Flume性能优化实践

Flume性能优化实践

发表评论

最近公司落地Flume日志采集着实反复了好久,简单记录一下性能优化的核心思路。

初始配置所有batch size、transaction size都是1000,channel的capactiy是10000。

版本一

最初我是按Memory Channel做压测,Taildir的source采集增量日志,Memory Channel缓冲数据,Kafka Sink发送数据。

这里面的瓶颈是Kafka Sink,因为Kafka Sink是单线程同步发送,网络延迟就会导致吞吐上不去,大概10MB+的一个吞吐就封顶了。

版本二

翻看了官方文档,打算试验一下sink group来实现多个kafka sink同时发送,结果性能仍旧10MB+。

分析原理,原来sink group仍旧是个单线程sink,相当于多个kafka sink的代理而已,仅仅实现了轮转负载均衡功能。

一个kafka sink的发送延迟高,轮转压根没有意义。

版本三

于是琢磨如何实现多线程跑多个Kafka Sink,于是仍旧使用1个Memory Channel,配置对应3个Kafka Sink,结果带宽可以升高到30MB的样子,但是极不稳定,来回跳跃。

此时发现Memory Channel的填充率接近90%+,应该是因为容量经常塞满导致的流水线阻塞,通过增加memory channel的capacity到10万,batch size和transaction size增加到1万,吞吐提升到60MB~80MB+,填充率小于10%,已经满足需求。

在transaction size=1000的情况下memory channel被填满,而transaction size=1万的情况下memory channel就不会被填满,其实是通过增加channel批处理的包大小,降低了channel访问的频次,解决的是memory channel的锁瓶颈。

同时,这个优化思路也带来了问题,更大的memory channel capacity带来了更大的数据丢失风险,因为宕机时memory channel里缓冲的数据都会丢失。

版本四

实现多个memory channel轮转,每个memory channel由一个kafka sink消费。

这样做目的有2个:

  1. 由多个sink竞争消费1个channel改为各自消费1个channel,锁瓶颈解决。
  2. 因为锁瓶颈变小,所以可以仍旧保持较小的channel capacity来保障数据可靠性,比如每个channel容量10000,那么3个channel丢失3万,仍旧优于”版本三”。

实现该功能需要自己开发channel selector插件,实现source流量的轮转分发,可以翻看我之前写的博客。

版本五

同事要求使用file channel,保障队列中数据的可靠性,但是经过测试发现吞吐只能跑到10MB左右,上述所说优化手段均无效。

更换SSD盘也没有带来任何提升,File channel自身填充率极低。

个人怀疑瓶颈在File Channel自身,其事务的提交效率太低,阻塞了source的投递动作,无论如何增加channel数量也无济于事,因为source是单线程的,轮转发往多个File Channel的速度仍旧等于单个File Channel速度,导致后续Sink没有足够数据消费,吞吐无法提升。

从FileChannel代码来看,磁盘读写的相关代码全部被加锁处理:

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

    synchronized FlumeEventPointer put(ByteBuffer buffer) throws IOException {

      if (encryptor != null) {

        buffer = ByteBuffer.wrap(encryptor.encrypt(buffer.array()));

      }

      Pair<Integer, Integer> pair = write(buffer);

      return new FlumeEventPointer(pair.getLeft(), pair.getRight());

    }

    synchronized void take(ByteBuffer buffer) throws IOException {

      if (encryptor != null) {

        buffer = ByteBuffer.wrap(encryptor.encrypt(buffer.array()));

      }

      write(buffer);

    }

    synchronized void rollback(ByteBuffer buffer) throws IOException {

      if (encryptor != null) {

        buffer = ByteBuffer.wrap(encryptor.encrypt(buffer.array()));

      }

      write(buffer);

    }

    synchronized void commit(ByteBuffer buffer) throws IOException {

      if (encryptor != null) {

        buffer = ByteBuffer.wrap(encryptor.encrypt(buffer.array()));

      }

      write(buffer);

      dirty = true;

      lastCommitPosition = position();

    }

另外,日志文件的sync刷盘策略分为两种选项,一种是每次提交事务都刷新,另外一个是定时线程刷新(下面是定时线程):

1

2

3

4

5

6

7

8

9

10

11

        syncExecutor.scheduleWithFixedDelay(new Runnable() {

          @Override

          public void run() {

            try {

              sync();

            } catch (Throwable ex) {

              LOG.error("Data file, " + getFile().toString() + " could not " +

                  "be synced to disk due to an error.", ex);

            }

          }

        }, fsyncInterval, fsyncInterval, TimeUnit.SECONDS);

而这个sync()刷盘操作同样被锁保护的,会占用大量的锁时间:

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

    /**

     * Sync the underlying log file to disk. Expensive call,

     * should be used only on commits. If a sync has already happened after

     * the last commit, this method is a no-op

     *

     * @throws IOException

     * @throws LogFileRetryableIOException - if this log file is closed.

     */

    synchronized void sync() throws IOException {

      if (!fsyncPerTransaction && !dirty) {

        if (LOG.isDebugEnabled()) {

          LOG.debug(

              "No events written to file, " + getFile().toString() +

                  " in last " + fsyncInterval + " or since last commit.");

        }

        return;

      }

      if (!isOpen()) {

        throw new LogFileRetryableIOException("File closed " + file);

      }

      if (lastSyncPosition < lastCommitPosition) {

        getFileChannel().force(false);

        lastSyncPosition = position();

        syncCount++;

        dirty = false;

      }

    }

降低sync()的调用频率,理论上可以降低锁占用时间,让出更多的锁时间给put与take操作。

flume可以配置这些参数,只是官方文档里并没有说明:

1

2

3

4

5

  public static final String FSYNC_PER_TXN = "fsyncPerTransaction";

  public static final boolean DEFAULT_FSYNC_PRE_TXN = true;

  public static final String FSYNC_INTERVAL = "fsyncInterval";

  public static final int DEFAULT_FSYNC_INTERVAL = 5; // seconds.

默认是每个事务都sync,这样当然是为了保障数据可靠性,否则也就没必要用FileChannel了。

我尝试改成了定时sync(),发现吞吐仍旧无法提升,那么我继续猜测问题在于事务的commit部分,也就是Sink做的事情:

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

/**

   * Synchronization not required as this method is atomic

   *

   * @param transactionID

   * @param type

   * @throws IOException

   */

  private void commit(long transactionID, short type) throws IOException {

    Preconditions.checkState(open, "Log is closed");

    Commit commit = new Commit(transactionID, WriteOrderOracle.next(), type);

    ByteBuffer buffer = TransactionEventRecord.toByteBuffer(commit);

    int logFileIndex = nextLogWriter(transactionID);

    long usableSpace = logFiles.get(logFileIndex).getUsableSpace();

    long requiredSpace = minimumRequiredSpace + buffer.limit();

    if (usableSpace <= requiredSpace) {

      throw new IOException("Usable space exhausted, only " + usableSpace +

          " bytes remaining, required " + requiredSpace + " bytes");

    }

    boolean error = true;

    try {

      try {

        LogFile.Writer logFileWriter = logFiles.get(logFileIndex);

        // If multiple transactions are committing at the same time,

        // this ensures that the number of actual fsyncs is small and a

        // number of them are grouped together into one.

        logFileWriter.commit(buffer);

        logFileWriter.sync();

        error = false;

      } catch (LogFileRetryableIOException e) {

        if (!open) {

          throw e;

        }

        roll(logFileIndex, buffer);

        LogFile.Writer logFileWriter = logFiles.get(logFileIndex);

        logFileWriter.commit(buffer);

        logFileWriter.sync();

        error = false;

      }

提交事务也只是写入一条日志标记对应的事务完结了,这样宕机重放日志时就会跳过该事务。

我们发现这个操作总是sync(),虽然这个操作不需要锁保护的样子,但是它占用了sink线程的时间,估计吞吐无法提升也离不开它的关系。

关于File Channel瓶颈,有同学有JAVA调优经验的可以具体给FileChannel加一些调试日志,看看到底慢在哪个环节。

我个人会优先选择使用capacity较小(1万-10万)的的memory channel配合多个sink来实现高吞吐,至于对宕机的那点担心实在没有必要,因为大多数时候memory channel的填充率不足1%,也就是丢失10万*0.01=100条而已。

猜你喜欢

转载自blog.csdn.net/wangshuminjava/article/details/83931340