Performance analysis of the HP Serviceguard Storage Management Suite for Oracle Database

14
Test analysis summary
The results demonstrate clearly that CFS scales as well as Raw SLVM does in a RAC environment if the
database is partitioned appropriately, i.e. partitioning of the I/O write intensive tables.
Let’s first consider the one node RAC test scenario. There is no CFS or database interconnect network
traffic and the storage does not cause contention on the I/O server side. Therefore the difference in
the TPMs rate can only be explained by the added overhead of the CFS/ODM stack.
On average the performance impact of ODM was relatively small: 8 to 15% drop in TPMs depending
on the workload. Obviously the higher the intensity of the workload the more stressed the I/O
subsystem is. When we increase the number of RAC nodes that TPM drop is subdued: 7 to 10% with
2 instances and 5 to 8% with 3 instances. The main reason is that I/Os are now spread across
multiple systems and Oracle could have passed the requested block via the cache fusion avoiding
some I/Os.
Overall there are no big surprises in the results, which validates our methodology approach.
Graphically when you look at the data we choose to evaluate (TPMs, CPU, I/O, db file sequential
read, unicast network traffic and CR block receive time), you see some degree of parallelism between
CFS and SLVM Raw with Raw consistently performing slightly better.
Recommendation
During this test campaign we showed that the HP Serviceguard Storage Management Suite is indeed
very reliable software-wise and performance is predictable. This is an excellent solution for customers
looking for improved manageability if they can live with a relatively small performance impact. By
properly sizing systems, storage and networking the impact should be minimal and acceptable.