use of org.apache.flink.optimizer.plan.WorksetIterationPlanNode in project flink by apache.
the class ConnectedComponentsCoGroupTest method testWorksetConnectedComponents.
@Test
public void testWorksetConnectedComponents() {
Plan plan = getConnectedComponentsCoGroupPlan();
plan.setExecutionConfig(new ExecutionConfig());
OptimizedPlan optPlan = compileNoStats(plan);
OptimizerPlanNodeResolver or = getOptimizerPlanNodeResolver(optPlan);
if (PRINT_PLAN) {
PlanJSONDumpGenerator dumper = new PlanJSONDumpGenerator();
String json = dumper.getOptimizerPlanAsJSON(optPlan);
System.out.println(json);
}
SourcePlanNode vertexSource = or.getNode(VERTEX_SOURCE);
SourcePlanNode edgesSource = or.getNode(EDGES_SOURCE);
SinkPlanNode sink = or.getNode(SINK);
WorksetIterationPlanNode iter = or.getNode(ITERATION_NAME);
DualInputPlanNode neighborsJoin = or.getNode(JOIN_NEIGHBORS_MATCH);
DualInputPlanNode cogroup = or.getNode(MIN_ID_AND_UPDATE);
// --------------------------------------------------------------------
// Plan validation:
//
// We expect the plan to go with a sort-merge join, because the CoGroup
// sorts and the join in the successive iteration can re-exploit the sorting.
// --------------------------------------------------------------------
// test all drivers
Assert.assertEquals(DriverStrategy.NONE, sink.getDriverStrategy());
Assert.assertEquals(DriverStrategy.NONE, vertexSource.getDriverStrategy());
Assert.assertEquals(DriverStrategy.NONE, edgesSource.getDriverStrategy());
Assert.assertEquals(DriverStrategy.INNER_MERGE, neighborsJoin.getDriverStrategy());
Assert.assertEquals(set0, neighborsJoin.getKeysForInput1());
Assert.assertEquals(set0, neighborsJoin.getKeysForInput2());
Assert.assertEquals(DriverStrategy.CO_GROUP, cogroup.getDriverStrategy());
Assert.assertEquals(set0, cogroup.getKeysForInput1());
Assert.assertEquals(set0, cogroup.getKeysForInput2());
// test all the shipping strategies
Assert.assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
Assert.assertEquals(ShipStrategyType.PARTITION_HASH, iter.getInitialSolutionSetInput().getShipStrategy());
Assert.assertEquals(set0, iter.getInitialSolutionSetInput().getShipStrategyKeys());
Assert.assertEquals(ShipStrategyType.PARTITION_HASH, iter.getInitialWorksetInput().getShipStrategy());
Assert.assertEquals(set0, iter.getInitialWorksetInput().getShipStrategyKeys());
// workset
Assert.assertEquals(ShipStrategyType.FORWARD, neighborsJoin.getInput1().getShipStrategy());
// edges
Assert.assertEquals(ShipStrategyType.PARTITION_HASH, neighborsJoin.getInput2().getShipStrategy());
Assert.assertEquals(set0, neighborsJoin.getInput2().getShipStrategyKeys());
Assert.assertTrue(neighborsJoin.getInput2().getTempMode().isCached());
// min id
Assert.assertEquals(ShipStrategyType.PARTITION_HASH, cogroup.getInput1().getShipStrategy());
// solution set
Assert.assertEquals(ShipStrategyType.FORWARD, cogroup.getInput2().getShipStrategy());
// test all the local strategies
Assert.assertEquals(LocalStrategy.NONE, sink.getInput().getLocalStrategy());
Assert.assertEquals(LocalStrategy.NONE, iter.getInitialSolutionSetInput().getLocalStrategy());
// the sort for the neighbor join in the first iteration is pushed out of the loop
Assert.assertEquals(LocalStrategy.SORT, iter.getInitialWorksetInput().getLocalStrategy());
// workset
Assert.assertEquals(LocalStrategy.NONE, neighborsJoin.getInput1().getLocalStrategy());
// edges
Assert.assertEquals(LocalStrategy.SORT, neighborsJoin.getInput2().getLocalStrategy());
Assert.assertEquals(LocalStrategy.SORT, cogroup.getInput1().getLocalStrategy());
// solution set
Assert.assertEquals(LocalStrategy.NONE, cogroup.getInput2().getLocalStrategy());
// check the caches
Assert.assertTrue(TempMode.CACHED == neighborsJoin.getInput2().getTempMode());
JobGraphGenerator jgg = new JobGraphGenerator();
jgg.compileJobGraph(optPlan);
}
use of org.apache.flink.optimizer.plan.WorksetIterationPlanNode in project flink by apache.
the class PregelCompilerTest method testPregelCompilerWithBroadcastVariable.
@SuppressWarnings("serial")
@Test
public void testPregelCompilerWithBroadcastVariable() {
try {
final String BC_VAR_NAME = "borat variable";
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(DEFAULT_PARALLELISM);
// compose test program
{
DataSet<Long> bcVar = env.fromElements(1L);
DataSet<Vertex<Long, Long>> initialVertices = env.fromElements(new Tuple2<>(1L, 1L), new Tuple2<>(2L, 2L)).map(new Tuple2ToVertexMap<Long, Long>());
DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple2<>(1L, 2L)).map(new MapFunction<Tuple2<Long, Long>, Edge<Long, NullValue>>() {
public Edge<Long, NullValue> map(Tuple2<Long, Long> edge) {
return new Edge<>(edge.f0, edge.f1, NullValue.getInstance());
}
});
Graph<Long, Long, NullValue> graph = Graph.fromDataSet(initialVertices, edges, env);
VertexCentricConfiguration parameters = new VertexCentricConfiguration();
parameters.addBroadcastSet(BC_VAR_NAME, bcVar);
DataSet<Vertex<Long, Long>> result = graph.runVertexCentricIteration(new CCCompute(), null, 100, parameters).getVertices();
result.output(new DiscardingOutputFormat<Vertex<Long, Long>>());
}
Plan p = env.createProgramPlan("Pregel Connected Components");
OptimizedPlan op = compileNoStats(p);
// check the sink
SinkPlanNode sink = op.getDataSinks().iterator().next();
assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
// check the iteration
WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
// check the solution set delta
PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
assertTrue(ssDelta instanceof SingleInputPlanNode);
SingleInputPlanNode ssFlatMap = (SingleInputPlanNode) ((SingleInputPlanNode) (ssDelta)).getInput().getSource();
assertEquals(DEFAULT_PARALLELISM, ssFlatMap.getParallelism());
assertEquals(ShipStrategyType.FORWARD, ssFlatMap.getInput().getShipStrategy());
// check the computation coGroup
DualInputPlanNode computationCoGroup = (DualInputPlanNode) (ssFlatMap.getInput().getSource());
assertEquals(DEFAULT_PARALLELISM, computationCoGroup.getParallelism());
assertEquals(ShipStrategyType.FORWARD, computationCoGroup.getInput1().getShipStrategy());
assertEquals(ShipStrategyType.PARTITION_HASH, computationCoGroup.getInput2().getShipStrategy());
assertTrue(computationCoGroup.getInput2().getTempMode().isCached());
assertEquals(new FieldList(0), computationCoGroup.getInput2().getShipStrategyKeys());
// check that the initial partitioning is pushed out of the loop
assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput1().getShipStrategy());
assertEquals(new FieldList(0), iteration.getInput1().getShipStrategyKeys());
} catch (Exception e) {
System.err.println(e.getMessage());
e.printStackTrace();
fail(e.getMessage());
}
}
use of org.apache.flink.optimizer.plan.WorksetIterationPlanNode in project flink by apache.
the class SpargelCompilerTest method testSpargelCompilerWithBroadcastVariable.
@SuppressWarnings("serial")
@Test
public void testSpargelCompilerWithBroadcastVariable() {
try {
final String BC_VAR_NAME = "borat variable";
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(DEFAULT_PARALLELISM);
// compose test program
{
DataSet<Long> bcVar = env.fromElements(1L);
DataSet<Vertex<Long, Long>> initialVertices = env.fromElements(new Tuple2<>(1L, 1L), new Tuple2<>(2L, 2L)).map(new Tuple2ToVertexMap<Long, Long>());
DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple2<>(1L, 2L)).map(new MapFunction<Tuple2<Long, Long>, Edge<Long, NullValue>>() {
public Edge<Long, NullValue> map(Tuple2<Long, Long> edge) {
return new Edge<>(edge.f0, edge.f1, NullValue.getInstance());
}
});
Graph<Long, Long, NullValue> graph = Graph.fromDataSet(initialVertices, edges, env);
ScatterGatherConfiguration parameters = new ScatterGatherConfiguration();
parameters.addBroadcastSetForScatterFunction(BC_VAR_NAME, bcVar);
parameters.addBroadcastSetForGatherFunction(BC_VAR_NAME, bcVar);
DataSet<Vertex<Long, Long>> result = graph.runScatterGatherIteration(new ConnectedComponents.CCMessenger<Long, Long>(BasicTypeInfo.LONG_TYPE_INFO), new ConnectedComponents.CCUpdater<Long, Long>(), 100).getVertices();
result.output(new DiscardingOutputFormat<Vertex<Long, Long>>());
}
Plan p = env.createProgramPlan("Spargel Connected Components");
OptimizedPlan op = compileNoStats(p);
// check the sink
SinkPlanNode sink = op.getDataSinks().iterator().next();
assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
// check the iteration
WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
// check the solution set join and the delta
PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
// this is only true if the update functions preserves the partitioning
assertTrue(ssDelta instanceof DualInputPlanNode);
DualInputPlanNode ssJoin = (DualInputPlanNode) ssDelta;
assertEquals(DEFAULT_PARALLELISM, ssJoin.getParallelism());
assertEquals(ShipStrategyType.PARTITION_HASH, ssJoin.getInput1().getShipStrategy());
assertEquals(new FieldList(0), ssJoin.getInput1().getShipStrategyKeys());
// check the workset set join
DualInputPlanNode edgeJoin = (DualInputPlanNode) ssJoin.getInput1().getSource();
assertEquals(DEFAULT_PARALLELISM, edgeJoin.getParallelism());
assertEquals(ShipStrategyType.PARTITION_HASH, edgeJoin.getInput1().getShipStrategy());
assertEquals(ShipStrategyType.FORWARD, edgeJoin.getInput2().getShipStrategy());
assertTrue(edgeJoin.getInput1().getTempMode().isCached());
assertEquals(new FieldList(0), edgeJoin.getInput1().getShipStrategyKeys());
// check that the initial partitioning is pushed out of the loop
assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput1().getShipStrategy());
assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput2().getShipStrategy());
assertEquals(new FieldList(0), iteration.getInput1().getShipStrategyKeys());
assertEquals(new FieldList(0), iteration.getInput2().getShipStrategyKeys());
} catch (Exception e) {
System.err.println(e.getMessage());
e.printStackTrace();
fail(e.getMessage());
}
}
use of org.apache.flink.optimizer.plan.WorksetIterationPlanNode in project flink by apache.
the class GSACompilerTest method testGSACompiler.
@Test
public void testGSACompiler() {
try {
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(DEFAULT_PARALLELISM);
// compose test program
{
DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple3<>(1L, 2L, NullValue.getInstance())).map(new Tuple3ToEdgeMap<Long, NullValue>());
Graph<Long, Long, NullValue> graph = Graph.fromDataSet(edges, new InitVertices(), env);
DataSet<Vertex<Long, Long>> result = graph.runGatherSumApplyIteration(new GatherNeighborIds(), new SelectMinId(), new UpdateComponentId(), 100).getVertices();
result.output(new DiscardingOutputFormat<Vertex<Long, Long>>());
}
Plan p = env.createProgramPlan("GSA Connected Components");
OptimizedPlan op = compileNoStats(p);
// check the sink
SinkPlanNode sink = op.getDataSinks().iterator().next();
assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
assertEquals(PartitioningProperty.HASH_PARTITIONED, sink.getGlobalProperties().getPartitioning());
// check the iteration
WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
// check the solution set join and the delta
PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
// this is only true if the update function preserves the partitioning
assertTrue(ssDelta instanceof DualInputPlanNode);
DualInputPlanNode ssJoin = (DualInputPlanNode) ssDelta;
assertEquals(DEFAULT_PARALLELISM, ssJoin.getParallelism());
assertEquals(ShipStrategyType.PARTITION_HASH, ssJoin.getInput1().getShipStrategy());
assertEquals(new FieldList(0), ssJoin.getInput1().getShipStrategyKeys());
// check the workset set join
SingleInputPlanNode sumReducer = (SingleInputPlanNode) ssJoin.getInput1().getSource();
SingleInputPlanNode gatherMapper = (SingleInputPlanNode) sumReducer.getInput().getSource();
DualInputPlanNode edgeJoin = (DualInputPlanNode) gatherMapper.getInput().getSource();
assertEquals(DEFAULT_PARALLELISM, edgeJoin.getParallelism());
// input1 is the workset
assertEquals(ShipStrategyType.FORWARD, edgeJoin.getInput1().getShipStrategy());
// input2 is the edges
assertEquals(ShipStrategyType.PARTITION_HASH, edgeJoin.getInput2().getShipStrategy());
assertTrue(edgeJoin.getInput2().getTempMode().isCached());
assertEquals(new FieldList(0), edgeJoin.getInput2().getShipStrategyKeys());
} catch (Exception e) {
System.err.println(e.getMessage());
e.printStackTrace();
fail(e.getMessage());
}
}
use of org.apache.flink.optimizer.plan.WorksetIterationPlanNode in project flink by apache.
the class IterationsCompilerTest method testTwoIterationsDirectlyChained.
@Test
public void testTwoIterationsDirectlyChained() throws Exception {
try {
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(8);
DataSet<Tuple2<Long, Long>> verticesWithInitialId = env.fromElements(new Tuple2<Long, Long>(1L, 2L));
DataSet<Tuple2<Long, Long>> edges = env.fromElements(new Tuple2<Long, Long>(1L, 2L));
DataSet<Tuple2<Long, Long>> bulkResult = doBulkIteration(verticesWithInitialId, edges);
DataSet<Tuple2<Long, Long>> depResult = doDeltaIteration(bulkResult, edges);
depResult.output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
Plan p = env.createProgramPlan();
OptimizedPlan op = compileNoStats(p);
assertEquals(1, op.getDataSinks().size());
assertTrue(op.getDataSinks().iterator().next().getInput().getSource() instanceof WorksetIterationPlanNode);
WorksetIterationPlanNode wipn = (WorksetIterationPlanNode) op.getDataSinks().iterator().next().getInput().getSource();
BulkIterationPlanNode bipn = (BulkIterationPlanNode) wipn.getInput1().getSource();
// the hash partitioning has been pushed out of the delta iteration into the bulk iteration
assertEquals(ShipStrategyType.FORWARD, wipn.getInput1().getShipStrategy());
// since the work has been pushed out of the bulk iteration, it has to guarantee the hash partitioning
for (Channel c : bipn.getRootOfStepFunction().getInputs()) {
assertEquals(ShipStrategyType.PARTITION_HASH, c.getShipStrategy());
}
assertEquals(DataExchangeMode.BATCH, wipn.getInput1().getDataExchangeMode());
assertEquals(DataExchangeMode.BATCH, wipn.getInput2().getDataExchangeMode());
assertEquals(TempMode.NONE, wipn.getInput1().getTempMode());
assertEquals(TempMode.NONE, wipn.getInput2().getTempMode());
new JobGraphGenerator().compileJobGraph(op);
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
Aggregations