Copilot commented on code in PR #13115:
URL: https://github.com/apache/gluten/pull/13115#discussion_r4244547131
##########
cpp/velox/compute/VeloxBackend.cc:
##########
@@ -254,6 +255,10 @@ void VeloxBackend::init(
std::make_unique<folly::CPUThreadPoolExecutor>(ioThreads,
folly::CPUThreadPoolExecutor::makeLifoSemQueue());
}
+ const auto broadcastHashTableBuildThreads =
+ std::max<int32_t>(1,
backendConf_->get<int32_t>(kBroadcastHashTableBuildThreads, 1));
+ broadcastHashTableBuildExecutor_ =
std::make_unique<folly::CPUThreadPoolExecutor>(broadcastHashTableBuildThreads);
Review Comment:
The dedicated `CPUThreadPoolExecutor` is created unconditionally with at
least one thread per backend instance, even when parallel broadcast hash-table
building may never be used. This increases thread count and idle resource
overhead per executor process. Consider lazily constructing this executor only
when parallel join build is enabled/needed, or allow an explicit config value
(e.g., 0) to keep it `nullptr` and fall back to inline execution when parallel
build is disabled.
##########
docs/velox-configuration.md:
##########
@@ -16,6 +16,7 @@ nav_order: 16
| spark.gluten.sql.columnar.backend.velox.abandonPartialAggregationMinPct
| 🔄 Dynamic | 90 | If partial aggregation aggregationPct
greater than this value, partial aggregation may be early abandoned. Note: this
option only works when flushable partial aggregation is enabled. Ignored when
spark.gluten.sql.columnar.backend.velox.flushablePartialAggregation=false.
|
| spark.gluten.sql.columnar.backend.velox.abandonPartialAggregationMinRows
| 🔄 Dynamic | 100000 | If partial aggregation input rows
number greater than this value, partial aggregation may be early abandoned.
Note: this option only works when flushable partial aggregation is enabled.
Ignored when
spark.gluten.sql.columnar.backend.velox.flushablePartialAggregation=false.
|
| spark.gluten.sql.columnar.backend.velox.asyncTimeoutOnTaskStopping
| âš“ Static | 30000ms | Timeout in milliseconds when waiting
for runtime-scoped async work to finish during teardown.
|
+| spark.gluten.sql.columnar.backend.velox.broadcastHashTableBuildThreads
| âš“ Static | <undefined> | Size of the dedicated CPU thread pool
for the parallel broadcast hash-table build. By default, the value is the same
as the maximum task slots per Spark executor, with a minimum of one thread.
|
Review Comment:
The table's 'Default' column says `<undefined>`, but the description states
there is an effective default derived from task slots (min 1). To avoid
confusing users, update the 'Default' column to reflect the derived default
behavior (e.g., `maxTaskSlotsPerExecutor (min 1)`) or clarify that the Spark
conf is optional but the native layer computes a default.
##########
gluten-ut/spark35/src/test/backends-velox/org/apache/spark/sql/execution/joins/GlutenParallelBroadcastJoinSuite.scala:
##########
@@ -0,0 +1,64 @@
+/*
+ * 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.joins
+
+import org.apache.gluten.execution.BroadcastHashJoinExecTransformer
+
+import org.apache.spark.SparkConf
+import org.apache.spark.sql.GlutenSQLTestsTrait
+import org.apache.spark.sql.execution.ColumnarBroadcastExchangeExec
+
+class GlutenParallelBroadcastJoinSuite extends GlutenSQLTestsTrait {
+
+ override def sparkConf: SparkConf = super.sparkConf
+ .set("spark.sql.adaptive.enabled", "false")
+ .set("spark.gluten.sql.columnar.backend.velox.IOThreads", "0")
+ .set("spark.gluten.velox.broadcast.build.targetBytesPerThread", "8KB")
+
+ for (buildOnDriver <- Seq(false, true)) {
+ testGluten(s"parallel broadcast hash table build with IO disabled,
driver=$buildOnDriver") {
+ withSQLConf(
+
"spark.gluten.sql.columnar.backend.velox.driverSideBroadcastHashTableBuild" ->
+ buildOnDriver.toString,
+ "spark.gluten.velox.minTableRowsForParallelJoinBuild" -> "1"
+ ) {
+ withTable("parallel_broadcast_build") {
+ spark.range(4096)
+ .selectExpr("id", "sha2(cast(id as string), 256) AS value")
+ .write.saveAsTable("parallel_broadcast_build")
+
+ val result = spark.sql("""
+ |SELECT /*+ BROADCAST(b) */ p.id, b.value
+ |FROM parallel_broadcast_build p JOIN
parallel_broadcast_build b
+ |ON p.id = b.id
+ |""".stripMargin)
+ val expected =
spark.table("parallel_broadcast_build").collect().toSeq
+ checkAnswer(result, expected)
+
+ val plan = result.queryExecution.executedPlan
+ assert(plan.collect { case j: BroadcastHashJoinExecTransformer => j
}.nonEmpty)
+ val exchanges = plan.collect { case e: ColumnarBroadcastExchangeExec
=> e }
+ assert(exchanges.size == 1)
+ assert(exchanges.head.metrics("buildThreads").value > 1)
Review Comment:
This assertion can be brittle because the suite never pins the broadcast
hash-table build thread-pool size. If the default resolves to 1 (e.g., due to
task-slot/core settings in the test environment), the test will fail even
though the feature works. Consider setting
`spark.gluten.sql.columnar.backend.velox.broadcastHashTableBuildThreads`
explicitly (e.g., to 2+) within `sparkConf` or `withSQLConf` for this test so
the `> 1` expectation is guaranteed.
--
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]
