Giraph源码分析（八）—— 统计每个SuperStep中参与计算的顶点数目

HamaWhite 原创，转载请注明出处！欢迎大家加入Giraph 技术交流群： 228591158

目的：科研中，需要分析在每次迭代过程中参与计算的顶点数目，来进一步优化系统。比如，在SSSP的compute()方法最后一行，都会把当前顶点voteToHalt，即变为InActive状态。所以每次迭代完成后，所有顶点都是InActive状态。在大同步后，收到消息的顶点会被激活，变为Active状态，然后调用顶点的compute()方法。本文的目的就是统计每次迭代过程中，参与计算的顶点数目。下面附上SSSP的compute()方法：

@Override
  public void compute(Iterable messages) {
    if (getSuperstep() == 0) {
      setValue(new DoubleWritable(Double.MAX_VALUE));
    }
    double minDist = isSource() ? 0d : Double.MAX_VALUE;
    for (DoubleWritable message : messages) {
      minDist = Math.min(minDist, message.get());
    }
    if (minDist < getValue().get()) {
      setValue(new DoubleWritable(minDist));
      for (Edge edge : getEdges()) {
        double distance = minDist + edge.getValue().get();
        sendMessage(edge.getTargetVertexId(), new DoubleWritable(distance));
      }
    }
	//把顶点置为InActive状态
    voteToHalt();
  }

附：giraph中算法的终止条件是：没有活跃顶点且worker间没有消息传递。

hama-0.6.0中算法的终止条件只是：判断是否有活跃顶点。不是真正的pregel思想，半成品。

修改过程如下：

1. org.apache.giraph.partition. PartitionStats 类

添加变量和方法，用来统计每个Partition在每个超步中参与计算的顶点数目。添加的变量和方法如下：

/** computed vertices in this partition */
private long computedVertexCount=0;

/**
* Increment the computed vertex count by one.
*/
public void incrComputedVertexCount() {
    ++ computedVertexCount;
}

/**
 * @return the computedVertexCount
 */
public long getComputedVertexCount() {
	return computedVertexCount;
}

修改readFields()和write()方法，每个方法追加最后一句。当每个Partition计算完成后，会把自己的computedVertexCount发送给Master，Mater再读取汇总。

@Override
public void readFields(DataInput input) throws IOException {
    partitionId = input.readInt();
    vertexCount = input.readLong();
    finishedVertexCount = input.readLong();
    edgeCount = input.readLong();
    messagesSentCount = input.readLong();
    //添加下条语句
    computedVertexCount=input.readLong();
}

@Override
public void write(DataOutput output) throws IOException {
    output.writeInt(partitionId);
    output.writeLong(vertexCount);
    output.writeLong(finishedVertexCount);
    output.writeLong(edgeCount);
    output.writeLong(messagesSentCount);
    //添加下条语句
    output.writeLong(computedVertexCount);
}

2. org.apache.giraph.graph. GlobalStats 类

  添加变量和方法，用来统计每个超步中参与计算的顶点总数目，包含每个Worker上的所有Partitions。

 /** computed vertices in this partition 
  *  Add by BaiSong 
  */
  private long computedVertexCount=0;
	 /**
	 * @return the computedVertexCount
	 */
	public long getComputedVertexCount() {
		return computedVertexCount;
	}

修改addPartitionStats(PartitionStats partitionStats)方法，增加统计computedVertexCount功能。

/**
  * Add the stats of a partition to the global stats.
  *
  * @param partitionStats Partition stats to be added.
  */
  public void addPartitionStats(PartitionStats partitionStats) {
    this.vertexCount += partitionStats.getVertexCount();
    this.finishedVertexCount += partitionStats.getFinishedVertexCount();
    this.edgeCount += partitionStats.getEdgeCount();
    //Add by BaiSong，添加下条语句
    this.computedVertexCount+=partitionStats.getComputedVertexCount();
 }

当然为了Debug方便，也可以修改该类的toString()方法（可选），修改后的如下：

public String toString() {
		return "(vtx=" + vertexCount + ", computedVertexCount="
				+ computedVertexCount + ",finVtx=" + finishedVertexCount
				+ ",edges=" + edgeCount + ",msgCount=" + messageCount
				+ ",haltComputation=" + haltComputation + ")";
	}

3. org.apache.giraph.graph. ComputeCallable

添加统计功能。在computePartition()方法中，添加下面一句。

if (!vertex.isHalted()) {
        context.progress();
        TimerContext computeOneTimerContext = computeOneTimer.time();
        try {
            vertex.compute(messages);
	    //添加下面一句，当顶点调用完compute()方法后，就把该Partition的computedVertexCount加1
            partitionStats.incrComputedVertexCount();
        } finally {
           computeOneTimerContext.stop();
        }
……

4. 添加Counters统计，和我的博客 Giraph源码分析（七）—— 添加消息统计功能 类似，此处不再详述。添加的类为：org.apache.giraph.counters.GiraphComputedVertex，下面附上该类的源码：

package org.apache.giraph.counters;

import java.util.Iterator;
import java.util.Map;

import org.apache.hadoop.mapreduce.Mapper.Context;
import com.google.common.collect.Maps;

/**
 * Hadoop Counters in group "Giraph Messages" for counting every superstep
 * message count.
 */

public class GiraphComputedVertex extends HadoopCountersBase {
	/** Counter group name for the giraph Messages */
	public static final String GROUP_NAME = "Giraph Computed Vertex";

	/** Singleton instance for everyone to use */
	private static GiraphComputedVertex INSTANCE;

	/** superstep time in msec */
	private final Map superstepVertexCount;

	private GiraphComputedVertex(Context context) {
		super(context, GROUP_NAME);
		superstepVertexCount = Maps.newHashMap();
	}

	/**
	 * Instantiate with Hadoop Context.
	 * 
	 * @param context
	 *            Hadoop Context to use.
	 */
	public static void init(Context context) {
		INSTANCE = new GiraphComputedVertex(context);
	}

	/**
	 * Get singleton instance.
	 * 
	 * @return singleton GiraphTimers instance.
	 */
	public static GiraphComputedVertex getInstance() {
		return INSTANCE;
	}

	/**
	 * Get counter for superstep messages
	 * 
	 * @param superstep
	 * @return
	 */
	public GiraphHadoopCounter getSuperstepVertexCount(long superstep) {
		GiraphHadoopCounter counter = superstepVertexCount.get(superstep);
		if (counter == null) {
			String counterPrefix = "Superstep: " + superstep+" ";
			counter = getCounter(counterPrefix);
			superstepVertexCount.put(superstep, counter);
		}
		return counter;
	}

	@Override
	public Iterator iterator() {
		return superstepVertexCount.values().iterator();
	}
}

5. 实验结果，运行程序后。会在终端输出每次迭代参与计算的顶点总数目。 测试SSSP（SimpleShortestPathsVertex类），输入图中共有9个顶点和12条边。输出结果如下：

上图测试中，共有6次迭代。红色框中，显示出了每次迭代过冲参与计算的顶点数目，依次是：9,4,4,3,4,0

解释：在第0个超步，每个顶点都是活跃的，所有共有9个顶点参与计算。在第5个超步，共有0个顶点参与计算，那么就不会向外发送消息，加上每个顶点都是不活跃的，所以算法迭代终止。

完！

本人原创，转载请注明出处！ 本人QQ：530422429，欢迎大家指正、讨论。