[
https://issues.apache.org/jira/browse/RATIS-1717?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17615963#comment-17615963
]
Song Ziyang edited comment on RATIS-1717 at 10/11/22 4:45 PM:
--------------------------------------------------------------
[~szetszwo] Oops, we {*}cannot directly serialize LogEntryProto to
directByteBuffer due to checksum{*}. Checksum class in JDK8 cannot calculate
checksum on byte sequence sitting in direct memory. Therefore, a copy to heap
byte array seems inevitable, unless we dare to *use Unsafe operations* to
checksum on direct memory :).
Still, we can allocate a global buffer (4M) instead of allocating a temp array
every write. Simulation results show that this did reduce young GC. I've
submitted https://issues.apache.org/jira/secure/attachment/13050789/idea.patch,
please take a look. Also, Should we try Unsafe operations?
was (Author: JIRAUSER281912):
[~szetszwo] Oops, we {*}cannot directly serialize LogEntryProto to
directByteBuffer due to checksum{*}. Checksum class in JDK8 cannot calculate
checksum on byte sequence sitting in direct memory. Therefore, a copy to heap
byte array seems inevitable, unless we dare to *use Unsafe operations* to
checksum on direct memory :).
Still, we can allocate a global buffer (4M) instead of allocating a temp array
every write. Simulation results show that this did reduce young GC. I've
submitted an [idea patch]([^idea.patch)]). Please take a look. Also, Should we
try Unsafe operations?
> Reduce data copy when serialize LogEntry to RaftLog
> ---------------------------------------------------
>
> Key: RATIS-1717
> URL: https://issues.apache.org/jira/browse/RATIS-1717
> Project: Ratis
> Issue Type: Improvement
> Components: performance
> Reporter: Song Ziyang
> Priority: Major
> Attachments: idea.patch
>
>
> In current implementation, before a log entry is serialized to RaftLog, it
> first needs to be serialized to a byte array[1], and then be written to
> FileChannel using directByteBuffer[2]. It requires two data copies, one from
> LogEntryProto to byte array, one from byte array to direct ByteBuffer.
> We can reduce the data copy by *serializing LogEntry to directByteBuffer,
> removing the byte array:*
> # We need a directByteBuffer large enough to hold a single log
> (maxBufferSize).
> # When we need to serialize a log entry to RaftLog, first we check whether
> there's enough space remained in directByteBuffer. If not, we flush the
> directByteBuffer to FileChannel. Otherwise we serialize the data directly
> into directByteBuffer.
> # When data buffered in directByteBuffer reaches the threshold (say 64KB),
> we flush the directByteBuffer, the same as current implementation.
> [1]
> [https://github.com/apache/ratis/blob/289db3ae64e8cf620eba882468dda661af0439bc/ratis-server/src/main/java/org/apache/ratis/server/raftlog/segmented/SegmentedRaftLogOutputStream.java#L96]
> [2]
> [https://github.com/apache/ratis/blob/289db3ae64e8cf620eba882468dda661af0439bc/ratis-server/src/main/java/org/apache/ratis/server/raftlog/segmented/BufferedWriteChannel.java#L65]
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
