One of the most common Exadata performance problems I see is that the direct path reads (and thus also Smart Scans) don’t sometimes kick in when running full scans in serial sessions. This is because in Oracle 11g onwards, the serial full segment scan IO path decision is done dynamically, at runtime, for every SQL execution – and for every segment (partition) separately. Whether you get a direct path read & smart scan, depends on the current buffer cache size, how big segment you’re about to scan and how much of that segment is actually cached at the moment. Note that the automatic IO path decision for index fast full scans is slightly different from table scans.
This dynamic decision unfortunately can cause unexpected surprises and variance in your report/batch job runtimes. Additionally, it looks like the SELECT part of your UPDATE/DELETE statements (the select part finds the rows to update/delete) does not ever automatically get direct path read/smart scan chosen – by design! So, when your SELECT statement may use smart scan and be really fast, the same select operation in an INSERT SELECT (or UPDATE/DELETE) context will not end up using smart scans by default. There’s even a bug explaining that – closed as “not a bug” (Exadata Smartscan Is Not Being Used On Insert As Select[Article ID 1348116.1]).
To work around these problems and force a direct path read/smart scan, you can either:
- Run your query in parallel as parallel full segment scans will use direct path reads, unless your parallel_degree_policy = AUTO, then you may still get buffered reads thanks to the dynamic in-memory parallel execution decision of Oracle 11.2
- Run your query in serial, but force the serial direct path reads by setting _serial_direct_read = TRUE (or ALWAYS in 188.8.131.52+)
Here are the valid options for this parameter in 184.108.40.206+