Running Systems

I have had an interesting experience with HPE servers, where the BIOS was defined to allow max performance (in contrast to ‘balanced’ mode), but still – the CPU was not at max all the time.

While we generally, strive to a greener computing, when having low-latency workload, we expect a deterministic performance. We want our database to provide results in a timely, and more important – a consistent way. We want to expect a certain query/job to take the same time, whenever we run it. CPU throttling and frequency manipulation, which are part of CPU power-management are not our friends. A certain query can take longer, if the CPU is currently throttled, and can switch to another, slower, core mid-task. This can result in a very complex performance tuning and server behaviour troubleshooting.

The goal of setting the BIOS to ‘max-performance’ is obvious, then, however – it does not have the desired effect. The CPU keeps on throttling, and, while some power-reducing features are disabled, not all. This is not our goal.

Adding to the boot options of the server (via the GRUB/GRUB2/GRUB-EFI/Whatever) the following parameters, should disable Intel CPU throttling, and all (i)relevant C_states. I have not tested, nor investigated AMD behaviour:

nosoftlockup intel_idle.max_cstate=0 mce=ignore_ce

Of course – a reboot is required for these settings to take effect.

If you want to control power management afterwards, you can manually disable a certain core/CPU by running a command such as this:

echo 0 > /sys/devices/system/cpu/cpu<number>/online

It will disable the CPU (and place it in deep power save) until woken by another CPU, by echoing ‘1’ to that same ‘file’.