@Override
public void configure(JobConf conf) {
this.threshold = conf.getFloat(PARAM_APS_THRESHOLD, DEFAULT_THRESHOLD);
int reducerID = conf.getInt("mapred.task.partition", -1);
int max = conf.getInt(PARAM_APS_MAXKEY, 0);
int nstripes = conf.getInt(PARAM_APS_STRIPES, 1);
int spread = conf.getInt(PARAM_APS_REDUCER_PER_STRIPE, 1);
if (reducerID < 0 || max == 0) {
LOG.error("Could not find stripe ID, reverting to whole rest file loading");
LOG.debug("reducer = " + reducerID + "\t max = " + max + "\t nstripes = " + nstripes);
// open the pruned part file in the DistrubutedCache
haspruned = FileUtils.readRestFile(conf, pruned);
} else {
int stripe = GenericKey.StripePartitioner.findStripe(reducerID, spread);
int from = GenericKey.StripePartitioner.minKeyInStripe(stripe, nstripes, max);
int to = from + GenericKey.StripePartitioner.numKeysInStripe(stripe, nstripes, max);
// read from 'from' included, to 'to' excluded
LOG.info(
"Reducer "
+ reducerID
+ " loading stripe "
+ stripe
+ " of "
+ nstripes
+ " ("
+ from
+ ","
+ (to - 1)
+ ")");
haspruned = FileUtils.readRestFile(conf, pruned, from, to);
}
if (!haspruned) LOG.warn("No pruned file provided in DistributedCache");
else LOG.info("Read " + pruned.size() + " entries from pruned file");
}
@Override
public void configure(JobConf conf) {
threshold = conf.getFloat(PARAM_APS_THRESHOLD, DEFAULT_THRESHOLD);
mos = new MultipleOutputs(conf);
// open the maxWeight_i file in the DistributedCache
boolean succeded = FileUtils.readMaxWiFile(conf, maxWi);
if (!succeded) throw new AssertionError("Could not read maxWi file");
}
@Override
public int run(String[] args) throws IOException {
OptionParser p = new OptionParser();
OptionSpec<String> maxwiOpt =
p.accepts(maxwiOptName, "location of maxWi map file (HDFS) REQUIRED")
.withRequiredArg()
.ofType(String.class);
OptionSpec<Float> thresholdOpt =
p.accepts(thresholdOptName, "similarity threshold")
.withRequiredArg()
.ofType(Float.class)
.defaultsTo(DEFAULT_THRESHOLD);
OptionSpec<Integer> stripesOpt =
p.accepts(stripesOptName, "number of stripes to divide the similarity matrix")
.withRequiredArg()
.ofType(Integer.class)
.defaultsTo(1);
OptionSpec<Integer> spreadOpt =
p.accepts(spreadOptName, "number of reducers per stripe")
.withRequiredArg()
.ofType(Integer.class)
.defaultsTo(DEFAULT_SPREAD);
OptionSpec<Integer> factorOpt =
p.accepts(factorOptName, "number of mappers per reducer")
.withRequiredArg()
.ofType(Integer.class)
.defaultsTo(DEFAULT_FACTOR);
OptionSpec<Integer> maxVectorIDOpt =
p.accepts(maxVectorIDOptName, "maximum vector ID").withRequiredArg().ofType(Integer.class);
p.acceptsAll(Arrays.asList("h", "?"), "show help");
OptionSet options = parseOptions(p, args);
// to distinguish indexes built in successive runs
DateFormat df = new SimpleDateFormat("yyyyMMdd-HHmmss");
Date date = new Date();
float threshold = options.valueOf(thresholdOpt); // threshold
if (threshold < 0 || threshold >= 1) {
System.err.println(thresholdOptName + " should be between 0 and 1");
System.exit(1);
}
int numStripes = options.valueOf(stripesOpt); // number of stripes
if (numStripes < 1) {
System.err.println(stripesOptName + " should be > 0");
System.exit(1);
}
// MapReduce parameters
int spread = options.valueOf(spreadOpt); // how many reducers per stripe
if (spread < 1) {
System.err.println(spreadOptName + " should be > 0");
System.exit(1);
}
int factor = options.valueOf(factorOpt); // how many mappers per reducer
if (factor < 1) {
System.err.println(factorOptName + " should be > 0");
System.exit(1);
}
int maxKey = 0;
if (options.has(maxVectorIDOpt)) {
maxKey = options.valueOf(maxVectorIDOpt); // maximum value of the vector ID
if (maxKey < 1) {
System.err.println(maxVectorIDOptName + " should be > 0");
System.exit(1);
}
}
int numReducers = GenericKey.StripePartitioner.numReducers(numStripes, spread);
int numMappers = numReducers * factor;
int numBuckets = numMappers;
// pick the file with max weights from command line
String maxWiDir = options.valueOf(maxwiOpt);
List<String> nonOptArgs = options.nonOptionArguments();
LOG.info("Threshold set to " + threshold);
LOG.info(
String.format(
"Buckets: %1$-10s Factor: %2$-10s Stripes: %3$-10s Spread: %4$-10s Reducers: %5$-10s",
numBuckets, factor, numStripes, spread, numReducers));
// start building the jobs
JobConf conf1 = new JobConf(getConf(), Similarity.class);
conf1.setFloat(PARAM_APS_THRESHOLD, threshold);
conf1.setInt(PARAM_APS_STRIPES, numStripes);
DistributedCache.addCacheFile(URI.create(maxWiDir), conf1);
Path inputPath = new Path(nonOptArgs.get(0));
Path indexPath =
new Path(
nonOptArgs.get(0) + "-index-" + threshold + "-s" + numStripes + "_" + df.format(date));
// index filtering pruned nested directory
Path indexOnlyPath = new Path(indexPath, "part*");
Path outputPath = new Path(nonOptArgs.get(1) + "-" + threshold + "-s" + numStripes);
FileInputFormat.setInputPaths(conf1, inputPath);
FileOutputFormat.setOutputPath(conf1, indexPath);
conf1.setInputFormat(SequenceFileInputFormat.class);
conf1.setOutputFormat(SequenceFileOutputFormat.class);
conf1.setMapOutputKeyClass(LongWritable.class);
conf1.setMapOutputValueClass(IndexItem.class);
conf1.setOutputKeyClass(LongWritable.class);
conf1.setOutputValueClass(IndexItemArrayWritable.class);
conf1.setMapperClass(IndexerMapper.class);
conf1.setReducerClass(IndexerReducer.class);
// assuming input is sorted according to the key (vectorID) so that the
// part files are locally sorted
MultipleOutputs.addNamedOutput(
conf1,
PRUNED,
SequenceFileOutputFormat.class,
IntWritable.class,
VectorComponentArrayWritable.class);
// remove the stuff we added from the job name
conf1.set(
"mapred.job.name",
"APS-" + indexPath.getName().substring(0, indexPath.getName().length() - 16));
conf1.setNumTasksToExecutePerJvm(-1); // JVM reuse
conf1.setSpeculativeExecution(false);
conf1.setCompressMapOutput(true);
// hash the posting lists in different buckets to distribute the load
conf1.setNumReduceTasks(numBuckets);
RunningJob job1 = JobClient.runJob(conf1);
// part 2
JobConf conf2 = new JobConf(getConf(), Similarity.class);
if (numStripes > 0) FileUtils.mergeRestFile(conf2, indexPath, PRUNED, INDEX_INTERVAL);
MultipleInputs.addInputPath(
conf2, indexOnlyPath, SequenceFileInputFormat.class, SimilarityMapperIndex.class);
MultipleInputs.addInputPath(
conf2, inputPath, SequenceFileInputFormat.class, SimilarityMapperInput.class);
FileOutputFormat.setOutputPath(conf2, outputPath);
conf2.setCombinerClass(SimilarityCombiner.class);
conf2.setReducerClass(SimilarityReducer.class);
conf2.setPartitionerClass(GenericKey.StripePartitioner.class);
conf2.setOutputKeyComparatorClass(GenericKey.Comparator.class);
conf2.setOutputValueGroupingComparator(GenericKey.PrimaryComparator.class);
conf2.setMapOutputKeyClass(GenericKey.class);
conf2.setMapOutputValueClass(GenericValue.class);
conf2.setOutputKeyClass(VectorPair.class);
conf2.setOutputValueClass(NullWritable.class);
Counter numDocs =
job1.getCounters()
.findCounter("org.apache.hadoop.mapred.Task$Counter", "MAP_INPUT_RECORDS");
maxKey = maxKey > 0 ? maxKey : (int) numDocs.getValue();
LOG.info("Setting max key value in input to " + maxKey);
conf2.setInt(PARAM_APS_MAXKEY, maxKey);
conf2.setInt(PARAM_APS_STRIPES, numStripes);
conf2.setFloat(PARAM_APS_THRESHOLD, threshold);
conf2.setInt(PARAM_APS_REDUCER_PER_STRIPE, spread);
conf2.set("mapred.job.name", "APS-" + outputPath.getName());
conf2.setNumTasksToExecutePerJvm(-1); // JVM reuse
conf2.setSpeculativeExecution(false);
conf2.setCompressMapOutput(true);
conf2.setNumReduceTasks(numReducers);
JobClient.runJob(conf2);
return 0;
}