[
https://issues.apache.org/jira/browse/HBASE-15180?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15126122#comment-15126122
]
Jingcheng Du commented on HBASE-15180:
--------------------------------------
To supplement more information.
+------------------+------------------+-----------------+-----------------+
| |mslab-pool-on | mslab-on | mslab-off |
+------------------+------------------+-----------------+-----------------+
| throughput | 242801 | 240464 | 225679 |
+------------------+------------------+-----------------+-----------------+
| latency(us) | 486 | 483 | 704
|
+------------------+------------------+-----------------+-----------------+
| gc pause(ms) | 8440 | 12837 | 17131 |
+------------------+------------------+-----------------+-----------------+
I used 2MB for the chunk size during the test.
Speaking of G1GC, G1GC splits the heap into 2000 regions. In my test (the heap
size is 64GB), each region will be 32MB.
If the chunk pool is off, there might be fragments after minor gc (2MB is a
fragment comparing to 32MB, and the fragments are compacted in mixed gc and
full gc).
If the chunk size is increased to 16MB or larger, I believe the performance can
be improved comparing to 2MB chunk size.
> Reduce garbage created while reading Cells from Codec Decoder
> -------------------------------------------------------------
>
> Key: HBASE-15180
> URL: https://issues.apache.org/jira/browse/HBASE-15180
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver
> Reporter: Anoop Sam John
> Assignee: Anoop Sam John
> Fix For: 2.0.0
>
> Attachments: HBASE-15180.patch, HBASE-15180_V2.patch
>
>
> In KeyValueDecoder#parseCell (Default Codec decoder) we use
> KeyValueUtil#iscreate to read cells from the InputStream. Here we 1st create
> a byte[] of length 4 and read the cell length and then an array of Cell's
> length and read in cell bytes into it and create a KV.
> Actually in server we read the reqs into a byte[] and CellScanner is created
> on top of a ByteArrayInputStream on top of this. By default in write path, we
> have MSLAB usage ON. So while adding Cells to memstore, we will copy the Cell
> bytes to MSLAB memory chunks (default 2 MB size) and recreate Cells over that
> bytes. So there is no issue if we create Cells over the RPC read byte[]
> directly here in Decoder. No need for 2 byte[] creation and copy for every
> Cell in request.
> My plan is to make a Cell aware ByteArrayInputStream which can read Cells
> directly from it.
> Same Codec path is used in client side also. There better we can avoid this
> direct Cell create and continue to do the copy to smaller byte[]s path. Plan
> to introduce some thing like a CodecContext associated with every Codec
> instance which can say the server/client context.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
