On Thu, 10 Nov 2005 08:33:39 +0000, Jonathan Lewis wrote:
> A basic requirement for RAC is that a highly concurrent
> application has to be designed to ensure separation of
> activity. Being able to select, update, and delete but haning
> on inserts is possible.
In the good, old times of OPS, there used to be a name for it: "functional
partitioning". Now, with the cache fusion, people are told that RAC is a
universal solution, good everywhere where a single instance is good. That
is not quite true. PCM still locks blocks and still has to coordinate two
machines. Communication between machines is still much slower then the
interprocess communication. RAC is a redundancy and survivability option,
not a performance option. Special cases are large data warehouse
configurations, which can utilize highly parallelized queries. Say you
have 8 HP SuperDome nodes, with 64 CPUs each, then RAC can give you
signficant performance advantage.
The freezing phenomenon can be observe from the log files: when one
node crashes, the other nodes must rebuild lock database. While at it,
no blocks can be locked, even in the share mode.
--
http://www.mgogala.com
Received on Sat Nov 12 2005 - 18:41:06 CST
