Keep buffer cache question
Date: Thu, 14 Mar 2019 21:22:05 +0000
Message-ID: <0D8F4CAC0F9D3C4AACC63F50FD9957F762D7D2B8_at_PRDTXWPEMLMB32.prod-am.ameritrade.com>
Odd issue here. One of our apps reported slightly increased latency on a heavily used database which started Monday evening and has been consistent since. This is not large from a database perspective but the increase of average response time from 1 to 3 microseconds has had a noticeable impact on their performance.
No execution plan changes, a slightly heavier load at peak times (up about 10% from last week) but nothing that I would expect to have such an impact. We still see the increased latency when the server is 90% idle and the load average is 5 (32 cpus, 16 cores).
In doing an AWR report comparison for comparable times one major difference I saw was that 2 frequently run queries were suddenly doing a lot of physical i/o. For a comparable 2 hour period they went from 1.5 million to 1.8 million executions but physical reads increased from 0 to 1.2 million. I sampled a few other random times and this was consistent. The queries are both doing index access. One is an index range scan and the other a unique scan against the primary key.
I checked with the app group and they have no explanation for why the app might suddenly be querying blocks that aren't in cache whereas they weren't last week.
I am reluctantly considering adding the tables to the keep cache even though they are huge (4-5G_at_). They are frequently accessed so I don't see any real downside to this other than the huge amount of memory I would be reserving for them.
We are going to test this in a non-production environment tomorrow but my two questions are:
- Is anyone aware of any problems this might cause? We have not used the keep cache in the past.
- Is there a way to preload the data once the db_keep_cache_size has been set, the sga increased to accommodate it, and the table assigned to the keep cache? A full table scan would usually bypass the buffer cache so I'm not sure of the best approach here.
- Any other ideas on what might cause this?
TIA!
Jay Miller
Sr. Oracle DBA
-- http://www.freelists.org/webpage/oracle-lReceived on Thu Mar 14 2019 - 22:22:05 CET