrc32<PureJavaCrc32C> {
@Override
public PureJavaCrc32C newAlgorithm() {
return new PureJavaCrc32C();
}
@Override
public DataChecksum.Type crcType() {
return DataChecksum.Type.CRC32C;
}
}
}
@ -114,9 +169,13 @@ public class Crc32PerformanceTest {
crcs.add(Crc32.Zip.class);
crcs.add(Crc32.PureJava.class);
crcs.add(Crc32.PureJavaC.class);
if (direct && NativeCrc32.isAvailable()) {
crcs.add(Crc32.Native.class);
if (NativeCrc32.isAvailable()) {
if (direct) {
crcs.add(Crc32.Native.class);
}
crcs.add(Crc32.NativeC.class);
((Log4JLogger)LogFactory.getLog(NativeCodeLoader.class))
.getLogger().setLevel(Level.ALL);
}
@ -131,13 +190,18 @@ public class Crc32PerformanceTest {
out.printf("Elapsed %.1fs\n", secondsElapsed(startTime));
}
public static void main(String args[]) throws Exception {
new Crc32PerformanceTest(64, 5, true).run();
public static void main(String[] args) throws Exception {
boolean isdirect = true;
if (args.length > 0) {
isdirect = false;
}
new Crc32PerformanceTest(64, 5, isdirect).run();
}
private static void printCell(String s, int width, PrintStream out) {
private static void printCell(String s, int width, PrintStream outCrc) {
final int w = s.length() > width? s.length(): width;
out.printf(" %" + w + "s |", s);
outCrc.printf(" %" + w + "s |", s);
}
private ByteBuffer allocateByteBuffer(int length) {
@ -155,16 +219,29 @@ public class Crc32PerformanceTest {
return dataBufs;
}
private ByteBuffer computeCrc(ByteBuffer dataBufs, int bytePerCrc) {
private ByteBuffer computeCrc(ByteBuffer dataBufs, int bytePerCrc,
DataChecksum.Type type) {
final int size = 4 * (dataBufs.remaining() - 1) / bytePerCrc + 1;
final ByteBuffer crcBufs = allocateByteBuffer(size);
final DataChecksum checksum = DataChecksum.newDataChecksum(
DataChecksum.Type.CRC32, bytePerCrc);
type, bytePerCrc);
checksum.calculateChunkedSums(dataBufs, crcBufs);
return crcBufs;
}
private void doBench(final List<Class<? extends Crc32>> crcs)
private ByteBuffer computeCrc(Class<? extends Crc32> clazz,
ByteBuffer dataBufs, int bytePerCrc) throws Exception {
final Constructor<? extends Crc32> ctor = clazz.getConstructor();
final Crc32 crc = ctor.newInstance();
final int size = 4 * (dataBufs.remaining() - 1) / bytePerCrc + 1;
final ByteBuffer crcBufs = allocateByteBuffer(size);
final DataChecksum checksum = DataChecksum.newDataChecksum(
crc.crcType(), bytePerCrc);
checksum.calculateChunkedSums(dataBufs, crcBufs);
return crcBufs;
}
private void doBench(final List<Class<? extends Crc32>> crcTargets)
throws Exception {
final ByteBuffer[] dataBufs = new ByteBuffer[16];
for(int i = 0; i < dataBufs.length; i++) {
@ -176,9 +253,9 @@ public class Crc32PerformanceTest {
out.printf(" (bpc: byte-per-crc in MB/sec; #T: #Theads)\n");
// Warm up implementations to get jit going.
final ByteBuffer[] crc32 = {computeCrc(dataBufs[0], 32)};
final ByteBuffer[] crc512 = {computeCrc(dataBufs[0], 512)};
for (Class<? extends Crc32> c : crcs) {
for (Class<? extends Crc32> c : crcTargets) {
final ByteBuffer[] crc32 = {computeCrc(c, dataBufs[0], 32)};
final ByteBuffer[] crc512 = {computeCrc(c, dataBufs[0], 512)};
doBench(c, 1, dataBufs, crc32, 32);
doBench(c, 1, dataBufs, crc512, 512);
}
@ -189,57 +266,69 @@ public class Crc32PerformanceTest {
}
}
private void doBench(final List<Class<? extends Crc32>> crcs,
final ByteBuffer[] dataBufs, final int bytePerCrc, final PrintStream out)
private void doBench(final List<Class<? extends Crc32>> crcTargets,
final ByteBuffer[] dataBufs, final int bytePerCrc,
final PrintStream outCrc)
throws Exception {
final ByteBuffer[] crcBufs = new ByteBuffer[dataBufs.length];
for(int i = 0; i < crcBufs.length; i++) {
crcBufs[i] = computeCrc(dataBufs[i], bytePerCrc);
final ByteBuffer[] crcBufsC = new ByteBuffer[dataBufs.length];
for(int i = 0; i < dataBufs.length; i++) {
crcBufs[i] = computeCrc(dataBufs[i], bytePerCrc,
DataChecksum.Type.CRC32);
crcBufsC[i] = computeCrc(dataBufs[i], bytePerCrc,
DataChecksum.Type.CRC32C);
}
final String numBytesStr = " bpc ";
final String numThreadsStr = "#T";
final String diffStr = "% diff";
out.print('|');
printCell(numBytesStr, 0, out);
printCell(numThreadsStr, 0, out);
for (int i = 0; i < crcs.size(); i++) {
final Class<? extends Crc32> c = crcs.get(i);
out.print('|');
printCell(c.getSimpleName(), 8, out);
for(int j = 0; j < i; j++) {
printCell(diffStr, diffStr.length(), out);
outCrc.print('|');
printCell(numBytesStr, 0, outCrc);
printCell(numThreadsStr, 0, outCrc);
for (int i = 0; i < crcTargets.size(); i++) {
final Class<? extends Crc32> c = crcTargets.get(i);
outCrc.print('|');
printCell(c.getSimpleName(), 8, outCrc);
if (i > 0) {
printCell(diffStr, diffStr.length(), outCrc);
}
}
out.printf("\n");
outCrc.printf("\n");
for(int numThreads = 1; numThreads <= dataBufs.length; numThreads <<= 1) {
out.printf("|");
printCell(String.valueOf(bytePerCrc), numBytesStr.length(), out);
printCell(String.valueOf(numThreads), numThreadsStr.length(), out);
outCrc.printf("|");
printCell(String.valueOf(bytePerCrc), numBytesStr.length(), outCrc);
printCell(String.valueOf(numThreads), numThreadsStr.length(), outCrc);
final List<BenchResult> previous = new ArrayList<BenchResult>();
for(Class<? extends Crc32> c : crcs) {
for(Class<? extends Crc32> c : crcTargets) {
System.gc();
final BenchResult result = doBench(c, numThreads, dataBufs, crcBufs,
bytePerCrc);
final BenchResult result;
final Constructor<? extends Crc32> ctor = c.getConstructor();
final Crc32 crc = ctor.newInstance();
if (crc.crcType() == DataChecksum.Type.CRC32) {
result = doBench(c, numThreads, dataBufs, crcBufs, bytePerCrc);
} else {
result = doBench(c, numThreads, dataBufs, crcBufsC, bytePerCrc);
}
printCell(String.format("%9.1f", result.mbps),
c.getSimpleName().length() + 1, out);
c.getSimpleName().length() + 1, outCrc);
//compare result with previous
for(BenchResult p : previous) {
//compare result with the last previous.
final int size = previous.size();
if (size > 0) {
BenchResult p = previous.get(size - 1);
final double diff = (result.mbps - p.mbps) / p.mbps * 100;
printCell(String.format("%5.1f%%", diff), diffStr.length(), out);
printCell(String.format("%5.1f%%", diff), diffStr.length(), outCrc);
}
previous.add(result);
}
out.printf("\n");
outCrc.printf("\n");
}
}
private BenchResult doBench(Class<? extends Crc32> clazz,
final int numThreads, final ByteBuffer[] dataBufs,
final ByteBuffer[] crcBufs, final int bytePerCrc)
@ -248,36 +337,34 @@ public class Crc32PerformanceTest {
final Thread[] threads = new Thread[numThreads];
final BenchResult[] results = new BenchResult[threads.length];
{
final Constructor<? extends Crc32> ctor = clazz.getConstructor();
final Constructor<? extends Crc32> ctor = clazz.getConstructor();
for(int i = 0; i < threads.length; i++) {
final Crc32 crc = ctor.newInstance();
final long byteProcessed = dataBufs[i].remaining() * trials;
final int index = i;
threads[i] = new Thread() {
@Override
public void run() {
final long startTime = System.nanoTime();
for (int i = 0; i < trials; i++) {
dataBufs[index].mark();
crcBufs[index].mark();
try {
crc.verifyChunked(dataBufs[index], bytePerCrc, crcBufs[index],
crc.getClass().getSimpleName(), dataBufs[index].position());
} catch (Throwable t) {
results[index] = new BenchResult(t);
return;
} finally {
dataBufs[index].reset();
crcBufs[index].reset();
}
for(int i = 0; i < threads.length; i++) {
final Crc32 crc = ctor.newInstance();
final long byteProcessed = dataBufs[i].remaining() * trials;
final int index = i;
threads[i] = new Thread() {
@Override
public void run() {
final long startTime = System.nanoTime();
for (int i = 0; i < trials; i++) {
dataBufs[index].mark();
crcBufs[index].mark();
try {
crc.verifyChunked(dataBufs[index], bytePerCrc, crcBufs[index],
crc.getClass().getSimpleName(), dataBufs[index].position());
} catch (Throwable t) {
results[index] = new BenchResult(t);
return;
} finally {
dataBufs[index].reset();
crcBufs[index].reset();
}
final double secsElapsed = secondsElapsed(startTime);
results[index] = new BenchResult(byteProcessed/secsElapsed/MB);
}
};
}
final double secsElapsed = secondsElapsed(startTime);
results[index] = new BenchResult(byteProcessed/secsElapsed/MB);
}
};
}
for(Thread t : threads) {
@ -295,7 +382,7 @@ public class Crc32PerformanceTest {
}
private static class BenchResult {
/** Speed (MB per second) */
/** Speed (MB per second). */
final double mbps;
final Throwable thrown;
@ -321,7 +408,7 @@ public class Crc32PerformanceTest {
return (System.nanoTime() - startTime) / 1000000000.0d;
}
static void printSystemProperties(PrintStream out) {
static void printSystemProperties(PrintStream outCrc) {
final String[] names = {
"java.version",
"java.runtime.name",
@ -344,7 +431,7 @@ public class Crc32PerformanceTest {
final Properties p = System.getProperties();
for(String n : names) {
out.printf("%" + max + "s = %s\n", n, p.getProperty(n));
outCrc.printf("%" + max + "s = %s\n", n, p.getProperty(n));
}
}
}