Have host spill use the new HostAlloc API #9257
There are no files selected for viewing
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
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
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
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
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -22,7 +22,7 @@ import java.util.concurrent.ConcurrentHashMap | |
|
|
||
| import scala.collection.mutable | ||
|
|
||
| import ai.rapids.cudf.{Cuda, DeviceMemoryBuffer, HostColumnVector, HostMemoryBuffer, JCudfSerialization, MemoryBuffer, NvtxColor, NvtxRange} | ||
| import com.nvidia.spark.rapids.Arm.{closeOnExcept, freeOnExcept, withResource} | ||
| import com.nvidia.spark.rapids.SpillPriorities.{applyPriorityOffset, HOST_MEMORY_BUFFER_SPILL_OFFSET} | ||
| import com.nvidia.spark.rapids.StorageTier.StorageTier | ||
|
|
@@ -39,26 +39,12 @@ import org.apache.spark.sql.vectorized.ColumnarBatch | |
| * @param pageableMemoryPoolSize maximum size in bytes for the internal pageable memory pool | ||
| */ | ||
| class RapidsHostMemoryStore( | ||
| maxSize: Option[Long]) | ||
| extends RapidsBufferStore(StorageTier.HOST) { | ||
|
|
||
| override protected def spillableOnAdd: Boolean = false | ||
|
|
||
| override def getMaxSize: Option[Long] = maxSize | ||
|
|
||
| def addBuffer( | ||
| id: RapidsBufferId, | ||
|
|
@@ -102,21 +88,23 @@ class RapidsHostMemoryStore( | |
| buffer: RapidsBuffer, | ||
| catalog: RapidsBufferCatalog, | ||
| stream: Cuda.Stream): Boolean = { | ||
| maxSize.forall { ms => | ||
| // this spillStore has a maximum size requirement (host only). We need to spill from it | ||
| // in order to make room for `buffer`. | ||
| val targetTotalSize = ms - buffer.memoryUsedBytes | ||
| if (targetTotalSize <= 0) { | ||
jlowe marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| // lets not spill to host when the buffer we are about | ||
| // to spill is larger than our limit | ||
| false | ||
| } else { | ||
| val amountSpilled = synchronousSpill(targetTotalSize, catalog, stream) | ||
| if (amountSpilled != 0) { | ||
| logDebug(s"Spilled $amountSpilled bytes from ${name} to make room for ${buffer.id}") | ||
| TrampolineUtil.incTaskMetricsDiskBytesSpilled(amountSpilled) | ||
| } | ||
| // if after spill we can fit the new buffer, return true | ||
| buffer.memoryUsedBytes <= (ms - currentSize) | ||
| } | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -125,53 +113,58 @@ class RapidsHostMemoryStore( | |
| catalog: RapidsBufferCatalog, | ||
| stream: Cuda.Stream): Option[RapidsBufferBase] = { | ||
| val wouldFit = trySpillToMaximumSize(other, catalog, stream) | ||
| if (!wouldFit) { | ||
|
There was a problem hiding this comment. @revans2 this will change the behavior in 23.10. Before this, we would still to host. Now we are going to skip to disk if we can't fit given the limit. Just making sure that was intended. There was a problem hiding this comment. I personally think that is what we want, but if anyone has a different opinion I am all ears. We want to get to the point where we have a hard limit on host memory. The reason we don't want to make the changes piecemeal is to reduce the pain a customer might see when running a job. This change is only going to show up when a customer wants to spill from GPU memory, and it is larger than the spill store size. I think that is fairly rare, so I am willing to take the hit. There was a problem hiding this comment. By default the host store is configured to There was a problem hiding this comment. That said, with such a small spill store, we are bound to spill to disk very often anyway... so I don't know if we are saving too much. |
||
| // skip host | ||
| logWarning(s"Buffer $other with size ${other.memoryUsedBytes} does not fit " + | ||
| s"in the host store, skipping tier.") | ||
| None | ||
| } else { | ||
| withResource(other.getCopyIterator) { otherBufferIterator => | ||
| val isChunked = otherBufferIterator.isChunked | ||
| val totalCopySize = otherBufferIterator.getTotalCopySize | ||
| closeOnExcept(HostAlloc.allocHighPriority(totalCopySize)) { hb => | ||
| hb.map { hostBuffer => | ||
| withResource(new NvtxRange("spill to host", NvtxColor.BLUE)) { _ => | ||
| var hostOffset = 0L | ||
| val start = System.nanoTime() | ||
| while (otherBufferIterator.hasNext) { | ||
| val otherBuffer = otherBufferIterator.next() | ||
| withResource(otherBuffer) { _ => | ||
| otherBuffer match { | ||
| case devBuffer: DeviceMemoryBuffer => | ||
| hostBuffer.copyFromMemoryBufferAsync( | ||
| hostOffset, devBuffer, 0, otherBuffer.getLength, stream) | ||
| hostOffset += otherBuffer.getLength | ||
| case _ => | ||
| throw new IllegalStateException("copying from buffer without device memory") | ||
| } | ||
| } | ||
| } | ||
| stream.sync() | ||
| val end = System.nanoTime() | ||
| val szMB = (totalCopySize.toDouble / 1024.0 / 1024.0).toLong | ||
| val bw = (szMB.toDouble / ((end - start).toDouble / 1000000000.0)).toLong | ||
| logDebug(s"Spill to host (chunked=$isChunked) " + | ||
| s"size=$szMB MiB bandwidth=$bw MiB/sec") | ||
| } | ||
| new RapidsHostMemoryBuffer( | ||
| other.id, | ||
| totalCopySize, | ||
| other.meta, | ||
| applyPriorityOffset(other.getSpillPriority, HOST_MEMORY_BUFFER_SPILL_OFFSET), | ||
| hostBuffer) | ||
| }.orElse { | ||
| // skip host | ||
| logWarning(s"Buffer $other with size ${other.memoryUsedBytes} does not fit " + | ||
| s"in the host store, skipping tier.") | ||
| None | ||
| } | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| def numBytesFree: Option[Long] = maxSize.map(_ - currentSize) | ||
|
|
||
| class RapidsHostMemoryBuffer( | ||
| id: RapidsBufferId, | ||
|
|
||
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
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
Oops, something went wrong.
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.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nice catch