Back in 2003, I published the results of experimenting with various
values of _spin_count to improve throughput on latch congested systems.
You can find the original paper here.
The
study involved varying spin count on an Oracle 8.1.6 database suffering
from heavy cache buffer chains latch contention. The database was
running on a Windows 2000 server. The results of this study confirmed
that _spin_count can be very effective in improving the throughput of
latch contention-bound systems. In the test, throughput almost doubled when
spin_count was raised from the default of 2000 to 10000.
I
recently repeated these tests for Oracle 11g. This time, the database
was experiencing shared pool and library cache latch contention and was
running on RHEL 5.0. The results will be published soon in my contention series at ToadWorld.
To summarize however, increasing spin_count in 11g was equally
effective in reducing latch contention. As _spin_count increased,
latch waits reduced and throughput increased, up until CPU saturated,
at which point no further improvements were achieved.