Hi Chris,

On Wed, May 06, 2020 at 01:23:59PM +0200, Chris Hofer wrote:

Hi,

to measure the performance overhead of the Armv8 Genode VMM, I ran some
benchmarks (cyclictest and lmbench). I compared it against a native Linux
installation on the Imx8-evk board (same kernel version). For some
measurements (for example fork + executing /bin/sh from lmbench), I observe
counter-intuitive results in the sense that the virtualized version runs
*faster* than the native one (Vm 1,1ms vs. native 2,5ms).

My understanding is, that the virtual timer offset (cntvoff_el2) is not used
and thus, the guest sees wall clock time. I would expect that the
virtualization incurs an overhead and this overhead manifests itself in
longer (wall clock) time measurements.
 Correct, but you assume that native Linux and virtualized version run
with the same frequencies (cpu + timer) - don't you?

Actually, I think the timer frequency should be constant in both
scenarios, whereby at least the cpu will be controlled dynamically in
the native Linux case.

When comparing such complex scenarios to avoid mistakes I would
strongly encourage you to not measure time on-target. It often leads
to wrong assumptions. It is better to measure long-term scenarios from
another host. If you need to measure micro-benchmarks, you might loop
the operation quite often and measure the whole time-span outside, and
divide the result.

Alternatively, if you insist to measure on-target, you can instrument
Linux to use a constant CPU frequency (google for "linux cpu
governor"). Moreover, you might check whether the timer frequency is
the same in both settings (CNTFRQ_EL0) - well it should be.