[
https://issues.apache.org/jira/browse/HBASE-15721?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anoop Sam John updated HBASE-15721:
-----------------------------------
Description:
Before cells added to memstore CSLM, there is a clone of cell after copying it
to MSLAB chunk area. This is done not in an efficient way.
{code}
public static int appendToByteArray(final Cell cell, final byte[] output, final
int offset) {
int pos = offset;
pos = Bytes.putInt(output, pos, keyLength(cell));
pos = Bytes.putInt(output, pos, cell.getValueLength());
pos = appendKeyTo(cell, output, pos);
pos = CellUtil.copyValueTo(cell, output, pos);
if ((cell.getTagsLength() > 0)) {
pos = Bytes.putAsShort(output, pos, cell.getTagsLength());
pos = CellUtil.copyTagTo(cell, output, pos);
}
return pos;
}
{code}
Copied in 9 steps and we end up parsing all lengths. When the cell
implementation is backed by a single byte[] (Like KeyValue) this can be done in
single step copy.
Also avoid lots of garbage while doing this MSLAB copy. With every cell, we
used to create a ByteRange instance to pass the allocated chunk + offset from
MSLAB to Segment. Now we can make the MSLAB impl to do the allocation and copy
of the cell so that one object creation per cell addition can be avoided.
was:
Before cells added to memstore CSLM, there is a clone of cell after copying it
to MSLAB chunk area. This is done not in an efficient way.
{code}
public static int appendToByteArray(final Cell cell, final byte[] output, final
int offset) {
int pos = offset;
pos = Bytes.putInt(output, pos, keyLength(cell));
pos = Bytes.putInt(output, pos, cell.getValueLength());
pos = appendKeyTo(cell, output, pos);
pos = CellUtil.copyValueTo(cell, output, pos);
if ((cell.getTagsLength() > 0)) {
pos = Bytes.putAsShort(output, pos, cell.getTagsLength());
pos = CellUtil.copyTagTo(cell, output, pos);
}
return pos;
}
{code}
Copied in 9 steps and we end up parsing all lengths. When the cell
implementation is backed by a single byte[] (Like KeyValue) this can be done in
single step copy.
> Optimization in cloning cells into MSLAB
> ----------------------------------------
>
> Key: HBASE-15721
> URL: https://issues.apache.org/jira/browse/HBASE-15721
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver
> Reporter: Anoop Sam John
> Assignee: Anoop Sam John
> Fix For: 2.0.0
>
> Attachments: HBASE-15721.patch, HBASE-15721_V2.patch,
> HBASE-15721_V3.patch
>
>
> Before cells added to memstore CSLM, there is a clone of cell after copying
> it to MSLAB chunk area. This is done not in an efficient way.
> {code}
> public static int appendToByteArray(final Cell cell, final byte[] output,
> final int offset) {
> int pos = offset;
> pos = Bytes.putInt(output, pos, keyLength(cell));
> pos = Bytes.putInt(output, pos, cell.getValueLength());
> pos = appendKeyTo(cell, output, pos);
> pos = CellUtil.copyValueTo(cell, output, pos);
> if ((cell.getTagsLength() > 0)) {
> pos = Bytes.putAsShort(output, pos, cell.getTagsLength());
> pos = CellUtil.copyTagTo(cell, output, pos);
> }
> return pos;
> }
> {code}
> Copied in 9 steps and we end up parsing all lengths. When the cell
> implementation is backed by a single byte[] (Like KeyValue) this can be done
> in single step copy.
> Also avoid lots of garbage while doing this MSLAB copy. With every cell, we
> used to create a ByteRange instance to pass the allocated chunk + offset from
> MSLAB to Segment. Now we can make the MSLAB impl to do the allocation and
> copy of the cell so that one object creation per cell addition can be avoided.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
