When running many Linux guests under z/VM you should always enable the On Demand
Timer Patch to help z/VM manage its resources properly.
The On Demand Timer Patch is one of the most important performance
relevant items for Linux on z/Series. For a detailed description
about the impact of the Timer Patch see: Linux on IBM eServer zSeries and S/390: Performance Measurement
and Tuning Chapter 7.3 (there it is called "The Linux timer
patch"). We highly recommend reading this chapter to understand
the performance impact of the On Demand Timer Patch.
Here an extract:
Traditionally, the Linux kernel keeps track of time using a
timer that interrupts the kernel at a constant rate. On each
interrupt, the global variable “jiffies” is incremented
and various queues inspected for work. On dedicated hardware, this
has a minimal performance effect. However, when running Linux
running on z/VM, this is not a good idea.
Although this method of keeping track of time may not be the
most efficient, the biggest problem for Linux on z/VM is that the
Linux guest uses a little bit of CPU cycles every 10 ms. This
causes the z/VM scheduler to keep the guest “in queue”
(and therefore, unused memory pages in the Linux virtual machine
cannot be trimmed, as discussed in 2.3, “z/VM use of
memory” on page 13). With many Linux guests holding on to
their working sets, z/VM memory will fill up rather quickly.
This problem is addressed with the so called “on-demand
timer” patch that can be downloaded from the IBM
developerWorks Linux for zSeries and S/390 home page:
...
This patch does away with the 10 ms timer tick and sets the
timer to pop only when the kernel needs to wake up. Even though
this does not make a real “zero load idle Linux guest,”
the periods between two timer ticks are normally long enough for
z/VM to recognize the guest as idle (and start taking measures to
trim memory pages).
