Search in sources :

Example 1 with Cell

use of org.apache.sysml.runtime.controlprogram.parfor.util.Cell in project incubator-systemml by apache.

the class ResultMergeLocalFile method createBinaryCellStagingFile.

@SuppressWarnings("deprecation")
private static void createBinaryCellStagingFile(String fnameStaging, MatrixObject mo, long ID) throws IOException, DMLRuntimeException {
    JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
    Path path = new Path(mo.getFileName());
    FileSystem fs = IOUtilFunctions.getFileSystem(path, job);
    LinkedList<Cell> buffer = new LinkedList<>();
    MatrixIndexes key = new MatrixIndexes();
    MatrixCell value = new MatrixCell();
    MatrixCharacteristics mc = mo.getMatrixCharacteristics();
    int brlen = mc.getRowsPerBlock();
    int bclen = mc.getColsPerBlock();
    for (Path lpath : IOUtilFunctions.getSequenceFilePaths(fs, path)) {
        SequenceFile.Reader reader = new SequenceFile.Reader(fs, lpath, job);
        try {
            while (reader.next(key, value)) {
                Cell tmp = new Cell(key.getRowIndex(), key.getColumnIndex(), value.getValue());
                buffer.addLast(tmp);
                if (// periodic flush
                buffer.size() > StagingFileUtils.CELL_BUFFER_SIZE) {
                    appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                    buffer.clear();
                }
            }
            // final flush
            if (!buffer.isEmpty()) {
                appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                buffer.clear();
            }
        } finally {
            IOUtilFunctions.closeSilently(reader);
        }
    }
}
Also used : Path(org.apache.hadoop.fs.Path) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) RecordReader(org.apache.hadoop.mapred.RecordReader) LinkedList(java.util.LinkedList) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) SequenceFile(org.apache.hadoop.io.SequenceFile) FileSystem(org.apache.hadoop.fs.FileSystem) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) JobConf(org.apache.hadoop.mapred.JobConf) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Cell(org.apache.sysml.runtime.controlprogram.parfor.util.Cell)

Example 2 with Cell

use of org.apache.sysml.runtime.controlprogram.parfor.util.Cell in project incubator-systemml by apache.

the class ResultMergeLocalFile method createTextCellStagingFile.

private static void createTextCellStagingFile(String fnameStaging, MatrixObject mo, long ID) throws IOException, DMLRuntimeException {
    JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
    Path path = new Path(mo.getFileName());
    FileInputFormat.addInputPath(job, path);
    TextInputFormat informat = new TextInputFormat();
    informat.configure(job);
    InputSplit[] splits = informat.getSplits(job, 1);
    LinkedList<Cell> buffer = new LinkedList<>();
    LongWritable key = new LongWritable();
    Text value = new Text();
    MatrixCharacteristics mc = mo.getMatrixCharacteristics();
    int brlen = mc.getRowsPerBlock();
    int bclen = mc.getColsPerBlock();
    // long row = -1, col = -1; //FIXME needs reconsideration whenever textcell is used actively
    // NOTE MB: Originally, we used long row, col but this led reproducibly to JIT compilation
    // errors during runtime; experienced under WINDOWS, Intel x86-64, IBM JDK 64bit/32bit.
    // It works fine with int row, col but we require long for larger matrices.
    // Since, textcell is never used for result merge (hybrid/hadoop: binaryblock, singlenode:binarycell)
    // we just propose the to exclude it with -Xjit:exclude={package.method*}(count=0,optLevel=0)
    FastStringTokenizer st = new FastStringTokenizer(' ');
    for (InputSplit split : splits) {
        RecordReader<LongWritable, Text> reader = informat.getRecordReader(split, job, Reporter.NULL);
        try {
            while (reader.next(key, value)) {
                // reset tokenizer
                st.reset(value.toString());
                long row = st.nextLong();
                long col = st.nextLong();
                double lvalue = Double.parseDouble(st.nextToken());
                Cell tmp = new Cell(row, col, lvalue);
                buffer.addLast(tmp);
                if (// periodic flush
                buffer.size() > StagingFileUtils.CELL_BUFFER_SIZE) {
                    appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                    buffer.clear();
                }
            }
            // final flush
            if (!buffer.isEmpty()) {
                appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                buffer.clear();
            }
        } finally {
            IOUtilFunctions.closeSilently(reader);
        }
    }
}
Also used : Path(org.apache.hadoop.fs.Path) Text(org.apache.hadoop.io.Text) LinkedList(java.util.LinkedList) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) FastStringTokenizer(org.apache.sysml.runtime.util.FastStringTokenizer) TextInputFormat(org.apache.hadoop.mapred.TextInputFormat) LongWritable(org.apache.hadoop.io.LongWritable) JobConf(org.apache.hadoop.mapred.JobConf) InputSplit(org.apache.hadoop.mapred.InputSplit) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Cell(org.apache.sysml.runtime.controlprogram.parfor.util.Cell)

Example 3 with Cell

use of org.apache.sysml.runtime.controlprogram.parfor.util.Cell in project incubator-systemml by apache.

the class DataPartitionerLocal method appendCellBufferToStagingArea.

private void appendCellBufferToStagingArea(String dir, LinkedList<Cell> buffer, int brlen, int bclen) throws IOException {
    HashMap<Long, LinkedList<Cell>> sortedBuffer = new HashMap<>();
    // sort cells in buffer wrt key
    long key = -1;
    for (Cell c : buffer) {
        switch(_format) {
            case ROW_WISE:
                key = c.getRow();
                c.setRow(1);
                break;
            case ROW_BLOCK_WISE:
                key = (c.getRow() - 1) / brlen + 1;
                c.setRow((c.getRow() - 1) % brlen + 1);
                break;
            case COLUMN_WISE:
                key = c.getCol();
                c.setCol(1);
                break;
            case COLUMN_BLOCK_WISE:
                key = (c.getCol() - 1) / bclen + 1;
                c.setCol((c.getCol() - 1) % bclen + 1);
                break;
            default:
        }
        if (!sortedBuffer.containsKey(key))
            sortedBuffer.put(key, new LinkedList<Cell>());
        sortedBuffer.get(key).addLast(c);
    }
    // write lists of cells to local files
    for (Entry<Long, LinkedList<Cell>> e : sortedBuffer.entrySet()) {
        String pdir = LocalFileUtils.checkAndCreateStagingDir(dir + "/" + e.getKey());
        String pfname = pdir + "/" + "block_" + _seq.getNextID();
        StagingFileUtils.writeCellListToLocal(pfname, e.getValue());
    }
}
Also used : HashMap(java.util.HashMap) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Cell(org.apache.sysml.runtime.controlprogram.parfor.util.Cell) LinkedList(java.util.LinkedList)

Example 4 with Cell

use of org.apache.sysml.runtime.controlprogram.parfor.util.Cell in project incubator-systemml by apache.

the class DataPartitionerLocal method partitionTextCell.

private void partitionTextCell(String fname, String fnameStaging, String fnameNew, long rlen, long clen, int brlen, int bclen) {
    long row = -1;
    long col = -1;
    try {
        // STEP 1: read matrix from HDFS and write blocks to local staging area
        // check and add input path
        JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
        Path path = new Path(fname);
        FileInputFormat.addInputPath(job, path);
        TextInputFormat informat = new TextInputFormat();
        informat.configure(job);
        InputSplit[] splits = informat.getSplits(job, 1);
        LinkedList<Cell> buffer = new LinkedList<>();
        LongWritable key = new LongWritable();
        Text value = new Text();
        FastStringTokenizer st = new FastStringTokenizer(' ');
        for (InputSplit split : splits) {
            RecordReader<LongWritable, Text> reader = informat.getRecordReader(split, job, Reporter.NULL);
            try {
                while (reader.next(key, value)) {
                    // reset tokenizer
                    st.reset(value.toString());
                    row = st.nextLong();
                    col = st.nextLong();
                    double lvalue = st.nextDouble();
                    Cell tmp = new Cell(row, col, lvalue);
                    buffer.addLast(tmp);
                    if (// periodic flush
                    buffer.size() > StagingFileUtils.CELL_BUFFER_SIZE) {
                        appendCellBufferToStagingArea(fnameStaging, buffer, brlen, bclen);
                        buffer.clear();
                    }
                }
                // final flush
                if (!buffer.isEmpty()) {
                    appendCellBufferToStagingArea(fnameStaging, buffer, brlen, bclen);
                    buffer.clear();
                }
            } finally {
                IOUtilFunctions.closeSilently(reader);
            }
        }
        // STEP 2: read matrix blocks from staging area and write matrix to HDFS
        String[] fnamesPartitions = new File(fnameStaging).list();
        if (PARALLEL) {
            int len = Math.min(fnamesPartitions.length, _par);
            Thread[] threads = new Thread[len];
            for (int i = 0; i < len; i++) {
                int start = i * (int) Math.ceil(((double) fnamesPartitions.length) / len);
                int end = (i + 1) * (int) Math.ceil(((double) fnamesPartitions.length) / len) - 1;
                end = Math.min(end, fnamesPartitions.length - 1);
                threads[i] = new Thread(new DataPartitionerWorkerTextCell(job, fnameNew, fnameStaging, fnamesPartitions, start, end));
                threads[i].start();
            }
            for (Thread t : threads) t.join();
        } else {
            for (String pdir : fnamesPartitions) writeTextCellFileToHDFS(job, fnameNew, fnameStaging + "/" + pdir);
        }
    } catch (Exception e) {
        // post-mortem error handling and bounds checking
        if (row < 1 || row > rlen || col < 1 || col > clen) {
            throw new DMLRuntimeException("Matrix cell [" + (row) + "," + (col) + "] " + "out of overall matrix range [1:" + rlen + ",1:" + clen + "].");
        } else
            throw new DMLRuntimeException("Unable to partition text cell matrix.", e);
    }
}
Also used : Path(org.apache.hadoop.fs.Path) Text(org.apache.hadoop.io.Text) LinkedList(java.util.LinkedList) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) IOException(java.io.IOException) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) FastStringTokenizer(org.apache.sysml.runtime.util.FastStringTokenizer) TextInputFormat(org.apache.hadoop.mapred.TextInputFormat) LongWritable(org.apache.hadoop.io.LongWritable) JobConf(org.apache.hadoop.mapred.JobConf) InputSplit(org.apache.hadoop.mapred.InputSplit) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Cell(org.apache.sysml.runtime.controlprogram.parfor.util.Cell) SequenceFile(org.apache.hadoop.io.SequenceFile) File(java.io.File)

Example 5 with Cell

use of org.apache.sysml.runtime.controlprogram.parfor.util.Cell in project systemml by apache.

the class ResultMergeLocalFile method createTextCellStagingFile.

private static void createTextCellStagingFile(String fnameStaging, MatrixObject mo, long ID) throws IOException, DMLRuntimeException {
    JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
    Path path = new Path(mo.getFileName());
    FileInputFormat.addInputPath(job, path);
    TextInputFormat informat = new TextInputFormat();
    informat.configure(job);
    InputSplit[] splits = informat.getSplits(job, 1);
    LinkedList<Cell> buffer = new LinkedList<>();
    LongWritable key = new LongWritable();
    Text value = new Text();
    MatrixCharacteristics mc = mo.getMatrixCharacteristics();
    int brlen = mc.getRowsPerBlock();
    int bclen = mc.getColsPerBlock();
    // long row = -1, col = -1; //FIXME needs reconsideration whenever textcell is used actively
    // NOTE MB: Originally, we used long row, col but this led reproducibly to JIT compilation
    // errors during runtime; experienced under WINDOWS, Intel x86-64, IBM JDK 64bit/32bit.
    // It works fine with int row, col but we require long for larger matrices.
    // Since, textcell is never used for result merge (hybrid/hadoop: binaryblock, singlenode:binarycell)
    // we just propose the to exclude it with -Xjit:exclude={package.method*}(count=0,optLevel=0)
    FastStringTokenizer st = new FastStringTokenizer(' ');
    for (InputSplit split : splits) {
        RecordReader<LongWritable, Text> reader = informat.getRecordReader(split, job, Reporter.NULL);
        try {
            while (reader.next(key, value)) {
                // reset tokenizer
                st.reset(value.toString());
                long row = st.nextLong();
                long col = st.nextLong();
                double lvalue = Double.parseDouble(st.nextToken());
                Cell tmp = new Cell(row, col, lvalue);
                buffer.addLast(tmp);
                if (// periodic flush
                buffer.size() > StagingFileUtils.CELL_BUFFER_SIZE) {
                    appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                    buffer.clear();
                }
            }
            // final flush
            if (!buffer.isEmpty()) {
                appendCellBufferToStagingArea(fnameStaging, ID, buffer, brlen, bclen);
                buffer.clear();
            }
        } finally {
            IOUtilFunctions.closeSilently(reader);
        }
    }
}
Also used : Path(org.apache.hadoop.fs.Path) Text(org.apache.hadoop.io.Text) LinkedList(java.util.LinkedList) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) FastStringTokenizer(org.apache.sysml.runtime.util.FastStringTokenizer) TextInputFormat(org.apache.hadoop.mapred.TextInputFormat) LongWritable(org.apache.hadoop.io.LongWritable) JobConf(org.apache.hadoop.mapred.JobConf) InputSplit(org.apache.hadoop.mapred.InputSplit) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Cell(org.apache.sysml.runtime.controlprogram.parfor.util.Cell)

Aggregations

Cell (org.apache.sysml.runtime.controlprogram.parfor.util.Cell)18 MatrixCell (org.apache.sysml.runtime.matrix.data.MatrixCell)18 LinkedList (java.util.LinkedList)14 Path (org.apache.hadoop.fs.Path)14 SequenceFile (org.apache.hadoop.io.SequenceFile)12 File (java.io.File)10 FileSystem (org.apache.hadoop.fs.FileSystem)10 JobConf (org.apache.hadoop.mapred.JobConf)10 MatrixIndexes (org.apache.sysml.runtime.matrix.data.MatrixIndexes)8 IOException (java.io.IOException)6 RecordReader (org.apache.hadoop.mapred.RecordReader)6 DMLRuntimeException (org.apache.sysml.runtime.DMLRuntimeException)6 BufferedWriter (java.io.BufferedWriter)4 OutputStreamWriter (java.io.OutputStreamWriter)4 HashMap (java.util.HashMap)4 LongWritable (org.apache.hadoop.io.LongWritable)4 Text (org.apache.hadoop.io.Text)4 InputSplit (org.apache.hadoop.mapred.InputSplit)4 TextInputFormat (org.apache.hadoop.mapred.TextInputFormat)4 MatrixCharacteristics (org.apache.sysml.runtime.matrix.MatrixCharacteristics)4