前言
众所周知,在HDFS NameNode中,一直都有一个老生常谈的难题就是其扩展性的问题,而很多时候我们说HDFS的扩展性问题时我们很多时候都在谈的点在于里面全局锁的问题。一个很通常的场景是NameNode在高并发请求处理下存在着激烈的锁竞争,进而使得用户感觉到他们的请求被处理的有点慢。不过本文笔者不聊关于全局锁优化的问题,最近笔者遇到了另外一种NameNode性能慢的场景,这个事情是发生在锁外的,发生的场景也比较有意思,于是借此机会简单聊聊。
NameNode请求处理慢的场景
我们先来探讨几个能够引发NameNode请求处理慢的场景?
- NameNode处于严重Full GC阶段
- NameNode锁竞争激烈
- NameNode内部用于处理用户请求的Handler持续处于繁忙状态
上述3点是笔者能够快速联想到的场景,除去第一点是JVM层面的优化后,后面两点其实不太好优化,后面两点很多时候是和集群所面临的请求吞吐量有关联。有人可能会说了,上面第三点关于Handler繁忙的状态,我们不可用简单调大NameNode Handler count数量不就可用解决了嘛。调大Handler表面上能够增大throughput,但是这还意味着一个问题:在同一个时间,NameNode内部会有更为激烈的锁竞争问题。锁竞争意味着就会有锁等待。
所以针对上面Handler繁忙的问题,应该从以下几点切入点入手:
Handler为什么处于繁忙状态?Handler在处理 RPC call的某个阶段慢了?是因为在等锁所以导致Handler处理慢?
OK,下面笔者来好好聊聊笔者遇到的Handler处理慢的问题。
RPC返回response的Handler处理慢问题
在笔者遇到的生产场景中,时不时就会出现NameNode callqueue满的情况,但是它的process time其实并没有显示出特别慢的迹象。在最一开始,笔者也是自然怀疑的是NameNode锁的问题,后来经过了一系列的锁局部优化后,还是偶尔会遇到NameNode慢的问题。
这个时候,我们采用了一样profile利器:async-profile,来帮助我们一探究竟。
于是我们发现了一段开销较高的doResponse阶段,
从上图中,我们果然捕捉到了Handler的run方法处理阶段,然后我们再点击进入里面详细的堆栈信息方法,如下图所示:
从上图中,我们可以得到许多有用的信息点:
- Handler慢发生在返回response的操作里面,这个阶段已经是在锁外了,在RPC处理完毕之后了已经。
- Response返回操作里面又是主要花在了加密返回信息内容的步骤上。
这里需要说下背景,笔者的这个集群环境是开启了Security模式的,所以会有response加密的过程。一般集群如果没有开启安全认证这些的话,这步骤估计是没有的。
OK,找到了问题的根源后,我们立马想到了下面两项action:
- 第一点,是否有可能改进或替换现有的response加密算法,提升其加密效率。
- 第二点,是否能够将返回response阶段进行异步处理,不去阻塞Handler的处理。
鉴于第一种方法涉及到加密算法的改动,risk比较高,因此我们先考虑采用第二点的方法。还有一点是,第二点提到的特性在实际代码中已经可以支持,async editlog就是用了Hadoop RPC Server底层这个异步response特性来做的。现在的问题其实是如何让正常RPC call处理也能走异步response的方式,目前的HDFS RPC response返回都是同步的。
HDFS RPC call异步response改造
接下来,笔者仿照async editlog的处理方式,对局部RPC call进行了试验改造。
以下是相关核心代码的改动,主要仿照了HDFS-9198 IBR异步线程处理加上HADOOP-10300 RPC call postpone的处理逻辑。
首先定义好postpone线程专门处理这样的延时response call,
/**
* The flag that if we should skip the handle by PostponeRpcCallThread.
* This is only used for testing.
*/
@VisibleForTesting
static boolean SKIP_POSTPONE_HANDLE = false;
/**
* Whether to enable async rpc call response. If we enabled this feature,
* the rpc call response will be postponed return to client side.
*/
private boolean asyncResponse;
/**
* The thread specified for dealing with postponed rpc calls.
*/
private PostponeRpcCallQueueProcessingThread postponeRpcCallThread;
然后在这个线程内,会设置一个队里来存放需要延时返回response call的队列,外加一个加call进队列的方法和处理队列的方法操作。
/**
* The thread specified for dealing with postponed rpc calls.
*/
private class PostponeRpcCallQueueProcessingThread extends Thread {
private final BlockingQueue<Server.Call> postponeRpcCallqueue =
new ArrayBlockingQueue<>(1024);
private static final long MAX_LOCK_HOLD_MS = 4;
private long lastFullTime = 0;
PostponeRpcCallQueueProcessingThread() {
super("Postpone Rpc call processor");
setDaemon(true);
}
@Override
public void run() {
try {
processQueue();
} catch (Throwable t) {
ExitUtil.terminate(1, getName() +
" encountered fatal exception: " + t);
}
}
,,,
/**
* Insert rpc call to be postponed response.
* @param call
* @throws InterruptedException
*/
private void enqueue(Server.Call call) throws InterruptedException {
if (!postponeRpcCallqueue.offer(call)) {
long now = Time.monotonicNow();
if (now - lastFullTime > 4000) {
lastFullTime = now;
LOG.info("Postpone rpc call queue is full.");
}
postponeRpcCallqueue.put(call);
}
}
}
接下来我们看看这里NameNode如何加入RPC call进队列,笔者这里取了getBlockLocations这个call做为测试,
/**
* Get block locations within the specified range.
* @see ClientProtocol#getBlockLocations(String, long, long)
*/
LocatedBlocks getBlockLocations(String clientMachine, String srcArg,
long offset, long length) throws IOException {
final String operationName = "open";
checkOperation(OperationCategory.READ);
GetBlockLocationsResult res = null;
FSPermissionChecker pc = getPermissionChecker();
readLock();
try {
checkOperation(OperationCategory.READ);
res = getBlockLocations(pc, srcArg, offset, length, true, true);
} catch (AccessControlException e) {
logAuditEvent(false, operationName, srcArg);
throw e;
} finally {
readUnlock(operationName);
}
...
postponeResponse("getBlockLocations");
return blocks;
}
/**
* Postponed current rpc call response.
* @param methodName Rpc call operation name.
*/
private void postponeResponse(String methodName) {
// if we don't enabled async response, just return.
if (!asyncResponse) {
return;
}
// 获取当前的RPC call,即这里的getBlockLocations call
final Server.Call rpcCall = Server.getCurCall().get();
// 触发postponeResponse处理会使得此RPC的response会被延时,只有额外再触发一次sendResponse才会真正执行回复
rpcCall.postponeResponse();
try {
postponeRpcCallThread.enqueue(rpcCall);
} catch (InterruptedException e) {
String message = "Exception while edit logging: " + e.getMessage();
LOG.fatal("Exception while adding postpone rpc call", e);
ExitUtil.terminate(1, message);
}
LOG.info("Add " + methodName + " rpc call into postpone rpc call queue.");
}
在getBlockLocations最末尾端加postpone的处理是因为一旦前面执行操作抛出错误,这里不应该做postpone处理,延时处理只作用于正常PRC处理完成的情况下。
OK,最后我们再回到call队列处理的操作,
/**
* Process the postpone call queue.
*/
private void processQueue() {
// this only used for the testing
if (skipPostponeHandle) {
postponeRpcCallqueue.clear();
return;
}
while (isRunning()) {
NameNodeMetrics metrics = NameNode.getNameNodeMetrics();
try {
Server.Call call = postponeRpcCallqueue.take();
// batch as many operations in the write lock until the queue
// runs dry, or the max lock hold is reached.
writeLock();
metrics.setPostponeRpcCallQueued(postponeRpcCallqueue.size() + 1);
try {
long start = Time.monotonicNow();
do {
try {
// 此时才执行真正的返回response的操作
call.sendResponse();
LOG.info("Triggered postpone rpc call sendResponse function.");
} catch (IOException ioe) {
LOG.error("Trigger send response call error", ioe);
}
if (Time.monotonicNow() - start > MAX_LOCK_HOLD_MS) {
break;
}
call = postponeRpcCallqueue.poll();
} while (call != null);
} finally {
writeUnlock();
}
} catch (InterruptedException e) {
// ignore unless thread was specifically interrupted.
if (Thread.interrupted()) {
break;
}
}
}
postponeRpcCallqueue.clear();
}
鉴于此Thread是一个仅run在Active service上的服务,所以需要在startActiveService/stopActiveService方法里进行thread的启动和停止。
/**
* Start services required in active state
* @throws IOException
*/
void startActiveServices() throws IOException {
startingActiveService = true;
LOG.info("Starting services required for active state");
writeLock();
try {
...
if (asyncResponse) {
postponeRpcCallThread = new PostponeRpcCallQueueProcessingThread();
postponeRpcCallThread.start();
}
} finally {
startingActiveService = false;
checkSafeMode();
writeUnlock("startActiveServices");
}
}
/**
* Stop services required in active state
*/
void stopActiveServices() {
LOG.info("Stopping services started for active state");
writeLock();
try {
...
if (postponeRpcCallThread != null) {
try {
postponeRpcCallThread.interrupt();
postponeRpcCallThread.join(3000);
} catch (InterruptedException ie) {
}
}
initializedReplQueues = false;
} finally {
writeUnlock("stopActiveServices");
}
}
与此对应的unit test测试类代码,server端postpone行为对于client的影响就是block其请求结果的返回。
/**
* 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.server.namenode;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.CommonConfigurationKeys;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hdfs.*;
import org.apache.hadoop.hdfs.protocol.LocatedBlocks;
import org.junit.Assert;
import org.junit.Test;
import java.io.IOException;
import java.net.SocketTimeoutException;
/**
* The unit test for PostponeRpcCallQueueProcessingThread handling.
*/
public class TestPostponeResponseHandling {
private static final Log LOG = LogFactory.getLog(TestPostponeResponseHandling.class);
private Configuration conf = null;
private MiniDFSCluster cluster;
@Test
public void testEnableRPCResponsePostponeHandling() throws IOException {
FileSystem fs = null;
Path file = null;
try {
FSNamesystem.SKIP_POSTPONE_HANDLE = true;
conf = new Configuration();
conf.setBoolean(DFSConfigKeys.DFS_NAMENODE_ASYNC_RESPONSE_ENABLED, true);
conf.setInt(CommonConfigurationKeys.IPC_CLIENT_RPC_TIMEOUT_KEY,
10000);
cluster = new MiniDFSCluster.Builder(conf).numDataNodes(3).build();
cluster.waitActive();
fs = cluster.getFileSystem();
file = new Path("/testEnableRPCResponsePostponeHandling");
DFSTestUtil.createFile(fs, file, 100, (short) 1, 12345L);
final DFSClient client = cluster.getFileSystem().getClient();
try {
client.getLocatedBlocks("/testEnableRPCResponsePostponeHandling", 0, 100);
Assert.fail("The call getLocatedBlocks should be timeout and failed.");
} catch (Exception e) {
LOG.error("Client getLocatedBlocks error", e);
Assert.assertTrue(e instanceof SocketTimeoutException);
}
} finally {
if (fs != null) {
fs.delete(file, true);
}
cluster.close();
}
}
@Test
public void testDisableRPCResponsePostponeHandling() throws IOException {
FileSystem fs = null;
Path file = null;
try {
FSNamesystem.SKIP_POSTPONE_HANDLE = false;
conf = new Configuration();
conf.setBoolean(DFSConfigKeys.DFS_NAMENODE_ASYNC_RESPONSE_ENABLED, true);
conf.setInt(CommonConfigurationKeys.IPC_CLIENT_RPC_TIMEOUT_KEY,
10000);
cluster = new MiniDFSCluster.Builder(conf).numDataNodes(3).build();
cluster.waitActive();
fs = cluster.getFileSystem();
file = new Path("/testDisableRPCResponsePostponeHandling");
DFSTestUtil.createFile(fs, file, 100, (short) 1, 12345L);
final DFSClient client = cluster.getFileSystem().getClient();
LocatedBlocks blocks = client.getLocatedBlocks(
"/testDisableRPCResponsePostponeHandling", 0, 100);
Assert.assertNotNull(blocks);
} finally {
if (fs != null) {
fs.delete(file, true);
}
cluster.close();
}
}
}
以上改动均在FSNamesystem类内。笔者在测试集群内已经可以跑通上述的异步response的改动,但是还没有在生产环境测试其实际的功效。本文多次提及的RPC postpone处理可详细参阅笔者之前的文章:基于RPC Call延时返回的HDFS异步editlog原理。RPC postpone的处理并不是完全没有缺点的,server延时response同时意味着server会延时释放与client的connection,简单来说就是server要hold住比之前更多的连接数了。