Mainframe Migration - 7
degree of interactions with others in the mix (increased contention for resources) would also require a larger SEP:CPU ratio since the SEPs are likely to be blocked at sometime during the transaction execution.
Scalability Results and Analysis
Note that the TPC benchmark rules prohibit publicly disclosing TPC performance figures that have not been independently audited. Therefore, we are required to withhold from this paper any data that may be used to derive our TPC metrics. However, we can disclose relative benchmark results from the different CPU configurations to convey the scalability of the system. Tests were performed by varying the simulated user count (see following table) or the number of CPUs in the configuration (see subsequent table) and observing its effect on cumulative CPU utilization. As with many TPC-C tests 90th percentile response time requirements (under 5 seconds) led to the termination of the test before CPU was exhausted. Increases in response time were attributed to increased disk queue length and client customer networks.
Transaction Rate Ratio
Response Time (90th percentile)
The above table shows the effect of increasing workload in terms of concurrent users while maintaining a constant CPU configuration. As expected, the increased rate increased the transaction response time and CPU utilization consistent with the increased load.
Changes in the processor and disk subsystem hardware led to an increased MIPS rating. The earlier paper estimated 1347 mainframe MIPS equivalent for online transaction processing workloads. The 8-CPU system used in these tests equated to 1712 MIPS. The same versions of Windows, Enterprise Server with the Mainframe Transaction Option and IBM UDB were used here and in tests cited in the earlier paper. The hardware changes also allow significant increases in the number of concurrent users the system could sustain. In the earlier paper’s test the transaction response time requirements in the TPC-C spec limited the overall user count to 3600 for 1347 MIPS. Lower 90th percentile response times were noted in these tests allowing the number of concurrent users to be increased to a high of 5160 (1712 MIPS).
A second round of tests concerned scalability of the alternative mainframe environment. We varied the CPU count (see following table) while keeping the number of concurrent users constant (except for the smallest configuration which could not sustain the higher count).
WinHEC 2005 Version - April 20, 2005