Memory plugging and steal time

The cpuplugd tests with memory plugging revealed a serious problem when a kernel compile was used as workload.

The result is shown in Figure 1.
Figure 1. CMM Pools size and CPU steal time over time when compiling a Linux™ kernel
CMM Pools size and CPU steal time over time when compiling a Linux kernel

Observation

During the increase of the CMM pool the steal time frequently reaches values between 90% and 100%. The issue appears on z/VM® 5.4 and on z/VM 6.1 when the APAR described below is not installed.

Conclusion

It is not recommended using cpuplugd to manage the memory size of a guest without the APAR installed, described in the next section.

Figure 2 shows the behavior after installing the fix released in APAR VM65060.
Figure 2. CMM Pools size and CPU steal time over time when compiling a Linux kernel with the fix released in APAR VM65060 installed
Graph shows the behavior after installing the fix released in APAR VM65060.

Observation

After installing the fix the steal time is between 1% and 4% for most of the time instead of 90%-100% without the fix, see Figure 1. The graph also shows that the number of assigned CPUs increases during the compile phase and decreases afterwards during the link phase, in line with the processing power required.

Conclusion

Installing the z/VM APAR VM65060 fix is required when memory management with cpuplugd is planned to avoid excessive steal time numbers. It is available for z/VM 5.4, z/VM 6.1, and z/VM 6.2.