apache / datafusion-comet

Apache DataFusion Comet Spark Accelerator
https://datafusion.apache.org/comet
Apache License 2.0
825 stars 164 forks source link

CometShuffleMemoryAllocator should use Spark unified memory manager #1064

Closed viirya closed 1 week ago

viirya commented 2 weeks ago

What is the problem the feature request solves?

Currently Comet JVM shuffle's CometShuffleMemoryAllocator uses a separate unsafe memory allocator and track off-heap memory allocation internally. The memory size is configured by a Comet config.

A better approach should be to use Spark unified memory manager for CometShuffleMemoryAllocator. So we can unify this off-heap memory with Spark off-heap memory config, i.e., we don't need to separately configure Comet JVM shuffle memory.

Describe the potential solution

No response

Additional context

No response