The following examples use the hardware guidelines and Table 7-3 to provide illustrate how to use parallelism to meet performance goals:
The number of partitions for a table should be less than or equal to the number of devices. For the experiment showing scaling of engines and worker processes shown in Table 7-3, there were 30 devices available, so 30 partitions were used. Performance is optimal when each partition is placed on a separate physical device.
Determine the number of partitions based on the I/O throughput you want to achieve. If you know your disks and controllers can sustain 1MB per second per device, and you want a table scan on an 800MB table to complete in 30 seconds, you need to achieve approximately 27MB per second total throughput, so you would need at least 27 devices with one partition per device, and at least 27 worker processes, one for each partition. These figures are very close to the I/O rates in the example in Table 7-3.
Estimate the number of CPUs, based on the number of partitions, and then determine the optimum number by tracking both CPU utilization and I/O saturation. The example shown in Table 7-3 had 30 partitions available. Following the suggestions in the hardware guidelines of one CPU for each five devices suggests using six engines for CPU-intensive queries. At that level, I/O was not saturated, so adding more engines improved response time.
