Deferred Enabling Of ACPI CPUFreq Boost Support Can Help Boot Times For Large Servers
Linux kernel contributor and system administrator Stuart Hayes noticed that when the ACPI CPUFreq driver is loaded, CPU boosting support is enabled sequentially by setting the appropriate model specific registers (MSRs) before the driver is even registered with CPUFreq. This ends up being rather time consuming with today's high core count servers with the one-at-a-time per-core CPU boost enabling. It's also possible that the time can be completely wasted if CPUFreq registering of the driver fails.
Stuart sent out a patch instead to defer setting of the boost MSRs with the ACPI CPUFreq driver. He commented, "Since cpufreq already sets up a CPU hotplug startup callback if and when acpi-cpufreq is registered, set the boost MSRs in acpi_cpufreq_cpu_init(), which is called by the cpufreq cpuhp callback. This allows acpi-cpufreq to exit quickly if it is loaded but not needed."
This change is to the generic ACPI CPUFreq driver, which is used by various Arm server processors as well as AMD CPUs unless using the newer AMD P-State driver and the Intel server CPUs are generally on the Intel P-State driver. Hayes noted that the small code patch was able to speed up booting on a test server by about 30 seconds. The server mentioned had 192 CPU cores. Given 192 CPU cores is a bit of an odd combination right now, it's more than likely an AMD EPYC Genoa server that is already public knowledge as maxing out at 96 CPU cores per socket. In any event this change isn't specific to Genoa and will help any large core count systems making use of the ACPI CPUFreq driver.
This boot time saving improvement for large core count systems is currently out for review/comment on the linux-pm mailing list.