Add all stage metrics to tools output (#1151)

* Add Gpu metrics to profiling tool output Signed-off-by: Niranjan Artal <nartal@nvidia.com> * add file Signed-off-by: Niranjan Artal <nartal@nvidia.com> * addressed review comments * address review comments * addressed review comments, removed unwanted code. --------- Signed-off-by: Niranjan Artal <nartal@nvidia.com>
NVIDIA · Jul 3, 2024 · 91fed9d · 91fed9d
1 parent 93e512e
commit 91fed9d
Show file tree

Hide file tree

Showing 8 changed files with 197 additions and 6 deletions.
diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/analysis/AppSQLPlanAnalyzer.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/analysis/AppSQLPlanAnalyzer.scala
@@ -19,7 +19,7 @@ package com.nvidia.spark.rapids.tool.analysis
 import scala.collection.mutable.{AbstractSet, ArrayBuffer, HashMap, LinkedHashSet}
 
 import com.nvidia.spark.rapids.tool.planparser.SQLPlanParser
-import com.nvidia.spark.rapids.tool.profiling.{DataSourceCase, SQLAccumProfileResults, SQLMetricInfoCase, SQLStageInfoProfileResult, UnsupportedSQLPlan, WholeStageCodeGenResults}
+import com.nvidia.spark.rapids.tool.profiling.{AccumProfileResults, DataSourceCase, SQLAccumProfileResults, SQLMetricInfoCase, SQLStageInfoProfileResult, UnsupportedSQLPlan, WholeStageCodeGenResults}
 import com.nvidia.spark.rapids.tool.qualification.QualSQLPlanAnalyzer
 
 import org.apache.spark.sql.execution.SparkPlanInfo
@@ -342,6 +342,58 @@ class AppSQLPlanAnalyzer(app: AppBase, appIndex: Int) extends AppAnalysisBase(ap
       }
     }
   }
+
+  /**
+   * Generate the stage level metrics for the SQL plan including GPU metrics if applicable.
+   * Along with Spark defined metrics, below is the list of GPU metrics that are collected if they
+   * are present in the eventlog:
+   * gpuSemaphoreWait, gpuRetryCount, gpuSplitAndRetryCount, gpuRetryBlockTime,
+   * gpuRetryComputationTime, gpuSpillToHostTime, gpuSpillToDiskTime, gpuReadSpillFromHostTime,
+   * gpuReadSpillFromDiskTime
+   *
+   * @return a sequence of AccumProfileResults
+   */
+  def generateStageLevelAccums(): Seq[AccumProfileResults] = {
+
+    def computeStatistics(updates: Seq[Long]): Option[StatisticsMetrics] = {
+      // drop the metrics if there are no values
+      if (updates.isEmpty) {
+        None
+      } else if (updates.length == 1) {
+        Some(StatisticsMetrics(0L, 0L, 0L, updates.sum))
+      } else {
+        Some(StatisticsMetrics(
+          min = updates.head,
+          med = updates(updates.size / 2),
+          max = updates.last,
+          total = updates.sum
+        ))
+      }
+    }
+
+    // Process taskStageAccumMap to get all the accumulators
+    val stageLevelAccums = app.taskStageAccumMap.values.flatten
+    val groupedByAccumulatorId = stageLevelAccums.groupBy(_.accumulatorId)
+    groupedByAccumulatorId.flatMap { case (accumulatorId, accums) =>
+      // Extract and sort the update values, defaulting to 0 if not present
+      val sortedUpdates = accums.flatMap(_.update).toSeq.sorted
+
+      // Compute the statistics for the accumulator if applicable
+      computeStatistics(sortedUpdates).map { stats =>
+        val sampleAccum = accums.head
+        AccumProfileResults(
+          appIndex = appIndex,
+          stageId = sampleAccum.stageId.toString,
+          accumulatorId = accumulatorId,
+          name = sampleAccum.name.getOrElse("Unknown"),
+          min = stats.min,
+          median = stats.med,
+          max = stats.max,
+          total = stats.total
+        )
+      }
+    }.toSeq
+  }
 }
 
 object AppSQLPlanAnalyzer {

diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/ApplicationSummaryInfo.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/ApplicationSummaryInfo.scala
@@ -26,6 +26,7 @@ case class ApplicationSummaryInfo(
     rapidsProps: Seq[RapidsPropertyProfileResult],
     rapidsJar: Seq[RapidsJarProfileResult],
     sqlMetrics: Seq[SQLAccumProfileResults],
+    stageMetrics: Seq[AccumProfileResults],
     jobAggMetrics: Seq[JobAggTaskMetricsProfileResult],
     stageAggMetrics: Seq[StageAggTaskMetricsProfileResult],
     sqlTaskAggMetrics: Seq[SQLTaskAggMetricsProfileResult],

diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/CollectInformation.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/CollectInformation.scala
@@ -84,6 +84,11 @@ class CollectInformation(apps: Seq[ApplicationInfo]) extends Logging {
     ProfSQLPlanMetricsView.getRawView(apps)
   }
 
+  // Print all Stage level Metrics
+  def getStageLevelMetrics: Seq[AccumProfileResults] = {
+    ProfStageMetricView.getRawView(apps)
+  }
+
   /**
    * This function is meant to clean up Delta log execs so that you could align
    * SQL ids between CPU and GPU eventlogs. It attempts to remove any delta log
@@ -103,6 +108,12 @@ object CollectInformation extends Logging {
     }
   }
 
+  def generateStageLevelAccums(apps: Seq[ApplicationInfo]): Seq[AccumProfileResults] = {
+    apps.flatMap { app =>
+      app.planMetricProcessor.generateStageLevelAccums()
+    }
+  }
+
   def printSQLPlans(apps: Seq[ApplicationInfo], outputDir: String): Unit = {
     for (app <- apps) {
       val planFileWriter = new ToolTextFileWriter(s"$outputDir/${app.appId}",

diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/ProfileClassWarehouse.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/ProfileClassWarehouse.scala
@@ -219,6 +219,23 @@ case class SQLAccumProfileResults(appIndex: Int, sqlID: Long, nodeID: Long,
   }
 }
 
+case class AccumProfileResults(appIndex: Int, stageId: String, accumulatorId: Long,  name: String,
+    min: Long, median: Long, max: Long, total: Long) extends ProfileResult {
+  override val outputHeaders = Seq("appIndex", "stageId", "accumulatorId", "name", "min",
+    "median", "max", "total")
+
+  override def convertToSeq: Seq[String] = {
+    Seq(appIndex.toString, stageId, accumulatorId.toString, name, min.toString, median.toString,
+      max.toString, total.toString)
+  }
+
+  override def convertToCSVSeq: Seq[String] = {
+    Seq(appIndex.toString, StringUtils.reformatCSVString(stageId), accumulatorId.toString,
+      StringUtils.reformatCSVString(name), min.toString, median.toString, max.toString,
+      total.toString)
+  }
+}
+
 case class ResourceProfileInfoCase(
     val resourceProfileId: Int,
     val executorResources: Map[String, ExecutorResourceRequest],

diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/Profiler.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/profiling/Profiler.scala
@@ -329,6 +329,7 @@ class Profiler(hadoopConf: Configuration, appArgs: ProfileArgs, enablePB: Boolea
     val systemProps = collect.getSystemProperties
     val rapidsJar = collect.getRapidsJARInfo
     val sqlMetrics = collect.getSQLPlanMetrics
+    val stageMetrics = collect.getStageLevelMetrics
     val wholeStage = collect.getWholeStageCodeGenMapping
     // for compare mode we just add in extra tables for matching across applications
     // the rest of the tables simply list all applications specified
@@ -392,7 +393,7 @@ class Profiler(hadoopConf: Configuration, appArgs: ProfileArgs, enablePB: Boolea
     logInfo(s"Took ${endTime - startTime}ms to Process [${appInfo.head.appId}]")
     (ApplicationSummaryInfo(appInfo, dsInfo,
       collect.getExecutorInfo, collect.getJobInfo, rapidsProps,
-      rapidsJar, sqlMetrics, analysis.jobAggs, analysis.stageAggs,
+      rapidsJar, sqlMetrics, stageMetrics, analysis.jobAggs, analysis.stageAggs,
       analysis.sqlAggs, analysis.sqlDurAggs, analysis.taskShuffleSkew,
       failedTasks, failedStages, failedJobs, removedBMs, removedExecutors,
       unsupportedOps, sparkProps, collect.getSQLToStage, wholeStage, maxTaskInputInfo,
@@ -471,6 +472,7 @@ class Profiler(hadoopConf: Configuration, appArgs: ProfileArgs, enablePB: Boolea
         combineProps("rapids", appsSum).sortBy(_.key),
         appsSum.flatMap(_.rapidsJar).sortBy(_.appIndex),
         appsSum.flatMap(_.sqlMetrics).sortBy(_.appIndex),
+        appsSum.flatMap(_.stageMetrics).sortBy(_.appIndex),
         appsSum.flatMap(_.jobAggMetrics).sortBy(_.appIndex),
         appsSum.flatMap(_.stageAggMetrics).sortBy(_.appIndex),
         appsSum.flatMap(_.sqlTaskAggMetrics).sortBy(_.appIndex),
@@ -513,6 +515,8 @@ class Profiler(hadoopConf: Configuration, appArgs: ProfileArgs, enablePB: Boolea
         Some(ProfRapidsJarView.getDescription))
       profileOutputWriter.write(ProfSQLPlanMetricsView.getLabel, app.sqlMetrics,
         Some(ProfSQLPlanMetricsView.getDescription))
+      profileOutputWriter.write(ProfStageMetricView.getLabel, app.stageMetrics,
+        Some(ProfStageMetricView.getDescription))
       profileOutputWriter.write(ProfSQLCodeGenView.getLabel, app.wholeStage,
         Some(ProfSQLCodeGenView.getDescription))
       comparedRes.foreach { compareSum =>

diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/views/AppStageMetricsView.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/views/AppStageMetricsView.scala
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.nvidia.spark.rapids.tool.views
+
+import com.nvidia.spark.rapids.tool.analysis.{AppSQLPlanAnalyzer, ProfAppIndexMapperTrait}
+import com.nvidia.spark.rapids.tool.profiling.AccumProfileResults
+
+import org.apache.spark.sql.rapids.tool.AppBase
+import org.apache.spark.sql.rapids.tool.annotation.Since
+import org.apache.spark.sql.rapids.tool.profiling.ApplicationInfo
+
+trait AppStageMetricsViewTrait extends ViewableTrait[AccumProfileResults] {
+  override def getLabel: String = "Stage Level All Metrics"
+  override def getDescription: String = "Stage Level Metrics"
+
+  override def sortView(
+      rows: Seq[AccumProfileResults]): Seq[AccumProfileResults] = {
+    rows.sortBy(cols => (cols.stageId, cols.appIndex, cols.accumulatorId))
+  }
+}
+
+@Since("24.06.2")
+object ProfStageMetricView extends AppStageMetricsViewTrait with ProfAppIndexMapperTrait {
+
+  override def getRawView(app: AppBase, index: Int): Seq[AccumProfileResults] = {
+    app match {
+      case app: ApplicationInfo =>
+        app.planMetricProcessor.generateStageLevelAccums()
+      case _ => Seq.empty
+    }
+  }
+}
+
+object QualStageMetricView extends AppStageMetricsViewTrait with ProfAppIndexMapperTrait {
+  // Keep for the following refactor to customize the view based on the app type (Qual/Prof)
+  override def getRawView(app: AppBase, index: Int): Seq[AccumProfileResults] = {
+    Seq.empty
+  }
+
+  def getRawViewFromSqlProcessor(
+      sqlAnalyzer: AppSQLPlanAnalyzer): Seq[AccumProfileResults] = {
+    sortView(sqlAnalyzer.generateStageLevelAccums())
+  }
+}
diff --git a/core/src/main/scala/com/nvidia/spark/rapids/tool/views/QualRawReportGenerator.scala b/core/src/main/scala/com/nvidia/spark/rapids/tool/views/QualRawReportGenerator.scala
@@ -75,6 +75,9 @@ object QualRawReportGenerator {
       pWriter.write(QualExecutorView.getLabel, QualExecutorView.getRawView(Seq(app)))
       pWriter.write(QualAppJobView.getLabel, QualAppJobView.getRawView(Seq(app)))
       generateSQLProcessingView(pWriter, sqlPlanAnalyzer)
+      pWriter.write(QualStageMetricView.getLabel,
+        QualStageMetricView.getRawViewFromSqlProcessor(sqlPlanAnalyzer),
+        Some(QualStageMetricView.getDescription))
       pWriter.write(RapidsQualPropertiesView.getLabel,
         RapidsQualPropertiesView.getRawView(Seq(app)),
         Some(RapidsQualPropertiesView.getDescription))

diff --git a/core/src/test/scala/com/nvidia/spark/rapids/tool/profiling/ApplicationInfoSuite.scala b/core/src/test/scala/com/nvidia/spark/rapids/tool/profiling/ApplicationInfoSuite.scala
@@ -17,6 +17,7 @@
 package com.nvidia.spark.rapids.tool.profiling
 
 import java.io.File
+import java.nio.charset.StandardCharsets
 import java.nio.file.{Files, Paths, StandardOpenOption}
 
 import scala.collection.mutable.ArrayBuffer
@@ -202,6 +203,50 @@ class ApplicationInfoSuite extends FunSuite with Logging {
     ToolTestUtils.compareDataFrames(df, dfExpect)
   }
 
+  test("test GpuMetrics in eventlog") {
+    TrampolineUtil.withTempDir { outputDir =>
+      TrampolineUtil.withTempDir { tmpEventLogDir =>
+        val eventLogFilePath = Paths.get(tmpEventLogDir.getAbsolutePath, "gpu_metrics_eventlog")
+        // scalastyle:off line.size.limit
+        val eventLogContent =
+          """{"Event":"SparkListenerLogStart","Spark Version":"3.2.1"}
+            |{"Event":"SparkListenerApplicationStart","App Name":"GPUMetrics", "App ID":"local-16261043003", "Timestamp":123456, "User":"User1"}
+            |{"Event":"SparkListenerTaskEnd","Stage ID":10,"Stage Attempt ID":0,"Task Type":"ShuffleMapTask","Task End Reason":{"Reason":"Success"},"Task Info":{"Task ID":5073,"Index":5054,"Attempt":0,"Partition ID":5054,"Launch Time":1712248533994,"Executor ID":"100","Host":"10.154.65.143","Locality":"PROCESS_LOCAL","Speculative":false,"Getting Result Time":0,"Finish Time":1712253284920,"Failed":false,"Killed":false,"Accumulables":[{"ID":1010,"Name":"gpuSemaphoreWait","Update":"00:00:00.492","Value":"03:13:31.359","Internal":false,"Count Failed Values":true},{"ID":1018,"Name":"gpuSpillToHostTime","Update":"00:00:00.845","Value":"00:29:39.521","Internal":false,"Count Failed Values":true},{"ID":1016,"Name":"gpuSplitAndRetryCount","Update":"1","Value":"2","Internal":false,"Count Failed Values":true}]}}""".stripMargin
+        // scalastyle:on line.size.limit
+        Files.write(eventLogFilePath, eventLogContent.getBytes(StandardCharsets.UTF_8))
+
+        val profileArgs = Array(
+          "--output-directory", outputDir.getAbsolutePath,
+          eventLogFilePath.toString
+        )
+
+        val appArgs = new ProfileArgs(profileArgs)
+        val (exit, _) = ProfileMain.mainInternal(appArgs)
+        assert(exit == 0)
+
+        val apps = ArrayBuffer[ApplicationInfo]()
+        var index = 1
+
+        val eventLogPaths = appArgs.eventlog()
+        eventLogPaths.foreach { path =>
+          val eventLogInfo = EventLogPathProcessor.getEventLogInfo(path, hadoopConf).head._1
+          apps += new ApplicationInfo(hadoopConf, eventLogInfo, index)
+          index += 1
+        }
+        assert(apps.size == 1)
+
+        val collect = new CollectInformation(apps)
+        val gpuMetrics = collect.getStageLevelMetrics
+
+        // Sample eventlog has 3 gpu metrics, gpuSemaphoreWait,
+        // gpuSpillToHostTime, gpuSplitAndRetryCount
+        assert(gpuMetrics.size == 3)
+        val gpuSemaphoreWait = gpuMetrics.find(_.name == "gpuSemaphoreWait")
+        assert(gpuSemaphoreWait.isDefined)
+      }
+    }
+  }
+
   test("test printSQLPlans") {
     TrampolineUtil.withTempDir { tempOutputDir =>
       val apps: ArrayBuffer[ApplicationInfo] = ArrayBuffer[ApplicationInfo]()
@@ -790,7 +835,7 @@ class ApplicationInfoSuite extends FunSuite with Logging {
         f.endsWith(".csv")
       })
       // compare the number of files generated
-      assert(dotDirs.length === 19)
+      assert(dotDirs.length === 20)
       for (file <- dotDirs) {
         assert(file.getAbsolutePath.endsWith(".csv"))
         // just load each one to make sure formatted properly
@@ -824,7 +869,7 @@ class ApplicationInfoSuite extends FunSuite with Logging {
         f.endsWith(".csv")
       })
       // compare the number of files generated
-      assert(dotDirs.length === 15)
+      assert(dotDirs.length === 16)
       for (file <- dotDirs) {
         assert(file.getAbsolutePath.endsWith(".csv"))
         // just load each one to make sure formatted properly
@@ -861,7 +906,7 @@ class ApplicationInfoSuite extends FunSuite with Logging {
         f.endsWith(".csv")
       })
       // compare the number of files generated
-      assert(dotDirs.length === 19)
+      assert(dotDirs.length === 20)
       for (file <- dotDirs) {
         assert(file.getAbsolutePath.endsWith(".csv"))
         // just load each one to make sure formatted properly
@@ -898,7 +943,7 @@ class ApplicationInfoSuite extends FunSuite with Logging {
         f.endsWith(".csv")
       })
       // compare the number of files generated
-      assert(dotDirs.length === 17)
+      assert(dotDirs.length === 18)
       for (file <- dotDirs) {
         assert(file.getAbsolutePath.endsWith(".csv"))
         // just load each one to make sure formatted properly