[ 
https://issues.apache.org/jira/browse/HBASE-17343?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15983227#comment-15983227
 ] 

Anoop Sam John commented on HBASE-17343:
----------------------------------------

Did some PE perf testing with different value size (1K which is def in PE, 500 
bytes, 100 bytes)..   
When value size is lesser, the gain from Compaction is more..  Even on bigger 
sized cells case, there is no significant degrade.
Also I did test along with off heap memstore feature.. The gain is more when 
off heap also in place. (This is what my thinking from day one and it comes 
correct as per the tests)..  Off heap + Compacting Memstore comes as the best 
performance.
All tests use G1GC. I also did test with global memstore size as 42% as well as 
60%.  In case of 42%, the Initial Heap Occupancy Percentage (IHOP) is 50% where 
as in next case it is 65%.  In the second case, some times I got long GC pause 
so that master thinks RS is died. 1 out of 3 times I get this.  In case of 50% 
IHOP, this seems not coming.  Remember in all these tests MSLAB is ON (which is 
the default behave).
I also did test with MSLAB off.   Seems this case is better performing than 
MSLAB ON cases.  Specially when the cell size is smaller. In those , I could 
not reproduce the long GC pause.

I have a code level comment, (will say abt that in another comment) and a fix 
there avoided the long GC issue in any case.. 

MSLAB ON vs OFF I tested specifically and seems when we have higher global 
memstore size % and IHOP, the impact is more (MSLAB ON is lesser performing). 
When the IHOP is set to be 50% (This defaults to 45%) , MSLAB ON is having bit 
better performance (Though not too much of diff).    With G1GC, we have to 
check whether we should advice to turn off MSLAB..  Any way off heap memstore 
is implemented using off heap MSLAB pool.. So as such it is useful only. As I 
said, off heap memstore is better performing than on heap counter part and in 
memory compaction there makes things even better.

Did not make a perf graphs.. Will be coming up with that soon (in a day or 2)

> Make Compacting Memstore default in 2.0 with BASIC as the default type
> ----------------------------------------------------------------------
>
>                 Key: HBASE-17343
>                 URL: https://issues.apache.org/jira/browse/HBASE-17343
>             Project: HBase
>          Issue Type: Bug
>    Affects Versions: 2.0.0
>            Reporter: ramkrishna.s.vasudevan
>            Priority: Blocker
>             Fix For: 2.0.0
>
>         Attachments: HBASE-17343-V01.patch, HBASE-17343-V02.patch, 
> HBASE-17343-V04.patch
>
>
> FYI [~anastas], [~eshcar] and [~ebortnik].



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Reply via email to