CE/PE throughput and CPU load

Edit online

The measurement series with the maximum throughput number for each CE/PE guest CPU configuration was selected for the vertical scalability chart below.

Besides the throughput bars, two trend lines indicate the CPU load for the CE/PE and the database guest belonging to the throughput rates. The bars for the normalized transaction throughput are arranged accordingly to their workload level in the chart.

line and bar graph showing throughput, CE/PE CPU load and DB CPU load — Figure 1. CE/PE vertical scalability - transaction throughput and CPU load

The blue (dark) bars in chart show the maximum transaction throughput for all CPU configurations at increasing workload levels. The red and yellow (light) trend lines show the corresponding CPU load on the CE/PE and the DB (database) guest.

Observations

The maximum throughput rates start from 5 for the 1-1 CPU scenario to nearly 40 for the 8-2 CPU scenario. To drive each CPU scenario with the maximum workload almost all CPUs on the CE/PE guest are used. On the database server the CPU load grows smoothly from 0.1 to 1.5 IFLs used for the highest throughput rate at the 8-2 CPU scenario.

Each doubling of the CPUs on the CE/PE guest allowed us to achieve a nearly doubled workload. The throughput rate increase shows a nearly perfect scalability characteristics for the full workload bandwidth (see also Figure 1). Adding more CPUs to the CE/PE guest allows to drive higher throughput rates when the server is already running CPU bound. The CPU load on the database server should also be monitored, and in our scenarios the database server CPU load grows smoothly and not as fast as on the CE/PE server for the chosen ECM workload.

Line graph showing normalised throughput — Figure 2. CE/PE vertical scalability – CE/PE CPU load for transaction throughput rates

These tests illustrate the good vertical scalability characteristics for FileNet P8 on Linux for System z when increasing the number of CPUs for the scenarios and the workload for the CE/PE guest. The linear scalability range ranges from 1 used IFL through 6 IFLs, demonstrating very nice symmetric multiprocessing (SMP) performance, regardless how much virtual CPUs are configured for the CP/PE guest the used CPU capacity is the same. There is no additional CPU cost, just because there are more CPUs available.