jackylee-ch commented on code in PR #8209:
URL: https://github.com/apache/incubator-gluten/pull/8209#discussion_r1897537913
##########
shims/common/src/main/scala/org/apache/gluten/GlutenConfig.scala:
##########
@@ -483,6 +483,18 @@ class GlutenConfig(conf: SQLConf) extends Logging {
def enableCelebornFallback: Boolean = conf.getConf(CELEBORN_FALLBACK_ENABLED)
def enableHdfsViewfs: Boolean = conf.getConf(HDFS_VIEWFS_ENABLED)
+
+ def enableAutoAdjustStageResourceProfile: Boolean =
+ conf.getConf(AUTO_ADJUST_STAGE_RESOURCE_PROFILE_ENABLED) &&
conf.adaptiveExecutionEnabled
Review Comment:
nit: move `conf.adaptiveExecutionEnabled` to the [condition
check](https://github.com/apache/incubator-gluten/pull/8209/files#diff-0ebcffe98eb315f653dab59bd7638ca0c613e83d3733917832f7b0f2ca3f8edcR50).
##########
gluten-substrait/src/main/scala/org/apache/spark/sql/execution/GlutenAutoAdjustStageResourceProfile.scala:
##########
@@ -0,0 +1,169 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You 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 org.apache.spark.sql.execution
+
+import org.apache.gluten.GlutenConfig
+import org.apache.gluten.execution.{ColumnarToRowExecBase, GlutenPlan}
+import org.apache.gluten.logging.LogLevelUtil
+
+import org.apache.spark.annotation.Experimental
+import org.apache.spark.internal.Logging
+import org.apache.spark.resource.{ExecutorResourceRequest, ResourceProfile,
ResourceProfileManager, TaskResourceRequest}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.rules.Rule
+import
org.apache.spark.sql.execution.GlutenAutoAdjustStageResourceProfile.{applyNewResourceProfileIfPossible,
collectStagePlan}
+import org.apache.spark.sql.execution.adaptive.QueryStageExec
+import org.apache.spark.sql.execution.command.{DataWritingCommandExec,
ExecutedCommandExec}
+import org.apache.spark.sql.execution.exchange.Exchange
+import org.apache.spark.util.SparkTestUtil
+
+import scala.collection.mutable
+import scala.collection.mutable.ArrayBuffer
+
+/**
+ * This rule is used to dynamic adjust stage resource profile for following
purposes:
+ * 1. swap offheap and onheap memory size when whole stage fallback happened
2. increase executor
+ * heap memory if stage contains gluten operator and spark operator at
the same time. Note: we
+ * don't support set resource profile for final stage now. Todo: support
set resource profile
+ * for final stage.
+ */
+@Experimental
+case class GlutenAutoAdjustStageResourceProfile(glutenConf: GlutenConfig,
spark: SparkSession)
+ extends Rule[SparkPlan]
+ with LogLevelUtil {
+
+ override def apply(plan: SparkPlan): SparkPlan = {
+ if (!glutenConf.enableAutoAdjustStageResourceProfile) {
+ return plan
+ }
+ if (!plan.isInstanceOf[Exchange]) {
+ // todo: support set resource profile for final stage
+ return plan
+ }
+ val planNodes = collectStagePlan(plan)
+ if (planNodes.isEmpty) {
+ return plan
+ }
+ log.info(s"detailPlanNodes ${planNodes.map(_.nodeName).mkString("Array(",
", ", ")")}")
+
+ // one stage is considered as fallback if all node is not GlutenPlan
+ // or all GlutenPlan node is C2R node.
+ val wholeStageFallback = planNodes
+ .filter(_.isInstanceOf[GlutenPlan])
+ .count(!_.isInstanceOf[ColumnarToRowExecBase]) == 0
+
+ val rpManager = spark.sparkContext.resourceProfileManager
+ val defaultRP = rpManager.defaultResourceProfile
+
+ // initial resource profile config as default resource profile
+ val taskResource = mutable.Map.empty[String, TaskResourceRequest] ++=
defaultRP.taskResources
+ val executorResource =
+ mutable.Map.empty[String, ExecutorResourceRequest] ++=
defaultRP.executorResources
+ val memoryRequest = executorResource.get(ResourceProfile.MEMORY)
+ val offheapRequest = executorResource.get(ResourceProfile.OFFHEAP_MEM)
+ logInfo(s"default memory request $memoryRequest")
+ logInfo(s"default offheap request $offheapRequest")
+
+ // case 1: whole stage fallback to vanilla spark in such case we swap the
heap
+ // and offheap amount.
Review Comment:
can we calculate the new `onHeapMemory` with
`glutenConf.autoAdjustStageRPHeapRatio` that similar to case 2?
Memory swapping does not seem to be a particularly good idea, because we
cannot control the actual effect of the swap. In some scenarios, the onHeap is
set too small, for example, 1G, which results in only a small amount of memory
being obtained by offHeap after the swap. In other scenarios, the onHeap and
offHeap settings are not much different, and in the end, no better memory
settings are obtained.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
