[ https://issues.apache.org/jira/browse/HBASE-17623?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15925468#comment-15925468 ]
Chia-Ping Tsai commented on HBASE-17623: ---------------------------------------- ||statistics||master||HBASE-17623|| |elapsed(s)|5599|5501| |young GC count|2026|1778| |young total GC time(s)|338|273| |old GC count|495|415| |old total GC time(s)|679|540| |total allocation(GB)|1950|704| |total pause time(s)|369|296| # NO compaction # NO wal # NO split # 1 TB input # G1 # snappy # v2 patch I ran the tests 16 times. The more details are shown in the attachment. The patch avoids the extra copy if we disable the cache-on-write. The result showed that this patch reduces the memory allocation, GC, and runtime. Would you please take a look? [~anoop.hbase] > Reuse the bytes array when building the hfile block > --------------------------------------------------- > > Key: HBASE-17623 > URL: https://issues.apache.org/jira/browse/HBASE-17623 > Project: HBase > Issue Type: Improvement > Reporter: Chia-Ping Tsai > Assignee: Chia-Ping Tsai > Priority: Minor > Fix For: 2.0.0, 1.4.0 > > Attachments: after(snappy_hfilesize=5.04GB).png, > after(snappy_hfilesize=755MB).png, before(snappy_hfilesize=5.04GB).png, > before(snappy_hfilesize=755MB).png, HBASE-17623.branch-1.v0.patch, > HBASE-17623.branch-1.v1.patch, HBASE-17623.branch-1.v2.patch, > HBASE-17623.branch-1.v2.patch, HBASE-17623.v0.patch, HBASE-17623.v1.patch, > HBASE-17623.v1.patch, HBASE-17623.v2.patch, memory allocation measurement.xlsx > > > There are two improvements. > # The onDiskBlockBytesWithHeader should maintain a bytes array which can be > reused when building the hfile. > # The onDiskBlockBytesWithHeader is copied to an new bytes array only when we > need to cache the block. > # If no block need to be cached, the uncompressedBlockBytesWithHeader will > never be created. > {code:title=HFileBlock.java|borderStyle=solid} > private void finishBlock() throws IOException { > if (blockType == BlockType.DATA) { > this.dataBlockEncoder.endBlockEncoding(dataBlockEncodingCtx, > userDataStream, > baosInMemory.getBuffer(), blockType); > blockType = dataBlockEncodingCtx.getBlockType(); > } > userDataStream.flush(); > // This does an array copy, so it is safe to cache this byte array when > cache-on-write. > // Header is still the empty, 'dummy' header that is yet to be filled > out. > uncompressedBlockBytesWithHeader = baosInMemory.toByteArray(); > prevOffset = prevOffsetByType[blockType.getId()]; > // We need to set state before we can package the block up for > cache-on-write. In a way, the > // block is ready, but not yet encoded or compressed. > state = State.BLOCK_READY; > if (blockType == BlockType.DATA || blockType == BlockType.ENCODED_DATA) > { > onDiskBlockBytesWithHeader = dataBlockEncodingCtx. > compressAndEncrypt(uncompressedBlockBytesWithHeader); > } else { > onDiskBlockBytesWithHeader = defaultBlockEncodingCtx. > compressAndEncrypt(uncompressedBlockBytesWithHeader); > } > // Calculate how many bytes we need for checksum on the tail of the > block. > int numBytes = (int) ChecksumUtil.numBytes( > onDiskBlockBytesWithHeader.length, > fileContext.getBytesPerChecksum()); > // Put the header for the on disk bytes; header currently is > unfilled-out > putHeader(onDiskBlockBytesWithHeader, 0, > onDiskBlockBytesWithHeader.length + numBytes, > uncompressedBlockBytesWithHeader.length, > onDiskBlockBytesWithHeader.length); > // Set the header for the uncompressed bytes (for cache-on-write) -- > IFF different from > // onDiskBlockBytesWithHeader array. > if (onDiskBlockBytesWithHeader != uncompressedBlockBytesWithHeader) { > putHeader(uncompressedBlockBytesWithHeader, 0, > onDiskBlockBytesWithHeader.length + numBytes, > uncompressedBlockBytesWithHeader.length, > onDiskBlockBytesWithHeader.length); > } > if (onDiskChecksum.length != numBytes) { > onDiskChecksum = new byte[numBytes]; > } > ChecksumUtil.generateChecksums( > onDiskBlockBytesWithHeader, 0, onDiskBlockBytesWithHeader.length, > onDiskChecksum, 0, fileContext.getChecksumType(), > fileContext.getBytesPerChecksum()); > }{code} -- This message was sent by Atlassian JIRA (v6.3.15#6346)