HDFS-8968. Erasure coding: a comprehensive I/O throughput benchmark tool. Contributed by Rui Li.

Change-Id: I01ff5b04727fd79e6373582d4815c5e7b2096c67
This commit is contained in:
Zhe Zhang 2015-11-12 11:30:05 -08:00
parent 7f55a18071
commit 7b00c8e20e
3 changed files with 542 additions and 0 deletions

View File

@ -201,6 +201,9 @@ Trunk (Unreleased)
HDFS-9234. WebHdfs: getContentSummary() should give quota for storage types. HDFS-9234. WebHdfs: getContentSummary() should give quota for storage types.
(Surendra Singh Lilhore via xyao) (Surendra Singh Lilhore via xyao)
HDFS-8968. Erasure coding: a comprehensive I/O throughput benchmark tool.
(Rui Li via zhz)
OPTIMIZATIONS OPTIMIZATIONS
BUG FIXES BUG FIXES

View File

@ -0,0 +1,423 @@
/**
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you 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 org.apache.hadoop.hdfs;
import com.google.common.base.Preconditions;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.fs.FSDataInputStream;
import org.apache.hadoop.fs.FSDataOutputStream;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.fs.PathFilter;
import org.apache.hadoop.hdfs.client.HdfsClientConfigKeys;
import org.apache.hadoop.hdfs.protocol.ErasureCodingPolicy;
import org.apache.hadoop.hdfs.server.namenode.ErasureCodingPolicyManager;
import org.apache.hadoop.util.StopWatch;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;
import java.io.IOException;
import java.nio.ByteBuffer;
import java.text.DecimalFormat;
import java.util.ArrayList;
import java.util.List;
import java.util.Random;
import java.util.concurrent.Callable;
import java.util.concurrent.CompletionService;
import java.util.concurrent.ExecutorCompletionService;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
import java.util.concurrent.TimeUnit;
/**
* This class benchmarks the throughput of client read/write for both replica
* and Erasure Coding.
* <p/>
* Currently 4 operations are supported: read, write, generate and cleanup data.
* Users should specify an operation, the amount of data in MB for a single
* client, and which storage policy to use, i.e. EC or replication.
* Optionally, users can specify the number of clients to launch concurrently.
* The tool launches 1 thread for each client. Number of client is 1 by default.
* For reading, users can also specify whether stateful or positional read
* should be used. Stateful read is chosen by default.
* <p/>
* Each client reads and writes different files.
* For writing, client writes a temporary file at the desired amount, and the
* file will be cleaned up when the test finishes.
* For reading, each client tries to read the file specific to itself. And the
* client simply returns if such file does not exist. Therefore, users should
* generate the files before testing read. Generating data is essentially the
* same as writing, except that the files won't be cleared at the end.
* For example, if the user wants to test reading 1024MB data with 10 clients,
* he/she should firstly generate 1024MB data with 10 (or more) clients.
*/
public class ErasureCodeBenchmarkThroughput extends Configured implements Tool {
private static final int BUFFER_SIZE_MB = 128;
private static final String DFS_TMP_DIR = System.getProperty(
"test.benchmark.data", "/tmp/benchmark/data");
public static final String REP_DIR = DFS_TMP_DIR + "/replica";
public static final String EC_DIR = DFS_TMP_DIR + "/ec";
private static final String REP_FILE_BASE = "rep-file-";
private static final String EC_FILE_BASE = "ec-file-";
private static final String TMP_FILE_SUFFIX = ".tmp";
private static final ErasureCodingPolicy ecPolicy =
ErasureCodingPolicyManager.getSystemDefaultPolicy();
private static final byte[] data = new byte[BUFFER_SIZE_MB * 1024 * 1024];
static {
Random random = new Random();
random.nextBytes(data);
}
private final FileSystem fs;
public static ErasureCodingPolicy getEcPolicy() {
return ecPolicy;
}
public ErasureCodeBenchmarkThroughput(FileSystem fs) {
Preconditions.checkArgument(fs instanceof DistributedFileSystem);
this.fs = fs;
}
enum OpType {
READ, WRITE, GEN, CLEAN;
}
public static String getFilePath(int dataSizeMB, boolean isEc) {
String parent = isEc ? EC_DIR : REP_DIR;
String file = isEc ? EC_FILE_BASE : REP_FILE_BASE;
return parent + "/" + file + dataSizeMB + "MB";
}
private static void printUsage(String msg) {
if (msg != null) {
System.out.println(msg);
}
System.err.println("Usage: ErasureCodeBenchmarkThroughput " +
"<read|write|gen|clean> <size in MB> <ec|rep> [num clients] [stf|pos]\n"
+ "Stateful and positional option is only available for read.");
System.exit(1);
}
private List<Long> doBenchmark(boolean isRead, int dataSizeMB, int numClients,
boolean isEc, boolean statefulRead, boolean isGen) throws Exception {
CompletionService<Long> cs = new ExecutorCompletionService<Long>(
Executors.newFixedThreadPool(numClients));
for (int i = 0; i < numClients; i++) {
cs.submit(isRead ?
new ReadCallable(dataSizeMB, isEc, i, statefulRead) :
new WriteCallable(dataSizeMB, isEc, i, isGen));
}
List<Long> results = new ArrayList<>(numClients);
for (int i = 0; i < numClients; i++) {
results.add(cs.take().get());
}
return results;
}
private void setReadThreadPoolSize(int numClients) {
int numThread = numClients * ecPolicy.getNumDataUnits();
getConf().setInt(HdfsClientConfigKeys.StripedRead.THREADPOOL_SIZE_KEY,
numThread);
}
private DecimalFormat getDecimalFormat() {
return new DecimalFormat("#.##");
}
private void benchmark(OpType type, int dataSizeMB,
int numClients, boolean isEc, boolean statefulRead) throws Exception {
List<Long> sizes = null;
StopWatch sw = new StopWatch().start();
switch (type) {
case READ:
sizes = doBenchmark(true, dataSizeMB, numClients, isEc,
statefulRead, false);
break;
case WRITE:
sizes = doBenchmark(
false, dataSizeMB, numClients, isEc, statefulRead, false);
break;
case GEN:
sizes = doBenchmark(false, dataSizeMB, numClients, isEc,
statefulRead, true);
}
long elapsedSec = sw.now(TimeUnit.SECONDS);
double totalDataSizeMB = 0;
for (Long size : sizes) {
if (size >= 0) {
totalDataSizeMB += size.doubleValue() / 1024 / 1024;
}
}
double throughput = totalDataSizeMB / elapsedSec;
DecimalFormat df = getDecimalFormat();
System.out.println(type + " " + df.format(totalDataSizeMB) +
" MB data takes: " + elapsedSec + " s.\nTotal throughput: " +
df.format(throughput) + " MB/s.");
}
private void setUpDir() throws IOException {
DistributedFileSystem dfs = (DistributedFileSystem) fs;
dfs.mkdirs(new Path(DFS_TMP_DIR));
Path repPath = new Path(REP_DIR);
Path ecPath = new Path(EC_DIR);
if (!dfs.exists(repPath)) {
dfs.mkdirs(repPath);
} else {
Preconditions.checkArgument(
dfs.getClient().getErasureCodingPolicy(repPath.toString()) == null);
}
if (!dfs.exists(ecPath)) {
dfs.mkdirs(ecPath);
dfs.getClient().setErasureCodingPolicy(ecPath.toString(), ecPolicy);
} else {
Preconditions.checkArgument(
dfs.getClient().
getErasureCodingPolicy(ecPath.toString()).equals(ecPolicy));
}
}
@Override
public int run(String[] args) throws Exception {
OpType type = null;
int dataSizeMB = 0;
boolean isEc = true;
int numClients = 1;
boolean statefulRead = true;
if (args.length >= 3) {
if (args[0].equals("read")) {
type = OpType.READ;
} else if (args[0].equals("write")) {
type = OpType.WRITE;
} else if (args[0].equals("gen")) {
type = OpType.GEN;
} else if (args[0].equals("clean")) {
type = OpType.CLEAN;
} else {
printUsage("Unknown operation: " + args[0]);
}
try {
dataSizeMB = Integer.valueOf(args[1]);
if (dataSizeMB <= 0) {
printUsage("Invalid data size: " + dataSizeMB);
}
} catch (NumberFormatException e) {
printUsage("Invalid data size: " + e.getMessage());
}
isEc = args[2].equals("ec");
if (!isEc && !args[2].equals("rep")) {
printUsage("Unknown storage policy: " + args[2]);
}
} else {
printUsage(null);
}
if (args.length >= 4 && type != OpType.CLEAN) {
try {
numClients = Integer.valueOf(args[3]);
if (numClients <= 0) {
printUsage("Invalid num of clients: " + numClients);
}
} catch (NumberFormatException e) {
printUsage("Invalid num of clients: " + e.getMessage());
}
}
if (args.length >= 5 && type == OpType.READ) {
statefulRead = args[4].equals("stf");
if (!statefulRead && !args[4].equals("pos")) {
printUsage("Unknown read mode: " + args[4]);
}
}
setUpDir();
if (type == OpType.CLEAN) {
cleanUp(dataSizeMB, isEc);
} else {
if (type == OpType.READ && isEc) {
setReadThreadPoolSize(numClients);
}
benchmark(type, dataSizeMB, numClients, isEc, statefulRead);
}
return 0;
}
private void cleanUp(int dataSizeMB, boolean isEc) throws IOException {
final String fileName = getFilePath(dataSizeMB, isEc);
Path path = isEc ? new Path(EC_DIR) : new Path(REP_DIR);
FileStatus fileStatuses[] = fs.listStatus(path, new PathFilter() {
@Override
public boolean accept(Path path) {
return path.toString().contains(fileName);
}
});
for (FileStatus fileStatus : fileStatuses) {
fs.delete(fileStatus.getPath(), false);
}
}
/**
* A Callable that returns the number of bytes read/written
*/
private abstract class CallableBase implements Callable<Long> {
protected final int dataSizeMB;
protected final boolean isEc;
protected final int id;
public CallableBase(int dataSizeMB, boolean isEc, int id)
throws IOException {
this.dataSizeMB = dataSizeMB;
this.isEc = isEc;
this.id = id;
}
protected String getFilePathForThread() {
return getFilePath(dataSizeMB, isEc) + "_" + id;
}
}
private class WriteCallable extends CallableBase {
private final boolean isGen;
public WriteCallable(int dataSizeMB, boolean isEc, int id, boolean isGen)
throws IOException {
super(dataSizeMB, isEc, id);
this.isGen = isGen;
}
private long writeFile(Path path) throws IOException {
StopWatch sw = new StopWatch().start();
System.out.println("Writing " + path);
long dataSize = dataSizeMB * 1024 * 1024L;
long remaining = dataSize;
try (FSDataOutputStream outputStream = fs.create(path)) {
if (!isGen) {
fs.deleteOnExit(path);
}
int toWrite;
while (remaining > 0) {
toWrite = (int) Math.min(remaining, data.length);
outputStream.write(data, 0, toWrite);
remaining -= toWrite;
}
System.out.println("Finished writing " + path + ". Time taken: " +
sw.now(TimeUnit.SECONDS) + " s.");
return dataSize - remaining;
}
}
@Override
public Long call() throws Exception {
String pathStr = getFilePathForThread();
if (!isGen) {
pathStr += TMP_FILE_SUFFIX;
}
final Path path = new Path(pathStr);
if (fs.exists(path)) {
if (isGen) {
System.out.println("Data already generated at " + path);
} else {
System.out.println("Previous tmp data not cleaned " + path);
}
return 0L;
}
return writeFile(path);
}
}
private class ReadCallable extends CallableBase {
private final boolean statefulRead;
public ReadCallable(int dataSizeMB, boolean isEc, int id,
boolean statefulRead) throws IOException {
super(dataSizeMB, isEc, id);
this.statefulRead = statefulRead;
}
private long doStateful(FSDataInputStream inputStream) throws IOException {
long count = 0;
long bytesRead;
ByteBuffer buffer = ByteBuffer.allocate(BUFFER_SIZE_MB * 1024 * 1024);
while (true) {
bytesRead = inputStream.read(buffer);
if (bytesRead < 0) {
break;
}
count += bytesRead;
buffer.clear();
}
return count;
}
private long doPositional(FSDataInputStream inputStream)
throws IOException {
long count = 0;
long bytesRead;
byte buf[] = new byte[BUFFER_SIZE_MB * 1024 * 1024];
while (true) {
bytesRead = inputStream.read(count, buf, 0, buf.length);
if (bytesRead < 0) {
break;
}
count += bytesRead;
}
return count;
}
private long readFile(Path path) throws IOException {
try (FSDataInputStream inputStream = fs.open(path)) {
StopWatch sw = new StopWatch().start();
System.out.println((statefulRead ? "Stateful reading " :
"Positional reading ") + path);
long totalRead = statefulRead ? doStateful(inputStream) :
doPositional(inputStream);
System.out.println(
(statefulRead ? "Finished stateful read " :
"Finished positional read ") + path + ". Time taken: " +
sw.now(TimeUnit.SECONDS) + " s.");
return totalRead;
}
}
@Override
public Long call() throws Exception {
Path path = new Path(getFilePathForThread());
if (!fs.exists(path) || fs.isDirectory(path)) {
System.out.println("File not found at " + path +
". Call gen first?");
return 0L;
}
long bytesRead = readFile(path);
long dataSize = dataSizeMB * 1024 * 1024L;
Preconditions.checkArgument(bytesRead == dataSize,
"Specified data size: " + dataSize + ", actually read " + bytesRead);
return bytesRead;
}
}
public static void main(String[] args) throws Exception {
Configuration conf = new HdfsConfiguration();
FileSystem fs = FileSystem.get(conf);
int res = ToolRunner.run(conf,
new ErasureCodeBenchmarkThroughput(fs), args);
System.exit(res);
}
}

View File

@ -0,0 +1,116 @@
/**
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you 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 org.apache.hadoop.hdfs;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.fs.PathFilter;
import org.apache.hadoop.util.ToolRunner;
import org.junit.AfterClass;
import org.junit.Assert;
import org.junit.BeforeClass;
import org.junit.Test;
import java.io.IOException;
/**
* To test {@link org.apache.hadoop.hdfs.ErasureCodeBenchmarkThroughput}.
*/
public class TestErasureCodeBenchmarkThroughput {
private static MiniDFSCluster cluster;
private static Configuration conf;
private static FileSystem fs;
@BeforeClass
public static void setup() throws IOException {
conf = new HdfsConfiguration();
int numDN = ErasureCodeBenchmarkThroughput.getEcPolicy().getNumDataUnits() +
ErasureCodeBenchmarkThroughput.getEcPolicy().getNumParityUnits();
cluster = new MiniDFSCluster.Builder(conf).numDataNodes(numDN).build();
cluster.waitActive();
fs = cluster.getFileSystem();
}
@AfterClass
public static void tearDown() {
if (cluster != null) {
cluster.shutdown(true);
}
}
private static void runBenchmark(String[] args) throws Exception {
Assert.assertNotNull(conf);
Assert.assertNotNull(fs);
Assert.assertEquals(0, ToolRunner.run(conf,
new ErasureCodeBenchmarkThroughput(fs), args));
}
private static void verifyNumFile(final int dataSize, final boolean isEc,
int numFile) throws IOException {
Path path = isEc ? new Path(ErasureCodeBenchmarkThroughput.EC_DIR) :
new Path(ErasureCodeBenchmarkThroughput.REP_DIR);
FileStatus[] statuses = fs.listStatus(path, new PathFilter() {
@Override
public boolean accept(Path path) {
return path.toString().contains(
ErasureCodeBenchmarkThroughput.getFilePath(dataSize, isEc));
}
});
Assert.assertEquals(numFile, statuses.length);
}
@Test
public void testReplicaReadWrite() throws Exception {
Integer dataSize = 10;
Integer numClient = 3;
String[] args = new String[]{"write", dataSize.toString(), "rep",
numClient.toString()};
runBenchmark(args);
args[0] = "gen";
runBenchmark(args);
args[0] = "read";
runBenchmark(args);
}
@Test
public void testECReadWrite() throws Exception {
Integer dataSize = 5;
Integer numClient = 5;
String[] args = new String[]{"write", dataSize.toString(), "ec",
numClient.toString()};
runBenchmark(args);
args[0] = "gen";
runBenchmark(args);
args[0] = "read";
runBenchmark(args);
}
@Test
public void testCleanUp() throws Exception {
Integer dataSize = 5;
Integer numClient = 5;
String[] args = new String[]{"gen", dataSize.toString(), "ec",
numClient.toString()};
runBenchmark(args);
args[0] = "clean";
runBenchmark(args);
verifyNumFile(dataSize, true, 0);
}
}