Tools

Small scripts and prototypes for energy measurement

MSR configurations

turbo_boost.sh

This script can turn the TurboBoost feature in Intel CPUs on or off.

According to Brendan Greggs MSR-Tools Repo there might be some registers to be called when running on different Intel CPUs.

We could not test it on all the possbile, therefore we can only guarantee functionality for Intel Core i7 with Broadwell architecture.

If our script will not work check out the correct register with Brendan Greggs script.

RAPL energy filtering

The RAPL accuracy can be be lowered by either having SGX turned on OR by activating an option in Intel CPUs called "energy filtering".

Source: https://www.intel.com/content/www/us/en/developer/articles/technical/software-security-guidance/advisory-guidance/running-average-power-limit-energy-reporting.html#ipu-2020.2

What actually happens is that the energy information returned by RAPL is randomly modified and will be 0-50% different from the original signal.

This is enough to still have a broad idea of what the total energy budget is but prevents energy side-channel attacks.

We include a script that can check if the energy filtering switch has been set:

sudo check_energy_filtering_rapl.sh

Test SGX

The other possible option when energy filtering kicks in is if SGX is activated in the BIOS. You don't have to have the SGX libraries installed ... just the BIOS switch suffices.

To test wether SGX is activeuse the following script:

gcc test-sgx.c && ./a.out

The relevat values to look at are:

sgx available: 0
sgx launch control: 0

ONLY if both of them are 1, then SGX is active.

Credits: This beautiful script has been copied from @ayeks! Be sure to have a look at the repo!

cpu_relax

This script will trigger same logic as the cpu_relax() function in the linux kernel does.

Originally designed for spin-locks this function can set the processor cycling on the lowest power instruction without loosing control of it (C-State, scheduling etc.)

DVFS

This script will change the cpu frequency govenor in the linux system.

Currently it will just switch to schedutil / performance or powersave.

The possible govenors you can find in: /sys/devices/system/cpu/cpuX/cpufreq/scaling_available_governors

Check if the frequency was really changed by: watch -d -n 1 "cat /proc/cpuinfo | grep MHz"

Read CPU sysbench

In order for this script to work paranoid level should be set to 0:

sudo sysctl -w kernel.perf_event_paranoid=0

If you run multiple perf_events instances it may happen that the counters get multiplexed. A warning will be issued in this case.

The script was designed to find out the deviation between time sharding and instruction sharding when the system background load or is idling.

Currently we see the instruction share strongly deviationg from the time share of a process when the system is loaded.

Further research has to be done to conclude if one or the other metric is more helpful when trying to attribte cumulative metrics like energy to a process.

nop / pause

The nop instruction is the typical "do nothing" instruction in the x86 instruction set.

However Intel introduced also the pause instruction which should be used preferrably when having the CPU in a spin-lock state. This instruction will hint the processor that a spin-lock is happening and can prevent memory violations and in turn save energy.

In order to create this code you should compile the empty_loop.c with: gcc -S empty_loop.c option. This will create an assembler file and you have to inject the nop/pause statements at the correct line.

We have included sample *.s files to showcase that. However they might not compile directly on your system and should just serve as an example.

Once you have compiled your own *.s file create the executable via: gcc nop.s

On our test machine the energy difference is the following:

• sudo perf stat -a -e power/energy-pkg/ --timeout 10000 sleep 20
  • 21.80 Joules
• sudo perf stat -a -e power/energy-pkg/ ./nop
  • 87.37 Joules
  • 89.07 Joules
  • 86.97 Joules
• sudo perf stat -a -e power/energy-pkg/ --timeout 10000 ./pause
  • 88.05 Joules
  • 89.79 Joules
  • 88.51 Joules

Benchmarking an unknown CPU with the nop script

If you want to find out if the CPU speed advertised in the /proc/cpuinfo is really accurate you can just run the nop script and time it.

By dividing the runtime through the estimated instructions per cycle (either 3 or 4 on modern CPUs) you can find out if the CPU has the expected cycle count.

The core should get fully loaded and burst to its maximum frequency as the script contains no sleeps.