Oleg Kibirev created CASSANDRA-4726:
---------------------------------------
Summary: Streaming sstables during repair is limited by compression
Key: CASSANDRA-4726
URL: https://issues.apache.org/jira/browse/CASSANDRA-4726
Project: Cassandra
Issue Type: Bug
Components: Core
Affects Versions: 1.1.5
Reporter: Oleg Kibirev
In our benchmarks of node rebuilds, streaming SSTables is limited by sender
uncompressing them using snappy and compressing using lzf. The receiver has to
reverse this process and is likewise bottlenecked in compression code. We are
therefore only able to utilize less than 1/2 of receiver's 1gbps network link
in a test with one receiver and two senders.
If sstable is already compressed using snappy, it should be just sent to
receiver as is.
Here are hprof fragments excluding socket accepts and reads by idle threads:
Sender:
4 3.14% 95.51% 6128 302816 com.ning.compress.lzf.ChunkEncoder.tryCompress
5 1.93% 97.44% 3761 302815 java.net.SocketOutputStream.socketWrite0
6 1.27% 98.71% 2477 301463 java.util.zip.CRC32.updateBytes
7 0.41% 99.12% 798 302822 com.ning.compress.lzf.ChunkEncoder.tryCompress
8 0.31% 99.43% 602 301465 org.xerial.snappy.SnappyNative.rawUncompress
9 0.18% 99.60% 349 302820 java.net.SocketInputStream.socketRead0
10 0.14% 99.75% 282 302817 java.net.SocketOutputStream.socketWrite0
11 0.14% 99.89% 276 301468 java.io.FileInputStream.readBytes
12 0.03% 99.92% 52 301929 sun.nio.ch.FileDispatcher.preClose0
Receiver:
4 7.15% 84.63% 2782 300893 org.xerial.snappy.SnappyNative.rawCompress
5 3.16% 87.79% 1232 300895 java.util.zip.CRC32.updateBytes
6 2.83% 90.62% 1101 300896
com.ning.compress.lzf.LZFDecoder.decompressChunk
7 1.50% 92.12% 583 300916 java.io.DataInputStream.readFully
8 1.46% 93.57% 568 300921 java.io.FileDescriptor.sync
9 1.16% 94.73% 451 300910 java.io.RandomAccessFile.writeBytes
10 0.47% 95.20% 184 300949 sun.nio.ch.NativeThread.current
11 0.45% 95.66% 177 300926 java.io.RandomAccessFile.getFilePointer
12 0.31% 95.97% 121 300935
org.apache.cassandra.io.util.SequentialWriter.write
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
