Github user eyalfa commented on a diff in the pull request:
https://github.com/apache/spark/pull/18855#discussion_r131758308
--- Diff: core/src/main/scala/org/apache/spark/storage/DiskStore.scala ---
@@ -165,6 +147,62 @@ private[spark] class DiskStore(
}
+private class DiskBlockData(
+ conf: SparkConf,
+ file: File,
+ blockSize: Long) extends BlockData {
+
+ private val minMemoryMapBytes =
conf.getSizeAsBytes("spark.storage.memoryMapThreshold", "2m")
+
+ override def toInputStream(): InputStream = new FileInputStream(file)
+
+ /**
+ * Returns a Netty-friendly wrapper for the block's data.
+ *
+ * Please see `ManagedBuffer.convertToNetty()` for more details.
+ */
+ override def toNetty(): AnyRef = new DefaultFileRegion(file, 0, size)
+
+ override def toChunkedByteBuffer(allocator: (Int) => ByteBuffer):
ChunkedByteBuffer = {
+ Utils.tryWithResource(open()) { channel =>
+ var remaining = blockSize
+ val chunks = new ListBuffer[ByteBuffer]()
+ while (remaining > 0) {
+ val chunkSize = math.min(remaining, Int.MaxValue)
+ val chunk = allocator(chunkSize.toInt)
+ remaining -= chunkSize
+ JavaUtils.readFully(channel, chunk)
+ chunk.flip()
+ chunks += chunk
+ }
+ new ChunkedByteBuffer(chunks.toArray)
+ }
+ }
+
+ override def toByteBuffer(): ByteBuffer = {
--- End diff --
@cloud-fan
it took me roughly 4 hours, but I looked both at the shuffle cod path and
at `BlockManager.getRemoteBytes`:
it seems the first is robust to large blocks by using Netty's stream
capabilities,
the later seems to be broken as it's not using the Netty's streaming
capabilities and actually tries to copy the result buffer into a heap based
buffer. I think this deserves its own JIRA/PR.
I think these two places plus the external shuffle server cover most of the
relevant use cases (aside from local caching which i believe this PR completes
in terms of being 2GB proof).
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
