Search in sources :

Example 61 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class ResultMergeLocalFile method createBinaryBlockResultFile.

@SuppressWarnings("deprecation")
private void createBinaryBlockResultFile(String fnameStaging, String fnameStagingCompare, String fnameNew, MetaDataFormat metadata, boolean withCompare) throws IOException, DMLRuntimeException {
    JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
    Path path = new Path(fnameNew);
    FileSystem fs = IOUtilFunctions.getFileSystem(path, job);
    MatrixCharacteristics mc = metadata.getMatrixCharacteristics();
    long rlen = mc.getRows();
    long clen = mc.getCols();
    int brlen = mc.getRowsPerBlock();
    int bclen = mc.getColsPerBlock();
    // beware ca 50ms
    SequenceFile.Writer writer = new SequenceFile.Writer(fs, job, path, MatrixIndexes.class, MatrixBlock.class);
    try {
        MatrixIndexes indexes = new MatrixIndexes();
        for (long brow = 1; brow <= (long) Math.ceil(rlen / (double) brlen); brow++) for (long bcol = 1; bcol <= (long) Math.ceil(clen / (double) bclen); bcol++) {
            File dir = new File(fnameStaging + "/" + brow + "_" + bcol);
            File dir2 = new File(fnameStagingCompare + "/" + brow + "_" + bcol);
            MatrixBlock mb = null;
            if (dir.exists()) {
                if (// WITH COMPARE BLOCK
                withCompare && dir2.exists()) {
                    // copy only values that are different from the original
                    String[] lnames2 = dir2.list();
                    if (// there should be exactly 1 compare block
                    lnames2.length != 1)
                        throw new DMLRuntimeException("Unable to merge results because multiple compare blocks found.");
                    mb = LocalFileUtils.readMatrixBlockFromLocal(dir2 + "/" + lnames2[0]);
                    boolean appendOnly = mb.isInSparseFormat();
                    DenseBlock compare = DataConverter.convertToDenseBlock(mb, false);
                    for (String lname : dir.list()) {
                        MatrixBlock tmp = LocalFileUtils.readMatrixBlockFromLocal(dir + "/" + lname);
                        mergeWithComp(mb, tmp, compare);
                    }
                    // sort sparse due to append-only
                    if (appendOnly && !_isAccum)
                        mb.sortSparseRows();
                    // change sparsity if required after
                    mb.examSparsity();
                } else // WITHOUT COMPARE BLOCK
                {
                    // copy all non-zeros from all workers
                    boolean appendOnly = false;
                    for (String lname : dir.list()) {
                        if (mb == null) {
                            mb = LocalFileUtils.readMatrixBlockFromLocal(dir + "/" + lname);
                            appendOnly = mb.isInSparseFormat();
                        } else {
                            MatrixBlock tmp = LocalFileUtils.readMatrixBlockFromLocal(dir + "/" + lname);
                            mergeWithoutComp(mb, tmp, appendOnly);
                        }
                    }
                    // sort sparse due to append-only
                    if (appendOnly && !_isAccum)
                        mb.sortSparseRows();
                    // change sparsity if required after
                    mb.examSparsity();
                }
            } else {
                // NOTE: whenever runtime does not need all blocks anymore, this can be removed
                int maxRow = (int) (((brow - 1) * brlen + brlen < rlen) ? brlen : rlen - (brow - 1) * brlen);
                int maxCol = (int) (((bcol - 1) * bclen + bclen < clen) ? bclen : clen - (bcol - 1) * bclen);
                mb = new MatrixBlock(maxRow, maxCol, true);
            }
            // mb.examSparsity(); //done on write anyway and mb not reused
            indexes.setIndexes(brow, bcol);
            writer.append(indexes, mb);
        }
    } finally {
        IOUtilFunctions.closeSilently(writer);
    }
}
Also used : Path(org.apache.hadoop.fs.Path) MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) DenseBlock(org.apache.sysml.runtime.matrix.data.DenseBlock) SequenceFile(org.apache.hadoop.io.SequenceFile) FileSystem(org.apache.hadoop.fs.FileSystem) JobConf(org.apache.hadoop.mapred.JobConf) SequenceFile(org.apache.hadoop.io.SequenceFile) File(java.io.File) OutputStreamWriter(java.io.OutputStreamWriter) BufferedWriter(java.io.BufferedWriter)

Example 62 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class ResultMergeLocalFile method createBinaryCellResultFile.

@SuppressWarnings("deprecation")
private void createBinaryCellResultFile(String fnameStaging, String fnameStagingCompare, String fnameNew, MetaDataFormat metadata, boolean withCompare) throws IOException, DMLRuntimeException {
    JobConf job = new JobConf(ConfigurationManager.getCachedJobConf());
    Path path = new Path(fnameNew);
    FileSystem fs = IOUtilFunctions.getFileSystem(path, job);
    MatrixCharacteristics mc = metadata.getMatrixCharacteristics();
    long rlen = mc.getRows();
    long clen = mc.getCols();
    int brlen = mc.getRowsPerBlock();
    int bclen = mc.getColsPerBlock();
    MatrixIndexes indexes = new MatrixIndexes(1, 1);
    MatrixCell cell = new MatrixCell(0);
    // beware ca 50ms
    SequenceFile.Writer out = new SequenceFile.Writer(fs, job, path, MatrixIndexes.class, MatrixCell.class);
    try {
        boolean written = false;
        for (long brow = 1; brow <= (long) Math.ceil(rlen / (double) brlen); brow++) for (long bcol = 1; bcol <= (long) Math.ceil(clen / (double) bclen); bcol++) {
            File dir = new File(fnameStaging + "/" + brow + "_" + bcol);
            File dir2 = new File(fnameStagingCompare + "/" + brow + "_" + bcol);
            MatrixBlock mb = null;
            long row_offset = (brow - 1) * brlen + 1;
            long col_offset = (bcol - 1) * bclen + 1;
            if (dir.exists()) {
                if (// WITH COMPARE BLOCK
                withCompare && dir2.exists()) {
                    // copy only values that are different from the original
                    String[] lnames2 = dir2.list();
                    if (// there should be exactly 1 compare block
                    lnames2.length != 1)
                        throw new DMLRuntimeException("Unable to merge results because multiple compare blocks found.");
                    mb = StagingFileUtils.readCellList2BlockFromLocal(dir2 + "/" + lnames2[0], brlen, bclen);
                    boolean appendOnly = mb.isInSparseFormat();
                    DenseBlock compare = DataConverter.convertToDenseBlock(mb, false);
                    for (String lname : dir.list()) {
                        MatrixBlock tmp = StagingFileUtils.readCellList2BlockFromLocal(dir + "/" + lname, brlen, bclen);
                        mergeWithComp(mb, tmp, compare);
                    }
                    // sort sparse due to append-only
                    if (appendOnly && !_isAccum)
                        mb.sortSparseRows();
                    // change sparsity if required after
                    mb.examSparsity();
                } else // WITHOUT COMPARE BLOCK
                {
                    // copy all non-zeros from all workers
                    boolean appendOnly = false;
                    for (String lname : dir.list()) {
                        if (mb == null) {
                            mb = StagingFileUtils.readCellList2BlockFromLocal(dir + "/" + lname, brlen, bclen);
                            appendOnly = mb.isInSparseFormat();
                        } else {
                            MatrixBlock tmp = StagingFileUtils.readCellList2BlockFromLocal(dir + "/" + lname, brlen, bclen);
                            mergeWithoutComp(mb, tmp, appendOnly);
                        }
                    }
                    // sort sparse due to append-only
                    if (appendOnly && !_isAccum)
                        mb.sortSparseRows();
                    // change sparsity if required after
                    mb.examSparsity();
                }
            }
            // write the block to binary cell
            if (mb != null) {
                if (mb.isInSparseFormat()) {
                    Iterator<IJV> iter = mb.getSparseBlockIterator();
                    while (iter.hasNext()) {
                        IJV lcell = iter.next();
                        indexes.setIndexes(row_offset + lcell.getI(), col_offset + lcell.getJ());
                        cell.setValue(lcell.getV());
                        out.append(indexes, cell);
                        written = true;
                    }
                } else {
                    for (int i = 0; i < brlen; i++) for (int j = 0; j < bclen; j++) {
                        double lvalue = mb.getValueDenseUnsafe(i, j);
                        if (// for nnz
                        lvalue != 0) {
                            indexes.setIndexes(row_offset + i, col_offset + j);
                            cell.setValue(lvalue);
                            out.append(indexes, cell);
                            written = true;
                        }
                    }
                }
            }
        }
        if (!written)
            out.append(indexes, cell);
    } finally {
        IOUtilFunctions.closeSilently(out);
    }
}
Also used : Path(org.apache.hadoop.fs.Path) MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) DenseBlock(org.apache.sysml.runtime.matrix.data.DenseBlock) SequenceFile(org.apache.hadoop.io.SequenceFile) IJV(org.apache.sysml.runtime.matrix.data.IJV) FileSystem(org.apache.hadoop.fs.FileSystem) MatrixCell(org.apache.sysml.runtime.matrix.data.MatrixCell) Iterator(java.util.Iterator) JobConf(org.apache.hadoop.mapred.JobConf) SequenceFile(org.apache.hadoop.io.SequenceFile) File(java.io.File) OutputStreamWriter(java.io.OutputStreamWriter) BufferedWriter(java.io.BufferedWriter)

Example 63 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class ResultMergeLocalMemory method executeParallelMerge.

@Override
public MatrixObject executeParallelMerge(int par) {
    // always create new matrix object (required for nested parallelism)
    MatrixObject moNew = null;
    if (LOG.isTraceEnabled())
        LOG.trace("ResultMerge (local, in-memory): Execute parallel (par=" + par + ") " + "merge for output " + _output.hashCode() + " (fname=" + _output.getFileName() + ")");
    try {
        // get matrix blocks through caching
        MatrixBlock outMB = _output.acquireRead();
        ArrayList<MatrixObject> inMO = new ArrayList<>();
        for (MatrixObject in : _inputs) {
            // check for empty inputs (no iterations executed)
            if (in != null && in != _output)
                inMO.add(in);
        }
        if (// if there exist something to merge
        !inMO.isEmpty()) {
            // get old output matrix from cache for compare
            // NOTE: always in dense representation in order to allow for parallel unsynchronized access
            long rows = outMB.getNumRows();
            long cols = outMB.getNumColumns();
            MatrixBlock outMBNew = new MatrixBlock((int) rows, (int) cols, false);
            outMBNew.allocateDenseBlockUnsafe((int) rows, (int) cols);
            // create compare matrix if required (existing data in result)
            _compare = getCompareMatrix(outMB);
            if (_compare != null)
                outMBNew.copy(outMB);
            // parallel merge of all inputs
            // number of inputs can be lower than par
            int numThreads = Math.min(par, inMO.size());
            // ensure robustness for remote exec
            numThreads = Math.min(numThreads, InfrastructureAnalyzer.getLocalParallelism());
            Thread[] threads = new Thread[numThreads];
            for (// multiple waves if necessary
            int k = 0; // multiple waves if necessary
            k < inMO.size(); // multiple waves if necessary
            k += numThreads) {
                // create and start threads
                for (int i = 0; i < threads.length; i++) {
                    ResultMergeWorker rmw = new ResultMergeWorker(inMO.get(k + i), outMBNew);
                    threads[i] = new Thread(rmw);
                    threads[i].setPriority(Thread.MAX_PRIORITY);
                    // start execution
                    threads[i].start();
                }
                // wait for all workers to finish
                for (int i = 0; i < threads.length; i++) {
                    threads[i].join();
                }
            }
            // create new output matrix
            // (e.g., to prevent potential export<->read file access conflict in specific cases of
            // local-remote nested parfor))
            moNew = createNewMatrixObject(outMBNew);
        } else {
            // return old matrix, to prevent copy
            moNew = _output;
        }
        // release old output, and all inputs
        _output.release();
    } catch (Exception ex) {
        throw new DMLRuntimeException(ex);
    }
    return moNew;
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixObject(org.apache.sysml.runtime.controlprogram.caching.MatrixObject) ArrayList(java.util.ArrayList) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException)

Example 64 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class ResultMergeRemoteSpark method setRDDHandleForMerge.

@SuppressWarnings("unchecked")
private static void setRDDHandleForMerge(MatrixObject mo, SparkExecutionContext sec) {
    InputInfo iinfo = InputInfo.BinaryBlockInputInfo;
    JavaSparkContext sc = sec.getSparkContext();
    JavaPairRDD<MatrixIndexes, MatrixBlock> rdd = (JavaPairRDD<MatrixIndexes, MatrixBlock>) sc.hadoopFile(mo.getFileName(), iinfo.inputFormatClass, iinfo.inputKeyClass, iinfo.inputValueClass);
    RDDObject rddhandle = new RDDObject(rdd);
    rddhandle.setHDFSFile(true);
    mo.setRDDHandle(rddhandle);
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) InputInfo(org.apache.sysml.runtime.matrix.data.InputInfo) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) JavaPairRDD(org.apache.spark.api.java.JavaPairRDD) RDDObject(org.apache.sysml.runtime.instructions.spark.data.RDDObject) JavaSparkContext(org.apache.spark.api.java.JavaSparkContext)

Example 65 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class MultiReturnParameterizedBuiltinCPInstruction method processInstruction.

@Override
public void processInstruction(ExecutionContext ec) {
    // obtain and pin input frame
    FrameBlock fin = ec.getFrameInput(input1.getName());
    String spec = ec.getScalarInput(input2.getName(), input2.getValueType(), input2.isLiteral()).getStringValue();
    String[] colnames = fin.getColumnNames();
    // execute block transform encode
    Encoder encoder = EncoderFactory.createEncoder(spec, colnames, fin.getNumColumns(), null);
    // build and apply
    MatrixBlock data = encoder.encode(fin, new MatrixBlock(fin.getNumRows(), fin.getNumColumns(), false));
    FrameBlock meta = encoder.getMetaData(new FrameBlock(fin.getNumColumns(), ValueType.STRING));
    meta.setColumnNames(colnames);
    // release input and outputs
    ec.releaseFrameInput(input1.getName());
    ec.setMatrixOutput(getOutput(0).getName(), data, getExtendedOpcode());
    ec.setFrameOutput(getOutput(1).getName(), meta);
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) FrameBlock(org.apache.sysml.runtime.matrix.data.FrameBlock) Encoder(org.apache.sysml.runtime.transform.encode.Encoder)

Aggregations

MatrixBlock (org.apache.sysml.runtime.matrix.data.MatrixBlock)459 MatrixIndexes (org.apache.sysml.runtime.matrix.data.MatrixIndexes)142 DMLRuntimeException (org.apache.sysml.runtime.DMLRuntimeException)111 MatrixCharacteristics (org.apache.sysml.runtime.matrix.MatrixCharacteristics)102 CompressedMatrixBlock (org.apache.sysml.runtime.compress.CompressedMatrixBlock)48 SparkExecutionContext (org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext)48 IOException (java.io.IOException)44 MatrixObject (org.apache.sysml.runtime.controlprogram.caching.MatrixObject)41 ArrayList (java.util.ArrayList)40 Path (org.apache.hadoop.fs.Path)29 FrameBlock (org.apache.sysml.runtime.matrix.data.FrameBlock)24 FileSystem (org.apache.hadoop.fs.FileSystem)23 JavaPairRDD (org.apache.spark.api.java.JavaPairRDD)23 JobConf (org.apache.hadoop.mapred.JobConf)21 Tuple2 (scala.Tuple2)19 SequenceFile (org.apache.hadoop.io.SequenceFile)17 Row (org.apache.spark.sql.Row)14 SparseBlock (org.apache.sysml.runtime.matrix.data.SparseBlock)14 TestConfiguration (org.apache.sysml.test.integration.TestConfiguration)14 IndexedMatrixValue (org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue)13