richardstartin commented on a change in pull request #6320:
URL: https://github.com/apache/incubator-pinot/pull/6320#discussion_r536912358
##########
File path:
pinot-core/src/main/java/org/apache/pinot/core/segment/creator/impl/inv/OffHeapBitmapInvertedIndexCreator.java
##########
@@ -181,33 +188,49 @@ public void seal()
}
// Create bitmaps from inverted index buffers and serialize them to file
- try (DataOutputStream offsetDataStream = new DataOutputStream(
- new BufferedOutputStream(new FileOutputStream(_invertedIndexFile)));
- FileOutputStream bitmapFileStream = new
FileOutputStream(_invertedIndexFile);
- DataOutputStream bitmapDataStream = new DataOutputStream(new
BufferedOutputStream(bitmapFileStream))) {
- int bitmapOffset = (_cardinality + 1) * Integer.BYTES;
- offsetDataStream.writeInt(bitmapOffset);
- bitmapFileStream.getChannel().position(bitmapOffset);
-
+ ByteBuffer offsetBuffer = null;
+ ByteBuffer bitmapBuffer = null;
+ try (FileChannel channel = new RandomAccessFile(_invertedIndexFile,
"rw").getChannel()) {
Review comment:
I had a hard time believing it when this came up, but did enough
profiling to figure out what was going on. The difference is _up to_ 20x, and
it depends on the composition of the bitmap: `DataOutput` penalises dense
bitmaps more than others. `DataOutput` has to do things like manually extract
bytes from `int`, `long` etc. and this adds up to a lot compared to a simple
move instruction. Because of endianness, you get away without reversing the
bytes, and with `ByteBuffer` various bounds checks get eliminated. You can see
the difference even with a pre-sized `ByteArrayOutputStream`, so it's not a
question of unbuffered IO.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
