marin-ma commented on issue #11542: URL: https://github.com/apache/incubator-gluten/issues/11542#issuecomment-3840522300
> I'm guessing that spill works like this, is that correct? cc [@FelixYBW](https://github.com/FelixYBW) [@marin-ma](https://github.com/marin-ma) > > ``` > VeloxSortShuffleWriter::reclaimFixedSize --> VeloxSortShuffleWriter::evictAllPartitions --> VeloxSortShuffleWriter::evictPartition --> RssPartitionWriter::sortEvict > ^ | > | require memory to compress data > | | > ---------------------------------------- cannot recalim memory from shuffle writer again -------------------------------------------- > ``` > > We should allocate global memory instead of offheap memory during spill. [@marin-ma](https://github.com/marin-ma) . Did we ? > > [incubator-gluten/cpp/core/shuffle/rss/RssPartitionWriter.cc](https://github.com/apache/incubator-gluten/blob/1a05b03aee73335d41953b3472d79094dc844f69/cpp/core/shuffle/rss/RssPartitionWriter.cc#L72) > > Line 72 in [1a05b03](/apache/incubator-gluten/commit/1a05b03aee73335d41953b3472d79094dc844f69) > > codec_.get(), options_->compressionBufferSize, rssOs_, arrow::default_memory_pool())); > Currently, that should be the case, so is it impossible for us to avoid OOM kill by yarn? @FelixYBW As @wForget posted the data compressing already uses the `arrow::default_memory_pool()`. It shouldn't borrow memory here. -- 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]
