Looking at the 2 collection plugins in collectd for Host and VM (from the
hypervisor) CPU usage - measure the time it takes to collectd a CPU stat by instrumenting the CPU and the virt plugin with RTDSC around the piece of code that collects the stat then calculate the time it takes based on the CPU frequency. you can report the rtdsc cycle count as another metric from the plugin.

We follow the methodology here:
https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf
to instrument the read function around the calls to measure the CPU usage
(using CPUID, RDTSC and RDTSCP). We report out the value We collect as another
metric. We will also calculate the measurement overhead of the measurement
calls themselves and take that away from the cycle count values we are
reporting.

• Disable C states, P states and Turbo...
• Run a 1 min test and a 15min test. Measure the cost every 1 seconds
• Based on the reported cycle count calculate the time it takes to complete a
  reading
• Run on an idle system
• Run on a system under load (70-80% Utilized as a max)
• Run for 1 VM, and again for 5VMs
• Run on a platform that doesn't have hyper threading enabled.
• Run on a guest as well as a host