Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[backport] jvm-packages 1.6.1 (#7849)

Browse Source 
* [jvm-packages] move the dmatrix building into rabit context (#7823)

This fixes the QuantileDeviceDMatrix in distributed environment.

* [doc] update the jvm tutorial to 1.6.1 [skip ci] (#7834)

* [Breaking][jvm-packages] Use barrier execution mode (#7836)

With the introduction of the barrier execution mode. we don't need to kill SparkContext when some xgboost tasks failed. Instead, Spark will handle the errors for us. So in this PR, `killSparkContextOnWorkerFailure` parameter is deleted.

* [doc] remove the doc about killing SparkContext [skip ci] (#7840)

* [jvm-package] remove the coalesce in barrier mode (#7846)

* [jvm-packages] Fix model compatibility (#7845)

* Ignore all Java exceptions when looking for Linux musl support (#7844)

Co-authored-by: Bobby Wang <wbo4958@gmail.com>
Co-authored-by: Michael Allman <msa@allman.ms>
This commit is contained in:
Jiaming Yuan
2022-04-29 17:20:58 +08:00
committed by GitHub
parent f75c007f27
commit f4eb6b984e
24 changed files with 173 additions and 601 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
jvm-packages/xgboost4j-spark-gpu/src/main/scala/ml/dmlc/xgboost4j/scala/rapids/spark/GpuPreXGBoost.scala
View File

@@ -56,18 +56,20 @@ class GpuPreXGBoost extends PreXGBoostProvider {
}
/**
* Convert the Dataset[_] to RDD[Watches] which will be fed to XGBoost
* Convert the Dataset[_] to RDD[() => Watches] which will be fed to XGBoost
*
* @param estimator [[XGBoostClassifier]] or [[XGBoostRegressor]]
* @param dataset the training data
* @param params all user defined and defaulted params
* @return [[XGBoostExecutionParams]] => (RDD[[Watches]], Option[ RDD[_] ])
* RDD[Watches] will be used as the training input
* @return [[XGBoostExecutionParams]] => (Boolean, RDD[[() => Watches]], Option[ RDD[_] ])
* Boolean if building DMatrix in rabit context
* RDD[() => Watches] will be used as the training input
* Option[ RDD[_] ] is the optional cached RDD
*/
override def buildDatasetToRDD(estimator: Estimator[_],
dataset: Dataset[_],
params: Map[String, Any]): XGBoostExecutionParams => (RDD[Watches], Option[RDD[_]]) = {
params: Map[String, Any]):
XGBoostExecutionParams => (Boolean, RDD[() => Watches], Option[RDD[_]]) = {
GpuPreXGBoost.buildDatasetToRDD(estimator, dataset, params)
}
@@ -116,19 +118,21 @@ object GpuPreXGBoost extends PreXGBoostProvider {
}
/**
* Convert the Dataset[_] to RDD[Watches] which will be fed to XGBoost
* Convert the Dataset[_] to RDD[() => Watches] which will be fed to XGBoost
*
* @param estimator supports XGBoostClassifier and XGBoostRegressor
* @param dataset the training data
* @param params all user defined and defaulted params
* @return [[XGBoostExecutionParams]] => (RDD[[Watches]], Option[ RDD[_] ])
* RDD[Watches] will be used as the training input
* @return [[XGBoostExecutionParams]] => (Boolean, RDD[[() => Watches]], Option[ RDD[_] ])
* Boolean if building DMatrix in rabit context
* RDD[() => Watches] will be used as the training input to build DMatrix
* Option[ RDD[_] ] is the optional cached RDD
*/
override def buildDatasetToRDD(
estimator: Estimator[_],
dataset: Dataset[_],
params: Map[String, Any]): XGBoostExecutionParams => (RDD[Watches], Option[RDD[_]]) = {
params: Map[String, Any]):
XGBoostExecutionParams => (Boolean, RDD[() => Watches], Option[RDD[_]]) = {
val (Seq(labelName, weightName, marginName), feturesCols, groupName, evalSets) =
estimator match {
@@ -166,7 +170,7 @@ object GpuPreXGBoost extends PreXGBoostProvider {
xgbExecParams: XGBoostExecutionParams =>
val dataMap = prepareInputData(trainingData, evalDataMap, xgbExecParams.numWorkers,
xgbExecParams.cacheTrainingSet)
(buildRDDWatches(dataMap, xgbExecParams, evalDataMap.isEmpty), None)
(true, buildRDDWatches(dataMap, xgbExecParams, evalDataMap.isEmpty), None)
}
/**
@@ -403,14 +407,9 @@ object GpuPreXGBoost extends PreXGBoostProvider {
}
private def repartitionInputData(dataFrame: DataFrame, nWorkers: Int): DataFrame = {
// We can't check dataFrame.rdd.getNumPartitions == nWorkers here, since dataFrame.rdd is
// a lazy variable. If we call it here, we will not directly extract RDD[Table] again,
// instead, we will involve Columnar -> Row -> Columnar and decrease the performance
if (nWorkers == 1) {
dataFrame.coalesce(1)
} else {
dataFrame.repartition(nWorkers)
}
// we can't involve any coalesce operation here, since Barrier mode will check
// the RDD patterns which does not allow coalesce.
dataFrame.repartition(nWorkers)
}
private def repartitionForGroup(
@@ -448,7 +447,7 @@ object GpuPreXGBoost extends PreXGBoostProvider {
private def buildRDDWatches(
dataMap: Map[String, ColumnDataBatch],
xgbExeParams: XGBoostExecutionParams,
noEvalSet: Boolean): RDD[Watches] = {
noEvalSet: Boolean): RDD[() => Watches] = {
val sc = dataMap(TRAIN_NAME).rawDF.sparkSession.sparkContext
val maxBin = xgbExeParams.toMap.getOrElse("max_bin", 256).asInstanceOf[Int]
@@ -459,7 +458,7 @@ object GpuPreXGBoost extends PreXGBoostProvider {
GpuUtils.toColumnarRdd(dataMap(TRAIN_NAME).rawDF).mapPartitions({
iter =>
val iterColBatch = iter.map(table => new GpuColumnBatch(table, null))
Iterator(buildWatches(
Iterator(() => buildWatches(
PreXGBoost.getCacheDirName(xgbExeParams.useExternalMemory), xgbExeParams.missing,
colIndicesForTrain, iterColBatch, maxBin))
})
@@ -469,7 +468,7 @@ object GpuPreXGBoost extends PreXGBoostProvider {
val nameAndColIndices = dataMap.map(nc => (nc._1, nc._2.colIndices))
coPartitionForGpu(dataMap, sc, xgbExeParams.numWorkers).mapPartitions {
nameAndColumnBatchIter =>
Iterator(buildWatchesWithEval(
Iterator(() => buildWatchesWithEval(
PreXGBoost.getCacheDirName(xgbExeParams.useExternalMemory), xgbExeParams.missing,
nameAndColIndices, nameAndColumnBatchIter, maxBin))
}

12
jvm-packages/xgboost4j-spark-gpu/src/test/scala/ml/dmlc/xgboost4j/scala/rapids/spark/GpuTestSuite.scala
View File

@@ -39,13 +39,8 @@ trait GpuTestSuite extends FunSuite with TmpFolderSuite {
def enableCsvConf(): SparkConf = {
new SparkConf()
.set(RapidsConf.ENABLE_READ_CSV_DATES.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_BYTES.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_SHORTS.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_INTEGERS.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_LONGS.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_FLOATS.key, "true")
.set(RapidsConf.ENABLE_READ_CSV_DOUBLES.key, "true")
.set("spark.rapids.sql.csv.read.float.enabled", "true")
.set("spark.rapids.sql.csv.read.double.enabled", "true")
}
def withGpuSparkSession[U](conf: SparkConf = new SparkConf())(f: SparkSession => U): U = {
@@ -246,12 +241,13 @@ object SparkSessionHolder extends Logging {
Locale.setDefault(Locale.US)
val builder = SparkSession.builder()
.master("local[1]")
.master("local[2]")
.config("spark.sql.adaptive.enabled", "false")
.config("spark.rapids.sql.enabled", "false")
.config("spark.rapids.sql.test.enabled", "false")
.config("spark.plugins", "com.nvidia.spark.SQLPlugin")
.config("spark.rapids.memory.gpu.pooling.enabled", "false") // Disable RMM for unit tests.
.config("spark.sql.files.maxPartitionBytes", "1000")
.appName("XGBoost4j-Spark-Gpu unit test")
builder.getOrCreate()