Mailing-List: contact hdfs-issues-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Reply-To: hdfs-issues@hadoop.apache.org
Date: Tue, 21 Jun 2011 04:05:47 +0000 (UTC)
From: "Alberto Andreotti (JIRA)" <jira@apache.org>
To: hdfs-issues@hadoop.apache.org
Message-ID: 
 <1602555239.23312.1308629147374.JavaMail.tomcat@hel.zones.apache.org>
Subject: [jira] [Created] (HDFS-2091) Hadoop does not scale as expected
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit

Hadoop does not scale as expected
---------------------------------

                 Key: HDFS-2091
                 URL: https://issues.apache.org/jira/browse/HDFS-2091
             Project: Hadoop HDFS
          Issue Type: Bug
         Environment: Linux, 8 nodes.
            Reporter: Alberto Andreotti


The more nodes I add to this application, the slower it goes. This is the app's map,

 public void map(IntWritable linearPos, FloatWritable heat, Context context
                            ) throws IOException, InterruptedException {
       int myLinearPos = linearPos.get();

       //Distribute my value to the previous and the next
       linearPos.set(myLinearPos - 1);
       context.write(linearPos, heat);
       linearPos.set(myLinearPos + 1);
       context.write(linearPos, heat);

       //Distribute my value to the cells above and below
       linearPos.set(myLinearPos - MatrixData.Length());
       context.write(linearPos, heat);
       linearPos.set(myLinearPos + MatrixData.Length());
       context.write(linearPos, heat);

    }//end map

and this is the reduce,

public void reduce(IntWritable linearPos, Iterable<FloatWritable> fwValues,
                     Context context) throws IOException, InterruptedException {

       //Handle first and last "cold" boundaries
       if(linearPos.get()<0 || linearPos.get()>MatrixData.LinearSize()){
          return;
       }

       if(linearPos.get()==MatrixData.HeatSourceLinearPos()){
          context.write(linearPos, new FloatWritable(MatrixData.HeatSourceTemperature()));
          return;
       }

       float result = 0.0f;
       //Add all the values
       for(FloatWritable heat : fwValues) {
          result += heat.get();
       }

      context.write(linearPos, new FloatWritable(result/4) );
}

For example, with 6 nodes I get a running time of 15minutes, and with 4 nodes I get a running time of 8minutes!.
This is how I generated the input,

 public static void main(String[] args) throws IOException {
     //Write file in the local dir
     String uri = "/home/beto/mySeq";

     Configuration conf = new Configuration();
     FileSystem fs = FileSystem.get(URI.create(uri), conf);
     Path path = new Path(uri);

     IntWritable key = new IntWritable();
     FloatWritable value = new FloatWritable(0.0f);

     SequenceFile.Writer writer = null;
     try {
       writer = SequenceFile.createWriter(fs, conf, path, key.getClass(), value.getClass());

     int step = MatrixData.LinearSize()/10;
     int limit = step;
     for (int i = 0; i <= MatrixData.LinearSize(); i++) {
        key.set(i);
        if(i>limit){
             System.out.println("*");
             limit +=step;
        }
          if(i==MatrixData.HeatSourceLinearPos()) {
            writer.append(key, new FloatWritable(MatrixData.HeatSourceTemperature()));
            continue;
          }

        writer.append(key, value);

      }
    } finally {
      IOUtils.closeStream(writer);
    }
  }


I'm basically solving a heat transfer problem in a squared section. Pretty simple. The input data is being stored as a (key, value) pairs, read in this way, processed, and written again in the same format.
Any thoughts?

Alberto.


--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira