[ https://issues.apache.org/jira/browse/HADOOP-8148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13257987#comment-13257987 ]
Todd Lipcon commented on HADOOP-8148: ------------------------------------- Duplicating my comment from HADOOP-8258: {quote} In current versions of Hadoop, the read path for applications like HBase often looks like: allocate a byte array for an HFile block (~64kb) call read() into that byte array: copy 1: read() packets from the socket into a direct buffer provided by the DirectBufferPool copy 2: copy from the direct buffer pool into the provided byte[] call setInput on a decompressor copy 3: copy from the byte[] back to a direct buffer inside the codec implementation call decompress: JNI code accesses the input buffer and writes to the output buffer copy 4: from the output buffer back into the byte[] for the uncompressed hfile block ineffiency: HBase now does its own checksumming. Since it has to checksum the byte[], it can't easily use the SSE-enabled checksum path. Given the new direct-buffer read support introduced by HDFS-2834, we can remove copy #2 and #3 allocate a DirectBuffer for the compressed hfile block, and one for the uncompressed block (we know the size from the hfile block header) call read() into the direct buffer using the HDFS-2834 API copy 1: read() packets from the socket into that buffer call setInput() with that buffer. no copies necessary call decompress: JNI code accesses the input buffer and writes directly to the output buffer, with no copies HBase now has the uncompressed block as a direct buffer. It can use the SSE-enabled checksum for better efficiency This should improve the performance of HBase significantly. We may also be able to use the new API from within SequenceFile and other compressible file formats to avoid two copies from the read path. Similar applies to the write path, but in my experience the write path is less often CPU-constrained, so I'd prefer to concentrate on the read path first. {quote} > Zero-copy ByteBuffer-based compressor / decompressor API > -------------------------------------------------------- > > Key: HADOOP-8148 > URL: https://issues.apache.org/jira/browse/HADOOP-8148 > Project: Hadoop Common > Issue Type: New Feature > Components: io > Reporter: Tim Broberg > Assignee: Tim Broberg > Attachments: hadoop8148.patch > > > Per Todd Lipcon's comment in HDFS-2834, " > Whenever a native decompression codec is being used, ... we generally have > the following copies: > 1) Socket -> DirectByteBuffer (in SocketChannel implementation) > 2) DirectByteBuffer -> byte[] (in SocketInputStream) > 3) byte[] -> Native buffer (set up for decompression) > 4*) decompression to a different native buffer (not really a copy - > decompression necessarily rewrites) > 5) native buffer -> byte[] > with the proposed improvement we can hopefully eliminate #2,#3 for all > applications, and #2,#3,and #5 for libhdfs. > " > The interfaces in the attached patch attempt to address: > A - Compression and decompression based on ByteBuffers (HDFS-2834) > B - Zero-copy compression and decompression (HDFS-3051) > C - Provide the caller a way to know how the max space required to hold > compressed output. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira