[CELEBORN-1921] Broadcast large GetReducerFileGroupResponse to prevent Spark driver network exhausted #3158
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
turboFei
changed the title
turboFei
force-pushed
the
broadcast_rgf
branch
4 times, most recently
from
e2455f6 to
2379189
Compare
turboFei
force-pushed
the
broadcast_rgf
branch
2 times, most recently
from
0f00cde to
ba003e4
Compare
turboFei
changed the title
…t Spark driver network exhausted
turboFei
commented
Mar 20, 2025
| .doc("The size at which we use Broadcast to send the GetReducerFileGroupResponse to the executors.") | ||
| .version("0.6.0") | ||
| .bytesConf(ByteUnit.BYTE) | ||
| .createWithDefaultString("512k") |
There was a problem hiding this comment.
same default value with https://github.com/apache/spark/blob/8d260084b8a50ff59a127c7292c0cdb6737981b0/core/src/main/scala/org/apache/spark/internal/config/package.scala#L1718
private[spark] val SHUFFLE_MAPOUTPUT_MIN_SIZE_FOR_BROADCAST =
ConfigBuilder("spark.shuffle.mapOutput.minSizeForBroadcast")
.doc("The size at which we use Broadcast to send the map output statuses to the executors.")
.version("2.0.0")
.bytesConf(ByteUnit.BYTE)
.createWithDefaultString("512k")
turboFei
commented
Mar 20, 2025
| } | ||
| return null; | ||
| }); | ||
| } |
There was a problem hiding this comment.
turboFei
commented
Mar 20, 2025
| releaseLock(key) | ||
| } | ||
| } | ||
| } |
There was a problem hiding this comment.
turboFei
commented
Mar 20, 2025
| @@ -77,6 +77,7 @@ | |||
| <include>io.netty:*</include> | |||
| <include>org.apache.commons:commons-lang3</include> | |||
| <include>org.roaringbitmap:RoaringBitmap</include> | |||
| <include>commons-io:commons-io</include> | |||
There was a problem hiding this comment.
To fix the class not found:
rn.ApplicationMaster$$anon$2.run(ApplicationMaster.scala:732)
Caused by: java.lang.NoClassDefFoundError: org/apache/celeborn/shaded/org/apache/commons/io/output/ByteArrayOutputStream
at org.apache.spark.shuffle.celeborn.SparkShuffleManager.<init>(SparkShuffleManager.java:113)
... 20 more
Caused by: java.lang.ClassNotFoundException: org.apache.celeborn.shaded.org.apache.commons.io.output.ByteArrayOutputStream
at java.net.URLClassLoader.findClass(URLClassLoader.java:387)
at java.lang.ClassLoader.loadClass(ClassLoader.java:419)
at sun.misc.Launcher$AppClassLoader.loadClass(Launcher.java:365)
at java.lang.ClassLoader.loadClass(ClassLoader.java:352)
... 21 more
jar -tf client-spark/spark-3-shaded/target/celeborn-client-spark-3-shaded_2.12-0.6.0-SNAPSHOT.jar|grep org/apache/celeborn/shaded/org/apache/commons/io/output/ByteArrayOutputStream
org/apache/celeborn/shaded/org/apache/commons/io/output/ByteArrayOutputStream.class
There was a problem hiding this comment.
The org.apache.commons package is relocated.
celeborn/client-spark/spark-3-shaded/pom.xml
Lines 63 to 64 in 9e8f9f6
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
For spark celeborn application, if the GetReducerFileGroupResponse is larger than the threshold, Spark driver would broadcast the GetReducerFileGroupResponse to the executors, it prevents the driver from being the bottleneck in sending out multiple copies of the GetReducerFileGroupResponse (one per executor).
Why are the changes needed?
To prevent the driver from being the bottleneck in sending out multiple copies of the GetReducerFileGroupResponse (one per executor).
Does this PR introduce any user-facing change?
No, the feature is not enabled by defaults.
How was this patch tested?
UT.
Cluster testing with
spark.celeborn.client.spark.shuffle.getReducerFileGroup.broadcast.enabled=true.The broadcast response size should be always about 1kb.
Application succeed.