多region下的hbase写入问题

    原url地址为：http://koven2049.iteye.com/blog/1144526

    最近在集群上发现hbase写入性能受到较大下降，测试环境下没有该问题产生。而生产环境和测试环境的区别之一是生产环境的region数量远远多于测试环境，单台regionserver服务了约3500个region。
    通过jstack工具检查到大半写入线程BLOCKED状态在"public synchronized void reclaimMemStoreMemory() {"这一行，这是在put之前的一个检查过程。
    hbase在每次put以前，需要检查当前regionserver上的memstore是否超过总memstore阀值，如果超过，需要block住当前的写入，防止OOM，代码片段见下：

/**
   * Check if the regionserver's memstore memory usage is greater than the
   * limit. If so, flush regions with the biggest memstores until we're down
   * to the lower limit. This method blocks callers until we're down to a safe
   * amount of memstore consumption.
   */
  public synchronized void reclaimMemStoreMemory() {
    if (isAboveHighWaterMark()) {
      lock.lock();
      try {
        while (isAboveHighWaterMark() && !server.isStopped()) {
          wakeupFlushThread();
          try {
            // we should be able to wait forever, but we've seen a bug where
            // we miss a notify, so put a 5 second bound on it at least.
            flushOccurred.await(5, TimeUnit.SECONDS);
          } catch (InterruptedException ie) {
            Thread.currentThread().interrupt();
          }
        }
      } finally {
        lock.unlock();
      }
    } else if (isAboveLowWaterMark()) {
      wakeupFlushThread();
    }
  }

/**
   * Check if the regionserver's memstore memory usage is greater than the
   * limit. If so, flush regions with the biggest memstores until we're down
   * to the lower limit. This method blocks callers until we're down to a safe
   * amount of memstore consumption.
   */
  public synchronized void reclaimMemStoreMemory() {
    if (isAboveHighWaterMark()) {
      lock.lock();
      try {
        while (isAboveHighWaterMark() && !server.isStopped()) {
          wakeupFlushThread();
          try {
            // we should be able to wait forever, but we've seen a bug where
            // we miss a notify, so put a 5 second bound on it at least.
            flushOccurred.await(5, TimeUnit.SECONDS);
          } catch (InterruptedException ie) {
            Thread.currentThread().interrupt();
          }
        }
      } finally {
        lock.unlock();
      }
    } else if (isAboveLowWaterMark()) {
      wakeupFlushThread();
    }
  }

    这是一个同步操作，其中isAboveHighWaterMark()的代码如下：

private boolean isAboveHighWaterMark() {
  return server.getGlobalMemStoreSize() >= globalMemStoreLimit;
}

  private boolean isAboveHighWaterMark() {
    return server.getGlobalMemStoreSize() >= globalMemStoreLimit;
  }

    getGlobalMemStoreSize()里面的操作是遍历所有region，拿到每个region的memstore大小:

public long getGlobalMemStoreSize() {
  long total = 0;
  for (HRegion region : onlineRegions.values()) {
    total += region.memstoreSize.get();
  }
  return total;
}

  
  
  
  

   
   
   
   Java代码 
   
   
   
    
   
   
   
   
   
   
   
  
  
  
  

public long getGlobalMemStoreSize() {
  long total = 0;
  for (HRegion region : onlineRegions.values()) {
    total += region.memstoreSize.get();
  }
  return total;
}

 

    如果region数量很多就比较杯具了，在单台服务器3500个region的环境下通过btrace跟踪到这一步需要耗时0.4ms，也就是每一个put会block所有线程0.4ms，这样无法发挥出server端并行处理能力，同时可以计算出无论如何配置，写tps无法超过1000/0.4=2500!
    产生这个问题的根本原因是在0.90.x版本中，region无法拿到regionserver的信息，因此只能通过实时计算来得到rs上总的memstore大小。在0.92.0或trunk版本中修改了HRegion的数据结构，让HRegion在初始化时得到regionserver的信息，因此可以实时记录memstore的总大小，并让每个region对象能拿到该值，于是这个isAboveHighWaterMark()就不再block住所有线程了，在region较多的场景下写性能得到较大提升。
    参照 https://issues.apache.org/jira/browse/HBASE-3694，可以以此patch为参考生成相应的0.90.x版本的patch