_serial_direct_read will use direct path reads for full table scans & fast full index scans only and not the "random" single block reads done for index-based access. You can make your buffer cache really-really small and run your test then. Or use SLOB or Swingbench (again with really small buffer cache). Chris, so have you checked whether the plan_hash_values (and additionally the predicate sections of the plans) on each node are exactly the same - and the *amount* of IOs *per execution* these statements are the same on each node? (and if yes - what are the average IO wait event durations on different nodes?) -- *Tanel Poder* Enkitec (The Exadata Experts) Training <http://blog.tanelpoder.com/seminar/> | Troubleshooting<http://blog.tanelpoder.com/> | Exadata<http://www.amazon.com/Expert-Oracle-Exadata-Apress/dp/1430233923> | Voicee App <http://voic.ee/> On Mon, Apr 22, 2013 at 11:29 PM, Alex Fatkulin <afatkulin@xxxxxxxxx> wrote: > You need "_serial_direct_read"=always > On Mon, Apr 22, 2013 at 4:23 PM, <Christopher.Taylor2@xxxxxxxxxxxx> wrote: > > > I tried setting: > > "_serial_direct_read"=TRUE and that does not do it. > > > > -- > Alex Fatkulin, > http://afatkulin.blogspot.com > > Enkitec, > http://www.enkitec.com > > -- //www.freelists.org/webpage/oracle-l