3.7.0 HP StorageWorks HP Scalable NAS File Serving Software provisioning guide for Oracle HP Scalable NAS (AG513-96013, October 2009)

top five wait events in your STATSPACK reports, simply schedule a database reboot
and increase the number of writers.
NFS client network performance considerations
When configuring storage bandwidth for Oracle databases, there is no simple formula
that meets the needs of all workloads. However, there are some guidelines. For
Online Transaction Processing (OLTP), a commonly accepted ratio of database server
to I/O subsystem bandwidth is 100Mb I/O for every 1GHz CPU. While this is a
very rough formula, it does stand the test of time and field experience and is
applicable to DAS, NAS and SAN storage topologies. Consider that with OLTP, the
Oracle I/O profile consists largely of random single block transfers (e.g., db file
sequential read, DBWR writes), occasional multiblock reads (e.g., small table scans,
index range scans) and small sequential writes to the online redo log. Since 100Mb
I/O throughput is approximately 12MB/s, using a 4KB blocksize yields an
approximate maximum I/O rate of 3072 I/O operations per second (IOps). Given
these figures, the formula works out to be fairly accurate. Indeed, try to find a server
with a single 1GHz processor that drives Oracle OLTP at a rate of 3072 random
4KB transfers per second. Generally speaking, a single CPU clocked at 1GHz would
surely bottleneck long before it could drive I/O up to 3072 IOps due to transactional
overhead and activity on cached data.
A more modern, realistic example would be a server configured with two dual-core
processors each clocked at, say, 2.4GHz. In this case there would be some 9.6GHz
CPU bandwidth for the formula and thus 9600 Mb I/O requirement. With a little
rounding up, the formula suggests that the two-way, dual core server would require
a single GigE network path for NFS traffic to the cluster. The formula is not perfect,
but it is a good guideline. There is one significant error in the formula having to do
with the fact that there is no such thing as an OLTP-only Oracle deployment. The
same I/O subsystem that performs satisfactorily for OLTP might not be sufficient for
maintenance operations such as backup/restore, bulk data loads, archive log
switches, index creation, ad hoc parallel query and reporting.
The tremendous I/O bandwidth capability of HP Scalable NAS cannot be realized
if the NFS client network interfaces are under-configured. Given the relatively low
cost of network hardware, it is recommended that Oracle servers be configured with
a minimum of one bonded NFS path to each cluster NAS head regardless of the
“OLTP formula.” This will ensure that each Oracle server will be serviced with roughly
200MB/s for such activity as Parallel Query, Backup/Restore, index creation and
so on. Of course, larger Oracle servers may demand more throughput than 200MB/s
and options such as three-way bonded NICs and multiple mount points should be
considered.
HP Scalable NAS File Serving Software provisioning guide for Oracle 37