zhztheplayer commented on code in PR #8127: URL: https://github.com/apache/incubator-gluten/pull/8127#discussion_r1898481261
########## backends-velox/src/main/scala/org/apache/spark/sql/execution/unsafe/UnsafeColumnarBuildSideRelation.scala: ########## @@ -0,0 +1,312 @@ +/* + * 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.unsafe + +import org.apache.gluten.backendsapi.BackendsApiManager +import org.apache.gluten.columnarbatch.ColumnarBatches +import org.apache.gluten.iterator.Iterators +import org.apache.gluten.memory.arrow.alloc.ArrowBufferAllocators +import org.apache.gluten.runtime.Runtimes +import org.apache.gluten.sql.shims.SparkShimLoader +import org.apache.gluten.utils.ArrowAbiUtil +import org.apache.gluten.vectorized.{ColumnarBatchSerializerJniWrapper, NativeColumnarToRowJniWrapper} + +import org.apache.spark.{SparkEnv, TaskContext} +import org.apache.spark.internal.Logging +import org.apache.spark.memory.{TaskMemoryManager, UnifiedMemoryManager} +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions.{Attribute, Expression, UnsafeProjection, UnsafeRow} +import org.apache.spark.sql.catalyst.plans.physical.{BroadcastMode, IdentityBroadcastMode} +import org.apache.spark.sql.execution.joins.{BuildSideRelation, HashedRelationBroadcastMode} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.utils.SparkArrowUtil +import org.apache.spark.sql.vectorized.ColumnarBatch +import org.apache.spark.task.TaskResources +import org.apache.spark.util.Utils + +import com.esotericsoftware.kryo.{Kryo, KryoSerializable} +import com.esotericsoftware.kryo.io.{Input, Output} +import org.apache.arrow.c.ArrowSchema + +import java.io.{Externalizable, ObjectInput, ObjectOutput} + +import scala.collection.JavaConverters.asScalaIteratorConverter + +/** + * UnsafeColumnarBuildSideRelation should backed by offheap to avoid on-heap oom. Almost the same as + * ColumnarBuildSideRelation, we should remove ColumnarBuildSideRelation when + * UnsafeColumnarBuildSideRelation get matured. + * + * @param output + * @param batches + */ +case class UnsafeColumnarBuildSideRelation( + private var output: Seq[Attribute], + private var batches: UnsafeBytesBufferArray, + var mode: BroadcastMode) + extends BuildSideRelation + with Externalizable + with Logging + with KryoSerializable { + + // Needed for serialization + def this() = { + this(null, null.asInstanceOf[UnsafeBytesBufferArray], null) + } + + def this(output: Seq[Attribute], bytesBufferArray: Array[Array[Byte]], mode: BroadcastMode) = { + // only used in driver side when broadcast the whole batches + this( + output, + UnsafeBytesBufferArray( + bytesBufferArray.length, + bytesBufferArray.map(_.length), + bytesBufferArray.map(_.length.toLong).sum, + TaskContext.get().taskMemoryManager + ), + mode + ) + val batchesSize = bytesBufferArray.length + for (i <- 0 until batchesSize) { + val length = bytesBufferArray(i).length + log.debug(s"this $i--- $length") + batches.putBytesBuffer(i, bytesBufferArray(i)) + } + } + + // should only be used on driver to serialize this relation + override def writeExternal(out: ObjectOutput): Unit = Utils.tryOrIOException { + out.writeObject(output) + out.writeObject(mode) + out.writeInt(batches.arraySize) + out.writeObject(batches.bytesBufferLengths) + out.writeLong(batches.totalBytes) + for (i <- 0 until batches.arraySize) { + val bytes = batches.getBytesBuffer(i) + out.write(bytes) + log.debug(s"writeExternal index $i with length ${bytes.length}") + } + } + + // should only be used on driver to serialize this relation + override def write(kryo: Kryo, out: Output): Unit = Utils.tryOrIOException { + kryo.writeObject(out, output.toList) + kryo.writeObject(out, mode) + out.writeInt(batches.arraySize) + kryo.writeObject(out, batches.bytesBufferLengths) + out.writeLong(batches.totalBytes) + for (i <- 0 until batches.arraySize) { + val bytes = batches.getBytesBuffer(i) + out.write(bytes) + log.debug(s"write index $i with length ${bytes.length}") + } + } + + // should only be used on executor to deserialize this relation + override def readExternal(in: ObjectInput): Unit = Utils.tryOrIOException { + output = in.readObject().asInstanceOf[Seq[Attribute]] + mode = in.readObject().asInstanceOf[BroadcastMode] + val totalArraySize = in.readInt() + val bytesBufferLengths = in.readObject().asInstanceOf[Array[Int]] + val totalBytes = in.readLong() + + val taskMemoryManager = new TaskMemoryManager( + new UnifiedMemoryManager(SparkEnv.get.conf, Long.MaxValue, Long.MaxValue / 2, 1), + 0) Review Comment: @yikf Thank you for continuing the work. I was on vacation either today so sorry for the late reply. > @zhztheplayer Hi, terry is on vacation recently. What's your opinion on this idea? I can continue to follow up. I was thinking about using storage memory but the implementation is complicated? If so let's put that idea on hold. > we can directly use SparkEnv.get.memoryManager. 👍 Sounds good to me if it can utilize Spark's off-heap memory, since at least the memory is tracked. Though the release could still be a problem? Because Spark doesn't do `System.gc()` on off-heap OOM. If the issue does apply, maybe we can consider one of the following: Option 1. Hack into Spark to make sure `System.gc()` is triggered on OOM. (if there's an apporach without modifications on vanilla Spark's code) Option 2. Set a fixed capacity (e.g., 15% of off-heap memory) for the outstanding task memory manager used in the PR. When this part of memory is run out, trigger `System.gc()`, then if still unsatisified, throw an OOM. The above are just based on my assumption. Let me know if any other possible solutions. -- 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]
