Contents | Previous | Next

Comparison of CMS1 to FS8F

This section compares the new CMS1 workload to the FS8F0R workload it was derived from. These two workloads are described in CMS-Intensive (CMS1) and CMS-Intensive (FS8F). CMS1 has been set up so that it can be run with the TPNS driver running on a separate system (external TPNS) or with TPNS running in the measured system (internal TPNS) so both cases are shown in the comparison. FS8F0R is the all-minidisks (no SFS) variation of FS8F.

Methodology:  

The comparison measurements were obtained on a 9121-480 processor configured with 256M real storage and no expanded storage, running VM/ESA 2.4.0. Default MDC tuning was used.

Each measurement was done with the number of users adjusted so as to result in an average processor utilization of about 90%. Because CMS1 uses much more processor time per user, the CMS1 cases require a much lower number of users than FS8F. This would have resulted in the CMS1 cases running with zero paging while the FS8F case runs with significant paging. To avoid having this distort the comparison, we locked sufficient pages in the CMS1 cases so that the number of available pages per user was similar for all three measurements.

The CMS1 with internal TPNS case was run without collection of TPNS log data. This is to reflect how we normally run this case (it can also be run with TPNS logging enabled). Because of this, all of the TPNS-related measures (marked (T)) are not available for this case except for ETR (T), which is available from the 1-minute interval TPNS messages.

Results:  

The results are summarized in Table 1 (absolute results) and Table 2 (results relative to the FS8F measurement).

Some measures, such as think time, are intrinsic to the workload or are closely related to it. Other measures, such as response time, depend upon many factors (such as processor speed, I/O configuration, and degree of loading) in addition to the workload itself. In the results tables, an asterisk (*) denotes measures that primarily result from the workload itself ¹ and that are therefore especially useful for characterizing the differences between these workloads.

From the results tables, we can draw the following conclusions regarding how CMS1 with external TPNS compares to FS8F:

• Think time is lower.

• Processor requirements per command are higher because of CPU usage by the TPNS virtual machine.

• The above two factors mean that the number of users required for a given processor configuration is much lower.

• The TV ratio is lower, meaning that there is less CP processing and more emulation (virtual machine) processing.

• I/Os per command are higher. However, this is because each command does more processing. I/Os per CPU-second consumed by the system are similar. Because total system CPU utilization is about the same for all three measurements, this can be seen by comparing VIO RATE.

Most of these differences also apply to CMS1 with internal TPNS. However CMS1 with internal TPNS is different from CMS1 with external TPNS in some ways. All these differences result from the fact that TPNS is running in the measured system. The workload scripts run exactly the same in either case.

• Processor requirements per command are higher because of the TPNS virtual machine(s).

• I/Os per command or per second are lower because no I/Os are required to communicate with the system running TPNS. Instead of I/Os, this communication is between virtual machines using IUCV, which results in a large increase in PRIVOP/CMD.

AVG THINK (T) was not available for the internal TPNS measurement because TPNS log records were not collected. Other measurements with TPNS logging enabled confirm that think time is the same as for CMS1 with external TPNS.

Table 1. FS8F to CMS1 Comparison

Table 2. FS8F to CMS1 Comparison - Ratios

¹

For some of the marked measures, this is true only for the comparison shown. For example, NONPAGE RIO/CMD (V) would change substantially if one run used minidisk caching while another run did not. For these measurements, however, the same MDC tuning was used for all three runs and the MDC hit ratio was similar for all 3 cases.

Contents | Previous | Next