As discussed previously, in UNIX, the parameters
controlling shared memory usage are the limiting factor. In any case, before you
create the database, serious thought has to be given to how much you expect the
SGA to grow over the next year. Over specifying the shared memory parameters on
a UNIX platform will not do any harm and may save you and the system
administrator some headaches.The size of the SGA is
controlled by buffer sizes, and the buffer sizes are controlled by the database
block size, which is specified at database creation and cannot be changed
without rebuilding the database in Oracle8i but can vary on Oracle9i. This
usually defaults to 8 KB. I usually suggest at least 8 KB, although in some
cases 16 KB works best. If in doubt, set the block size to the largest supported
on your system.
The five major components of the SGA are the database
buffers, log buffers, LARGE POOL, Java pool, and the shared pool. The SGA also
contains the redo log buffers. The ideal situation would be to size the SGA to
hold the entire database in memory. For small systems, this may be a real
situation; for most, it is not feasible. However, with the new 64-bit
architectures 16 exabytes of storage can be directly addressed; only a few
exabytes can hold all of the world's printed, videotaped, and recorded data, so
in the future entire databases will reside easily in memory. Therefore, you must
decide how much to allocate.
In many cases, especially for development databases,
this will be a rough SWAG (scientific wild-assed guess). For systems already
designed with detailed data storage estimates, it may be better defined. A
general rule of thumb for a pure Oracle system (no other applications) is 50 to
60 percent of available RAM for your SGA. Note that for small databases this may
be overkill. In general, I have found that sizing the SGA data block buffers
(the product of DB_BLOCK_SIZE and DB_BLOCK_BUFFERS) to 1/50 to 1/100 of the
total physical size of the database is a good starting point. Under Oracle8i and
Oracle9i, the new default sizes in the supplied sample initialization file are
more realistic, but in general will still be too small for most production
databases, so use them with caution.
Oracle provides tools to analyze buffer performance.
Unfortunately, they can only be used once a system is operating and running
under a normal load; so for our discussion of installation, they are useless.
If you have no idea whatsoever, make the buffer area at
least 60 to 100 MB or so (you will usually outgrow the Oracle default rather
quickly) for a database that is near 1 gigabyte in total physical size, and up
to 400 MB for one that is around 20 gigabytes in size. For databases smaller
than 1 gigabyte physical size, the Oracle defaults may be usable. Make the
shared pool at least 50 to 100 MB.
We will discuss the actual parameters in the INIT.ORA
file that control SGA size when we get to the section on tuning. What you need
to know right now is that the default initialization file provided by Oracle has
three default ranges: way-too-small, too-small, and small. Unless you are
creating a database that will be less than 1 gigabyte in physical size, even the
large parameters in the example initialization file are woefully inadequate.
One thing to remember: If you over specify the shared
memory size on UNIX, you may get into a situation known as swapping. This is
where all or part of your application is swapped out to disk because physical
memory just isn't large enough to hold it all. Needless to say, this has a very
negative impact on performance. Usually, over specification of the SGA on UNIX
will lead to not being able to start the database.
and include the URL for the page.
