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-17343) Make Compacting Memstore default in 2.0 with BASIC as the default type
Date Wed, 26 Apr 2017 09:48:05 GMT

    [ https://issues.apache.org/jira/browse/HBASE-17343?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15984480#comment-15984480
] 

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

On the MSLAB, let us do some more tests.. Will discuss that in another issue.  I believe ur
tests were with MSLAB on but no pool.  With pool things are not that bad.  Some cases it is
bit better than no pool only..  But with G1,things do not show very useful (like CMS)
On the long GC pause issue which I saw with compaction.
This is coming in a test case where I have 42 GB of xmx and 60% as Global memstore size. 
I have MSLAB and pool ON.  The Initial Heap Occupancy Percentage is 65%.  The cell value size
is 100 bytes with each row having 10 cols.  In case of compacting memstore, I can see long
GC pause and it even exceeds 30 secs so that HM thinks RS is down and will get YouAreDeadException
and RS gets killed!    MSLAB / pool is good or not is another topic. But IMO in any case,
we should not land in such issues.. So this is some thing we should investigate. 
{code}
public boolean flatten(MemstoreSize memstoreSize) {
    if (isFlat()) return false;
    CellSet oldCellSet = getCellSet();
    int numOfCells = getCellsCount();

    // build the new (CellSet CellArrayMap based)
    CellSet  newCellSet = recreateCellArrayMapSet(numOfCells);
    type = Type.ARRAY_MAP_BASED;
    setCellSet(oldCellSet,newCellSet);

    // arrange the meta-data size, decrease all meta-data sizes related to SkipList
    // (recreateCellArrayMapSet doesn't take the care for the sizes)
    long newSegmentSizeDelta = -(numOfCells * ClassSize.CONCURRENT_SKIPLISTMAP_ENTRY);
    // add size of CellArrayMap and meta-data overhead per Cell
    newSegmentSizeDelta = newSegmentSizeDelta + numOfCells * ClassSize.CELL_ARRAY_MAP_ENTRY;
    incSize(0, newSegmentSizeDelta);
    if (memstoreSize != null) {
      memstoreSize.incMemstoreSize(0, newSegmentSizeDelta);
    }

    return true;
  }
{code}
we have the CSLM and we create the CellArray also  and then finally change the size (This
changes the global RS level size also).  We reduce the CSLM entry sizes and increase CellArray
entry sizes. Finally this will come as reduction in size only..  But temp the size is going
up.   I just did a change here like 1st I change the size with increasing (numOfCells * ClassSize.CELL_ARRAY_MAP_ENTRY)
and then only create CellArrayMap and then reduce as in this position.  Testing with that
patch on the same settings, I did never get this YouAreDeadException and RS killed.  

Also when the tests are with 42% as global MS size, and IHOP as 50%, I did not see RS killed
issue any time..  But the other case, it is very frequent (once in 3 tries)


> 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,
HBASE-17343-V05.patch
>
>
> FYI [~anastas], [~eshcar] and [~ebortnik].



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

Mime
View raw message