Enable kudo serializer by default #12222
Conversation
Signed-off-by: liurenjie1024 <liurenjie2008@gmail.com>
|
build |
There was a problem hiding this comment.
@liurenjie1024 can you provide performance numbers for NDS with and without MT shuffle?
|
build |
| @@ -500,7 +500,7 @@ class AdaptiveQueryExecSuite | |||
| val conf = new SparkConf() | |||
| .set(SQLConf.ADAPTIVE_EXECUTION_ENABLED.key, "true") | |||
| .set(SQLConf.LOCAL_SHUFFLE_READER_ENABLED.key, "true") | |||
| .set(SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key, "400") | |||
| .set(SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key, "50") | |||
There was a problem hiding this comment.
This change is required since shuffle size changed.
| @@ -99,7 +99,7 @@ class GpuLoreSuite extends SparkQueryCompareTestSuite with FunSuiteWithTempDir w | |||
| } | |||
|
|
|||
| test("AQE broadcast") { | |||
| doTestReplay("90[*]") { spark => | |||
| doTestReplay("93[*]") { spark => | |||
There was a problem hiding this comment.
Same as above, shuffle size change leads to plan change.
| @@ -2052,7 +2052,7 @@ val SHUFFLE_COMPRESSION_LZ4_CHUNK_SIZE = conf("spark.rapids.shuffle.compression. | |||
| .internal() | |||
| .startupOnly() | |||
| .booleanConf | |||
| .createWithDefault(false) | |||
| .createWithDefault(true) | |||
There was a problem hiding this comment.
For spark.rapids.shuffle.kudo.serializer.measure.buffer.copy.enabled, should we enable that by default?
There was a problem hiding this comment.
The changes themselves look fine. I mostly want to see the performance numbers to show that it is at least as good as the old code. I know we have done some of that in the past and that there have been a lot of optimizations recently so it should be good. But this is a big change so I want to see it.
Closes #12202 .
Enable kudo serializer by default, and contains several fix due to shuffle size change.