'enq: TX - index contention' Waits in a RAC Environment.

When we run OLTP systems in a Oracle RAC environment, ita€™s possible that we see high TX enqueue contention on index associated with  tables, which are having high concurrency from the application.  This usually happens when the application performs lot of INSERTs and DELETEs  concurrently from all the instances . 

The reason for this is the index block splits while inserting a new row into the index. The transactions will have to wait  for  TX lock in mode 4, until the session that is doing the block splits completes the operations.

A session will initiate a index block split, when it cana€™t find space in an index block where it needs to insert a new row. Before starting the split, it would clean out all the keys in the block to check whether there is enough sufficient space in the block.deleted

Splitter has to do the following activities:

    o          Allocate a new block.
    o          Copy a percentage of rows to the new buffer.
    o          Add the new buffer to the index structure and commit the operation.

In RAC environments, this could be an expensive operation, due to the global cache operations included. The impact will be more if the split is happening at a branch or root block level.

Most probable reasons are:

   o          Indexes on the tables which are being accessed heavily from the application.
   o          Indexes on table columns which are having values inserted by a monotonically increasing.

The indexes which are having contention can be identified from the AWR reports taken during the time of the issue.

Top 5 Timed Events:

Event                       Waits      Time(s)   Avg Wait(ms)  % Total Call  Time Wait Class 
en: TX - index contention   89,350     40,991    459           63.3          Concurrency 
db file sequential read     1,458,288  12,562    9             19.4          User I/O 
CPU time                               5,352                   8.3   

Instance Activity Stats:

Statistic                Total     per Second    per Trans 
branch node splits       945       0.26          0.00 
leaf node 90-10 splits   1,670     0.46          0.00 
leaf node splits         35,603    9.85          0.05                          

And the objects can be found either from V$SEGMENT_STATISTICS or from the 'Segments by Row Lock Waits' of the AWR reports.

Segments by Row Lock Waits:

Owner     Tablespace  Object Name            Obj.Type   Row Lock Waits  % of Capture 
ACSSPROD  ACSS_IDX03  ACSS_ORDER_HEADER_PK   INDEX      3,425           43.62 
ACSSPROD  ACSS_IDX03  ACSS_ORDER_HEADER_ST   INDEX      883             11.25 
ACSSPROD  ACSS_IDX03  ACSS_ORDER_HEADER_DT   INDEX      682             8.69 

 

Solution here is to tune the indexes avoid heavy access on a few set of blocks. 

Following are the options we could try:

o        Rebuild the as reverse key indexes or hash partition the indexes which are listed in the 'Segments by Row Lock Waits' of the AWR reports

From the Performance Tuning Guide - 

Reverse key indexes are designed to eliminate index hot spots on insert applications.  These indexes are excellent for insert performance.  But the downside of it is that, it may affect the performance of index range scans. 

The hash method can improve performance of indexes where a small number leaf blocks in the index have high contention in multiuser OLTP environment. In some OLTP applications, index insertions happen only at the right edge of the index. This could happen when the index is defined on monotonically increasing  columns. In such situations right edge of the index becomes a hotspot because of contention for index pages, buffers, latches for update, and additional index  maintenance activity, which results in performance degradation.

Its recommended to test the application performance,  after rebuilding the indexes as  reverse key or hash partitioned.

o        Consider increasing the CACHE size of the sequences.

When we use monotonically increasing sequences for populating column values, the leaf block which is having high sequence key will be changing with every  insert, which makes it a hot block and potential candidate for a block split. 

With CACHE SIZE (and probably with NOORDER option), each instance would use start using the sequence keys with a different range reduces the index keys getting insert same set of leaf blocks.