System Z Performance
Pegasystems has many years experience with producing high-performance applications with Pega BPM. Pega also regularly test system performance using simulated workloads for different combinations of hardware, applications, and use cases.
For its latest performance test, Pega partnered with IBM to perform a benchmark on System z to answer a critical set of questions:
- Can Pega BPM and CPM be effectively deployed across an enterprise (for over 10,000 users) while maintaining adequate performance levels?
- What are the System z memory and CPU resource requirements to adequately support the ensuing workload?
- How well does Pega BPM scale in the IBM System z environment both horizontally and vertically, and what is the optimal configuration to take advantage of the z architecture and scale?
Pega's aim was to produce linear or near-linear Performance Scalability Curves with respect to the MIPS required and work performed per unit of time within the load range supported by the z10 for Call Center, CPM, and Hybrid CPM-BPM workloads.
Benchmark goals were no more than 3 MIPS per user for a steady state load of 12,000 simulated, concurrent users and nd-to-end browser response time of < 2 seconds.
For each workflow, 5 to 7 test cases were run with increasing numbers of users to demonstrate vertical and horizontal scalability. For each workflow, the system was loaded up to the point of failure in the last test case.
Pega BPM met or exceeded all performance goals and success criteria. It scaled linearly both vertically within a single partition and horizontally across four partitions:
- Call Center scenario with 35,000 users required 0.82 MIPS per user and the end-to-end browser response time was 0.832 seconds.
- CPM scenario with 9,000 users required 3.24 MIPS per user and the end-to-end browser response time was 1.12 seconds.
- Hybrid CPM-BPM scenario with 12,000 users required 3.03 MIPS per user and the end-to-end browser response time was 1.22 seconds against targets of 3 MIPS per user and a 2.0 second browser response time.
Endurance tests for each workload (8-16 hours in length) showed the same performance numbers as the normal (1± hour) duration tests.