use of org.apache.sysml.lops.Lop in project incubator-systemml by apache.
the class Dag method getRecordReaderInstructions.
/**
* Method to get record reader instructions for a MR job.
*
* @param node low-level operator
* @param execNodes list of exec nodes
* @param inputStrings list of input strings
* @param recordReaderInstructions list of record reader instructions
* @param nodeIndexMapping node index mapping
* @param start_index start index
* @param inputLabels list of input labels
* @param inputLops list of input lops
* @param MRJobLineNumbers MR job line numbers
* @return -1 if problem
*/
private static int getRecordReaderInstructions(Lop node, ArrayList<Lop> execNodes, ArrayList<String> inputStrings, ArrayList<String> recordReaderInstructions, HashMap<Lop, Integer> nodeIndexMapping, int[] start_index, ArrayList<String> inputLabels, ArrayList<Lop> inputLops, ArrayList<Integer> MRJobLineNumbers) {
// if input source, return index
if (nodeIndexMapping.containsKey(node))
return nodeIndexMapping.get(node);
// not input source and not in exec nodes, then return.
if (!execNodes.contains(node))
return -1;
ArrayList<Integer> inputIndices = new ArrayList<>();
int max_input_index = -1;
// get mapper instructions
for (int i = 0; i < node.getInputs().size(); i++) {
// recurse
Lop childNode = node.getInputs().get(i);
int ret_val = getRecordReaderInstructions(childNode, execNodes, inputStrings, recordReaderInstructions, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers);
inputIndices.add(ret_val);
if (ret_val > max_input_index) {
max_input_index = ret_val;
// child_for_max_input_index = childNode;
}
}
// instructions
if ((node.getExecLocation() == ExecLocation.RecordReader)) {
int output_index = max_input_index;
// cannot reuse index if this is true
// need to add better indexing schemes
output_index = start_index[0];
start_index[0]++;
nodeIndexMapping.put(node, output_index);
// only Ranagepick lop can contribute to labels
if (node.getType() == Type.PickValues) {
PickByCount pbc = (PickByCount) node;
if (pbc.getOperationType() == PickByCount.OperationTypes.RANGEPICK) {
// always the second input is a scalar
int scalarIndex = 1;
// if data lop not a literal -- add label
if (node.getInputs().get(scalarIndex).getExecLocation() == ExecLocation.Data && !((Data) (node.getInputs().get(scalarIndex))).isLiteral()) {
inputLabels.add(node.getInputs().get(scalarIndex).getOutputParameters().getLabel());
inputLops.add(node.getInputs().get(scalarIndex));
}
// if not data lop, then this is an intermediate variable.
if (node.getInputs().get(scalarIndex).getExecLocation() != ExecLocation.Data) {
inputLabels.add(node.getInputs().get(scalarIndex).getOutputParameters().getLabel());
inputLops.add(node.getInputs().get(scalarIndex));
}
}
}
// get recordreader instruction.
if (node.getInputs().size() == 2) {
recordReaderInstructions.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), output_index));
if (DMLScript.ENABLE_DEBUG_MODE) {
MRJobLineNumbers.add(node._beginLine);
}
} else
throw new LopsException("Unexpected number of inputs while generating a RecordReader Instruction");
return output_index;
}
return -1;
}
use of org.apache.sysml.lops.Lop in project incubator-systemml by apache.
the class Dag method generateInstructionsForInputVariables.
/**
* Method to generate createvar instructions, which creates a new entry
* in the symbol table. One instruction is generated for every LOP that is
* 1) type Data and
* 2) persistent and
* 3) matrix and
* 4) read
*
* Transient reads needn't be considered here since the previous program
* block would already create appropriate entries in the symbol table.
*
* @param nodes_v list of nodes
* @param inst list of instructions
*/
private static void generateInstructionsForInputVariables(ArrayList<Lop> nodes_v, ArrayList<Instruction> inst) {
for (Lop n : nodes_v) {
if (n.getExecLocation() == ExecLocation.Data && !((Data) n).isTransient() && ((Data) n).getOperationType() == OperationTypes.READ && (n.getDataType() == DataType.MATRIX || n.getDataType() == DataType.FRAME)) {
if (!((Data) n).isLiteral()) {
try {
String inst_string = n.getInstructions();
CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(inst_string);
currInstr.setLocation(n);
inst.add(currInstr);
} catch (DMLRuntimeException e) {
throw new LopsException(n.printErrorLocation() + "error generating instructions from input variables in Dag -- \n", e);
}
}
}
}
}
use of org.apache.sysml.lops.Lop in project incubator-systemml by apache.
the class Dag method deleteUpdatedTransientReadVariables.
private static void deleteUpdatedTransientReadVariables(StatementBlock sb, ArrayList<Lop> nodeV, ArrayList<Instruction> inst) {
if (sb == null)
return;
if (LOG.isTraceEnabled())
LOG.trace("In delete updated variables");
// CANDIDATE list of variables which could have been updated in this statement block
HashMap<String, Lop> labelNodeMapping = new HashMap<>();
// ACTUAL list of variables whose value is updated, AND the old value of the variable
// is no longer accessible/used.
HashSet<String> updatedLabels = new HashSet<>();
HashMap<String, Lop> updatedLabelsLineNum = new HashMap<>();
// first capture all transient read variables
for (Lop node : nodeV) {
if (node.getExecLocation() == ExecLocation.Data && ((Data) node).isTransient() && ((Data) node).getOperationType() == OperationTypes.READ && ((Data) node).getDataType() == DataType.MATRIX) {
// "node" is considered as updated ONLY IF the old value is not used any more
// So, make sure that this READ node does not feed into any (transient/persistent) WRITE
boolean hasWriteParent = false;
for (Lop p : node.getOutputs()) {
if (p.getExecLocation() == ExecLocation.Data) {
// if the "p" is of type Data, then it has to be a WRITE
hasWriteParent = true;
break;
}
}
if (!hasWriteParent) {
// node has no parent of type WRITE, so this is a CANDIDATE variable
// add it to labelNodeMapping so that it is considered in further processing
labelNodeMapping.put(node.getOutputParameters().getLabel(), node);
}
}
}
// capture updated transient write variables
for (Lop node : nodeV) {
if (node.getExecLocation() == ExecLocation.Data && ((Data) node).isTransient() && ((Data) node).getOperationType() == OperationTypes.WRITE && ((Data) node).getDataType() == DataType.MATRIX && // check to make sure corresponding (i.e., with the same label/name) transient read is present
labelNodeMapping.containsKey(node.getOutputParameters().getLabel()) && // check to avoid cases where transient read feeds into a transient write
!labelNodeMapping.containsValue(node.getInputs().get(0))) {
updatedLabels.add(node.getOutputParameters().getLabel());
updatedLabelsLineNum.put(node.getOutputParameters().getLabel(), node);
}
}
// generate RM instructions
Instruction rm_inst = null;
for (String label : updatedLabels) {
rm_inst = VariableCPInstruction.prepareRemoveInstruction(label);
rm_inst.setLocation(updatedLabelsLineNum.get(label));
if (LOG.isTraceEnabled())
LOG.trace(rm_inst.toString());
inst.add(rm_inst);
}
}
use of org.apache.sysml.lops.Lop in project incubator-systemml by apache.
the class Dag method doGreedyGrouping.
/**
* Method to group a vector of sorted lops.
*
* @param sb statement block
* @param node_v list of low-level operators
* @return list of instructions
*/
private ArrayList<Instruction> doGreedyGrouping(StatementBlock sb, ArrayList<Lop> node_v) {
if (LOG.isTraceEnabled())
LOG.trace("Grouping DAG ============");
// nodes to be executed in current iteration
ArrayList<Lop> execNodes = new ArrayList<>();
// nodes that have already been processed
ArrayList<Lop> finishedNodes = new ArrayList<>();
// nodes that are queued for the following iteration
ArrayList<Lop> queuedNodes = new ArrayList<>();
ArrayList<ArrayList<Lop>> jobNodes = createNodeVectors(JobType.getNumJobTypes());
// list of instructions
ArrayList<Instruction> inst = new ArrayList<>();
// ArrayList<Instruction> preWriteDeleteInst = new ArrayList<Instruction>();
ArrayList<Instruction> writeInst = new ArrayList<>();
ArrayList<Instruction> deleteInst = new ArrayList<>();
ArrayList<Instruction> endOfBlockInst = new ArrayList<>();
// remove files for transient reads that are updated.
deleteUpdatedTransientReadVariables(sb, node_v, writeInst);
generateRemoveInstructions(sb, endOfBlockInst);
generateInstructionsForInputVariables(node_v, inst);
boolean done = false;
String indent = " ";
while (!done) {
if (LOG.isTraceEnabled())
LOG.trace("Grouping nodes in DAG");
execNodes.clear();
queuedNodes.clear();
clearNodeVectors(jobNodes);
gmrMapperFootprint = 0;
for (Lop node : node_v) {
// finished nodes don't need to be processed
if (finishedNodes.contains(node))
continue;
if (LOG.isTraceEnabled())
LOG.trace("Processing node (" + node.getID() + ") " + node.toString() + " exec nodes size is " + execNodes.size());
// its children nodes in execNodes
if (node.definesMRJob() && !compatibleWithChildrenInExecNodes(execNodes, node)) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing node " + node.toString() + " (code 1)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
continue;
}
// iteration
if (hasChildNode(node, queuedNodes)) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing node " + node.toString() + " (code 2)");
queuedNodes.add(node);
// if node has more than two inputs,
// remove children that will be needed in a future
// iterations
// may also have to remove parent nodes of these children
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
continue;
}
// if inputs come from different jobs, then queue
if (node.getInputs().size() >= 2) {
int jobid = Integer.MIN_VALUE;
boolean queueit = false;
for (int idx = 0; idx < node.getInputs().size(); idx++) {
int input_jobid = jobType(node.getInputs().get(idx), jobNodes);
if (input_jobid != -1) {
if (jobid == Integer.MIN_VALUE)
jobid = input_jobid;
else if (jobid != input_jobid) {
queueit = true;
break;
}
}
}
if (queueit) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing node " + node.toString() + " (code 3)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
continue;
}
}
// See if this lop can be eliminated
// This check is for "aligner" lops (e.g., group)
boolean eliminate = false;
eliminate = canEliminateLop(node, execNodes);
if (eliminate) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, eliminate);
continue;
}
// children that defines a MR Job are present in execNodes
if (node.definesMRJob()) {
if (hasMRJobChildNode(node, execNodes)) {
// this is because "group" can be pushed into the "Rand" job.
if (!(node.getType() == Lop.Type.Grouping && checkDataGenAsChildNode(node, execNodes))) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing node " + node.toString() + " (code 4)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
continue;
}
}
}
// not, queue "node"
if (node.getInputs().size() > 1 && hasChildNode(node, execNodes, ExecLocation.RecordReader)) {
// get the actual RecordReader lop
Lop rr_node = getChildNode(node, execNodes, ExecLocation.RecordReader);
// all inputs of "node" must be ancestors of rr_node
boolean queue_it = false;
for (Lop n : node.getInputs()) {
// each input should be ancestor of RecordReader lop
if (!n.equals(rr_node) && !isChild(rr_node, n, IDMap)) {
// i.e., "node" must be queued
queue_it = true;
break;
}
}
if (queue_it) {
// queue node
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 5)");
queuedNodes.add(node);
// TODO: does this have to be modified to handle
// recordreader lops?
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
continue;
} else {
// nothing here.. subsequent checks have to be performed
// on "node"
}
}
// only write nodes are kept in execnodes
if (node.getExecLocation() == ExecLocation.Data) {
Data dnode = (Data) node;
boolean dnode_queued = false;
if (dnode.getOperationType() == OperationTypes.READ) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding Data -" + node.toString());
// TODO: avoid readScalar instruction, and read it on-demand just like the way Matrices are read in control program
if (node.getDataType() == DataType.SCALAR && // TODO: LEO check the following condition is still needed
node.getOutputParameters().getFile_name() != null) {
// this lop corresponds to reading a scalar from HDFS file
// add it to execNodes so that "readScalar" instruction gets generated
execNodes.add(node);
// note: no need to add it to any job vector
}
} else if (dnode.getOperationType() == OperationTypes.WRITE) {
// Skip the transient write <code>node</code> if the input is a
// transient read with the same variable name. i.e., a dummy copy.
// Hence, <code>node</code> can be avoided.
// TODO: this case should ideally be handled in the language layer
// prior to the construction of Hops Dag
Lop input = dnode.getInputs().get(0);
if (dnode.isTransient() && input.getExecLocation() == ExecLocation.Data && ((Data) input).isTransient() && dnode.getOutputParameters().getLabel().equals(input.getOutputParameters().getLabel())) {
// do nothing, <code>node</code> must not processed any further.
} else if (execNodes.contains(input) && !isCompatible(node, input) && sendWriteLopToMR(node)) {
// input is in execNodes but it is not compatible with write lop. So, queue the write lop.
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString());
queuedNodes.add(node);
dnode_queued = true;
} else {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding Data -" + node.toString());
execNodes.add(node);
if (sendWriteLopToMR(node)) {
addNodeByJobType(node, jobNodes, execNodes, false);
}
}
}
if (!dnode_queued)
finishedNodes.add(node);
continue;
}
// map or reduce node, can always be piggybacked with parent
if (node.getExecLocation() == ExecLocation.MapOrReduce) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, false);
continue;
}
// RecordReader node, add, if no parent needs reduce, else queue
if (node.getExecLocation() == ExecLocation.RecordReader) {
// execNodes .. it has to be the first one in the job!
if (!hasChildNode(node, execNodes, ExecLocation.Map) && !hasChildNode(node, execNodes, ExecLocation.MapAndReduce)) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, false);
} else {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 6)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
}
continue;
}
// map node, add, if no parent needs reduce, else queue
if (node.getExecLocation() == ExecLocation.Map) {
boolean queueThisNode = false;
int subcode = -1;
if (node.usesDistributedCache()) {
// if an input to <code>node</code> comes from distributed cache
// then that input must get executed in one of the previous jobs.
int[] dcInputIndexes = node.distributedCacheInputIndex();
for (int dcInputIndex : dcInputIndexes) {
Lop dcInput = node.getInputs().get(dcInputIndex - 1);
if ((dcInput.getType() != Lop.Type.Data && dcInput.getExecType() == ExecType.MR) && execNodes.contains(dcInput)) {
queueThisNode = true;
subcode = 1;
}
}
// Limit the number of distributed cache inputs based on the available memory in mappers
double memsize = computeFootprintInMapper(node);
if (gmrMapperFootprint > 0 && !checkMemoryLimits(node, gmrMapperFootprint + memsize)) {
queueThisNode = true;
subcode = 2;
}
if (!queueThisNode)
gmrMapperFootprint += memsize;
}
if (!queueThisNode && !hasChildNode(node, execNodes, ExecLocation.MapAndReduce) && !hasMRJobChildNode(node, execNodes)) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, false);
} else {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 7 - " + "subcode " + subcode + ")");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
}
continue;
}
// reduce node, make sure no parent needs reduce, else queue
if (node.getExecLocation() == ExecLocation.MapAndReduce) {
// not define a job
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, eliminate);
continue;
}
// aligned reduce, make sure a parent that is reduce exists
if (node.getExecLocation() == ExecLocation.Reduce) {
if (compatibleWithChildrenInExecNodes(execNodes, node) && (hasChildNode(node, execNodes, ExecLocation.MapAndReduce) || hasChildNode(node, execNodes, ExecLocation.Map))) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding -" + node.toString());
execNodes.add(node);
finishedNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, false);
} else {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 8)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
}
continue;
}
// that will be executed in a MR job.
if (node.getExecLocation() == ExecLocation.ControlProgram) {
for (Lop lop : node.getInputs()) {
if (execNodes.contains(lop) && !(lop.getExecLocation() == ExecLocation.Data) && !(lop.getExecLocation() == ExecLocation.ControlProgram)) {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 9)");
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
break;
}
}
if (queuedNodes.contains(node))
continue;
if (LOG.isTraceEnabled())
LOG.trace(indent + "Adding - scalar" + node.toString());
execNodes.add(node);
addNodeByJobType(node, jobNodes, execNodes, false);
finishedNodes.add(node);
continue;
}
}
// no work to do
if (execNodes.isEmpty()) {
if (!queuedNodes.isEmpty())
throw new LopsException("Queued nodes should not be 0 at this point \n");
if (LOG.isTraceEnabled())
LOG.trace("All done! queuedNodes = " + queuedNodes.size());
done = true;
} else {
if (LOG.isTraceEnabled())
LOG.trace("Generating jobs for group -- Node count=" + execNodes.size());
// first process scalar instructions
generateControlProgramJobs(execNodes, inst, writeInst, deleteInst);
// copy unassigned lops in execnodes to gmrnodes
for (int i = 0; i < execNodes.size(); i++) {
Lop node = execNodes.get(i);
if (jobType(node, jobNodes) == -1) {
if (isCompatible(node, JobType.GMR)) {
if (node.hasNonBlockedInputs()) {
jobNodes.get(JobType.GMRCELL.getId()).add(node);
addChildren(node, jobNodes.get(JobType.GMRCELL.getId()), execNodes);
} else {
jobNodes.get(JobType.GMR.getId()).add(node);
addChildren(node, jobNodes.get(JobType.GMR.getId()), execNodes);
}
} else {
if (LOG.isTraceEnabled())
LOG.trace(indent + "Queueing -" + node.toString() + " (code 10)");
execNodes.remove(i);
finishedNodes.remove(node);
queuedNodes.add(node);
removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes);
}
}
}
// next generate MR instructions
if (!execNodes.isEmpty())
generateMRJobs(execNodes, inst, writeInst, deleteInst, jobNodes);
handleSingleOutputJobs(execNodes, jobNodes, finishedNodes);
}
}
// add write and delete inst at the very end.
// inst.addAll(preWriteDeleteInst);
inst.addAll(writeInst);
inst.addAll(deleteInst);
inst.addAll(endOfBlockInst);
return inst;
}
use of org.apache.sysml.lops.Lop in project incubator-systemml by apache.
the class Dag method createIDMapping.
private ArrayList<Lop> createIDMapping(Lop[] nodearray) {
// Copy sorted nodes into "v" and construct a mapping between Lop IDs and sequence of numbers
ArrayList<Lop> ret = new ArrayList<>();
IDMap.clear();
for (int i = 0; i < nodearray.length; i++) {
ret.add(nodearray[i]);
IDMap.put(nodearray[i].getID(), i);
}
/*
* Compute of All-pair reachability graph (Transitive Closure) of the DAG.
* - Perform a depth-first search (DFS) from every node $u$ in the DAG
* - and construct the list of reachable nodes from the node $u$
* - store the constructed reachability information in $u$.reachable[] boolean array
*/
for (int i = 0; i < nodearray.length; i++) {
dagDFS(nodearray[i], nodearray[i].create_reachable(nodearray.length));
}
// print the nodes in sorted order
if (LOG.isTraceEnabled()) {
for (Lop vnode : ret) {
StringBuilder sb = new StringBuilder();
sb.append(vnode.getID());
sb.append("(");
sb.append(vnode.getLevel());
sb.append(") ");
sb.append(vnode.getType());
sb.append("(");
for (Lop vin : vnode.getInputs()) {
sb.append(vin.getID());
sb.append(",");
}
sb.append("), ");
LOG.trace(sb.toString());
}
LOG.trace("topological sort -- done");
}
return ret;
}
Aggregations