List of usage examples for org.apache.hadoop.mapred.lib TotalOrderPartitioner getPartitionFile
@Deprecated public static String getPartitionFile(JobConf job)
From source file:com.alexholmes.hadooputils.sort.SortInputSampler.java
License:Apache License
public static <K, V> void writePartitionFile(JobConf job, Sampler<K, V> sampler) throws IOException { Configuration conf = job;/*from w ww.jav a 2 s . c o m*/ // Use the input format defined in the job. NOT, the one provided by // the parent class's writePartitionFile() method, which will be a plain // TextInputFormat, by default final InputFormat inf = job.getInputFormat(); int numPartitions = job.getNumReduceTasks(); K[] samples = (K[]) sampler.getSample(inf, job); RawComparator<K> comparator = (RawComparator<K>) job.getOutputKeyComparator(); Arrays.sort(samples, comparator); Path dst = new Path(TotalOrderPartitioner.getPartitionFile(job)); FileSystem fs = dst.getFileSystem(conf); if (fs.exists(dst)) { fs.delete(dst, false); } SequenceFile.Writer writer = SequenceFile.createWriter(fs, conf, dst, job.getMapOutputKeyClass(), NullWritable.class); NullWritable nullValue = NullWritable.get(); float stepSize = samples.length / (float) numPartitions; int last = -1; for (int i = 1; i < numPartitions; ++i) { int k = Math.round(stepSize * i); while (last >= k && comparator.compare(samples[last], samples[k]) == 0) { ++k; } writer.append(samples[k], nullValue); last = k; } writer.close(); }
From source file:com.ebay.erl.mobius.core.mapred.ConfigurableJob.java
License:Apache License
@Override protected synchronized void submit() { JobConf jobConf = this.getJobConf(); boolean isLocalHadoop = jobConf.get("mapred.job.tracker", "local").equals("local"); // the default partitioner is {@link com.ebay.erl.mobius.core.datajoin.DataJoinKeyPartitioner} // which is hash based. //// w w w .j a v a 2 s .c om // If user choose to use even partitioner, Mobius will use // {@link com.ebay.erl.mobius.core.datajoin.EvenlyPartitioner} which // is sampling based partitioner of attempting to balance the load // for each reducer. String partitioner = jobConf.get("mobius.partitioner", "default"); if (!isLocalHadoop && jobConf.getNumReduceTasks() != 0 && partitioner.equals("even")) { // this job needs reducer, perform sampling on the keys to // make load on reducers are almost evenly distributed. double freq = jobConf.getFloat("mobius.sampler.freq", 0.1F); int numSamples = jobConf.getInt("mobius.sampler.num.samples", 50000); int maxSplits = jobConf.getInt("mobius.sampler.max.slipts.sampled", 5); // log sampling parameters so that user knows. LOGGER.info("Sampling parameters { " + "mobius.sampler.freq:" + format.format(freq) + ", " + "mobius.sampler.num.samples:" + numSamples + ", " + "mobius.sampler.max.slipts.sampled:" + maxSplits + "}"); InputSampler.Sampler<?, ?> sampler = new MobiusInputSampler(freq, numSamples, maxSplits); writePartitionFile(jobConf, sampler); // add to distributed cache try { URI partitionUri = new URI(TotalOrderPartitioner.getPartitionFile(jobConf) + "#_partitions"); LOGGER.info("Adding partition uri to distributed cache:" + partitionUri.toString()); DistributedCache.addCacheFile(partitionUri, jobConf); DistributedCache.createSymlink(jobConf); jobConf.setPartitionerClass(EvenlyPartitioner.class); LOGGER.info("Using " + EvenlyPartitioner.class.getCanonicalName() + " to partiton the keys evenly among reducers."); } catch (URISyntaxException e) { LOGGER.error(e.getMessage(), e); throw new RuntimeException(e); } // adding -XX:-UseParallelOldGC, this will automatically set -XX:-UseParallelGC // according to Oracle's specification String jvmOpts = jobConf.get("mapred.child.java.opts", ""); if (jvmOpts.isEmpty()) { jvmOpts = "-XX:-UseParallelOldGC"; } else { if (jvmOpts.indexOf("-XX:-UseParallelOldGC") < 0) { // remove " jvmOpts = jvmOpts.replaceAll("\"", ""); jvmOpts = jvmOpts.concat(" -XX:-UseParallelOldGC"); } } jobConf.set("mapred.child.java.opts", jvmOpts); this.setJobConf(jobConf); } LOGGER.info("Submiting job:" + jobConf.getJobName()); super.submit(); }
From source file:com.ebay.erl.mobius.core.mapred.ConfigurableJob.java
License:Apache License
private static void writePartitionFile(JobConf job, Sampler sampler) { try {//ww w. ja v a2s. c o m //////////////////////////////////////////////// // first, getting samples from the data sources //////////////////////////////////////////////// LOGGER.info("Running local sampling for job [" + job.getJobName() + "]"); InputFormat inf = job.getInputFormat(); Object[] samples = sampler.getSample(inf, job); LOGGER.info("Samples retrieved, sorting..."); //////////////////////////////////////////////// // sort the samples //////////////////////////////////////////////// RawComparator comparator = job.getOutputKeyComparator(); Arrays.sort(samples, comparator); if (job.getBoolean("mobius.print.sample", false)) { PrintWriter pw = new PrintWriter( new OutputStreamWriter(new GZIPOutputStream(new BufferedOutputStream(new FileOutputStream( new File(job.get("mobius.sample.file", "./samples.txt.gz"))))))); for (Object obj : samples) { pw.println(obj); } pw.flush(); pw.close(); } //////////////////////////////////////////////// // start to write partition files //////////////////////////////////////////////// FileSystem fs = FileSystem.get(job); Path partitionFile = fs.makeQualified(new Path(TotalOrderPartitioner.getPartitionFile(job))); while (fs.exists(partitionFile)) { partitionFile = new Path(partitionFile.toString() + "." + System.currentTimeMillis()); } fs.deleteOnExit(partitionFile); TotalOrderPartitioner.setPartitionFile(job, partitionFile); LOGGER.info("write partition file to:" + partitionFile.toString()); int reducersNbr = job.getNumReduceTasks(); Set<Object> wroteSamples = new HashSet<Object>(); SequenceFile.Writer writer = SequenceFile.createWriter(fs, job, partitionFile, Tuple.class, NullWritable.class); float avgReduceSize = samples.length / reducersNbr; int lastBegin = 0; for (int i = 0; i < samples.length;) { // trying to distribute the load for every reducer evenly, // dividing the <code>samples</code> into a set of blocks // separated by boundaries, objects that selected from the // <code>samples</code> array, and each blocks should have // about the same size. // find the last index of element that equals to samples[i], as // such element might appear multiple times in the samples. int upperBound = Util.findUpperBound(samples, samples[i], comparator); int lowerBound = i;//Util.findLowerBound(samples, samples[i], comparator); // the repeat time of samples[i], if the key itself is too big // select it as boundary int currentElemSize = upperBound - lowerBound + 1; if (currentElemSize > avgReduceSize * 2) // greater than two times of average reducer size { // the current element is too big, greater than // two times of the <code>avgReduceSize</code>, // put itself as boundary writer.append(((DataJoinKey) samples[i]).getKey(), NullWritable.get()); wroteSamples.add(((DataJoinKey) samples[i]).getKey()); //pw.println(samples[i]); // immediate put the next element to the boundary, // the next element starts at <code> upperBound+1 // </code>, to prevent the current one consume even // more. if (upperBound + 1 < samples.length) { writer.append(((DataJoinKey) samples[upperBound + 1]).getKey(), NullWritable.get()); wroteSamples.add(((DataJoinKey) samples[upperBound + 1]).getKey()); //pw.println(samples[upperBound+1]); // move on to the next element of <code>samples[upperBound+1]/code> lastBegin = Util.findUpperBound(samples, samples[upperBound + 1], comparator) + 1; i = lastBegin; } else { break; } } else { // current element is small enough to be consider // with previous group int size = upperBound - lastBegin; if (size > avgReduceSize) { // by including the current elements, we have // found a block that's big enough, select it // as boundary writer.append(((DataJoinKey) samples[i]).getKey(), NullWritable.get()); wroteSamples.add(((DataJoinKey) samples[i]).getKey()); //pw.println(samples[i]); i = upperBound + 1; lastBegin = i; } else { i = upperBound + 1; } } } writer.close(); // if the number of wrote samples doesn't equals to number of // reducer minus one, then it means the key spaces is too small // hence TotalOrderPartitioner won't work, it works only if // the partition boundaries are distinct. // // we need to change the number of reducers if (wroteSamples.size() + 1 != reducersNbr) { LOGGER.info("Write complete, but key space is too small, sample size=" + wroteSamples.size() + ", reducer size:" + (reducersNbr)); LOGGER.info("Set the reducer size to:" + (wroteSamples.size() + 1)); // add 1 because the wrote samples define boundary, ex, if // the sample size is two with two element [300, 1000], then // there should be 3 reducers, one for handling i<300, one // for n300<=i<1000, and another one for 1000<=i job.setNumReduceTasks((wroteSamples.size() + 1)); } samples = null; } catch (IOException e) { LOGGER.error(e.getMessage(), e); throw new RuntimeException(e); } }