hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Anoop Sam John (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-15788) Use Offheap ByteBuffers from BufferPool to read RPC requests.
Date Wed, 09 Nov 2016 08:36:58 GMT

    [ https://issues.apache.org/jira/browse/HBASE-15788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15650192#comment-15650192
] 

Anoop Sam John commented on HBASE-15788:
----------------------------------------

bq.is an optimization that has nought to do w/ ByteBuffer writing?
Yes.  This is just another optimization. Instead of writing an int in 4 calls to an OS, it
can be done in one step.  So in OS like BAOS, BBOS, the checkSizeAndGrow step can be done
only once instead of 4 times.  Just few saving in #instructions executed.  Also we use Unsafe
way to write to BB/BA.  So instead of 4 calls one call using Unsafe.  As int write ops are
more (for every cell once) thought it would be better optimization.  I agree it wont fit correctly
in the context of the interface BBWriter.  What u say? We can get rid of that?  I can use
StreamUtils API to write int then.  Byte by byte it will write.

> Use Offheap ByteBuffers from BufferPool to read RPC requests.
> -------------------------------------------------------------
>
>                 Key: HBASE-15788
>                 URL: https://issues.apache.org/jira/browse/HBASE-15788
>             Project: HBase
>          Issue Type: Sub-task
>          Components: regionserver
>            Reporter: ramkrishna.s.vasudevan
>            Assignee: Anoop Sam John
>             Fix For: 2.0.0
>
>         Attachments: HBASE-15788.patch, HBASE-15788_V4.patch, HBASE-15788_V5.patch, HBASE-15788_V6.patch
>
>
> Right now, when an RPC request reaches RpcServer, we read the request into an on demand
created byte[]. When it is write request and including many mutations, the request size will
be some what larger and we end up creating many temp on heap byte[] and causing more GCs.
> We have a ByteBufferPool of fixed sized off heap BBs. This is used at RpcServer while
sending read response only. We can make use of the same while reading reqs also. Instead of
reading whole of the request bytes into a single BB, we can read into N BBs (based on the
req size). When BB is not available from pool, we will fall back to old way of on demand on
heap byte[] creation.
> Remember these are off heap BBs. We read many proto objects from this read request bytes
(like header, Mutation protos etc). Thanks to PB 3 and our shading work as it supports off
heap BB now.  Also the payload cells are also in these DBBs now. The codec decoder can work
on these and create off heap BBs. Whole of our write path work with Cells now. At the time
of addition to memstore, these cells are by default copied to MSLAB ( off heap based pooled
MSLAB issue to follow this one). If MSLAB copy is not possible, we will do a copy to on heap
byte[].
> One possible down side of this is :
> Before adding to Memstore, we do write to WAL. So the Cells created out of the offheap
BBs (Codec#Decoder) will be used to write to WAL. The default FSHLog works with an OS obtained
from DFSClient. This will have only standard OS write APIs which is byte[] based.  So just
to write to WAL, we will end up in temp on heap copy for each of the Cell. The other WAL imp
(ie. AsynWAL) supports writing offheap Cells directly. We have work in progress to make AsycnWAL
as default. Also we can raise HDFS req to support BB based write APIs in their client OS?
Until then, will try for a temp workaround solution. Patch to say more on this.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message