|X86_ENERGY_PERF_POLICY(8)||System Manager's Manual||X86_ENERGY_PERF_POLICY(8)|
scope: --cpu cpu-list | --pkg pkg-list
cpu-list, pkg-list: # | #,# | #-# | all
field: --all | --epb | --hwp-epp | --hwp-min | --hwp-max | --hwp-desired
other: (--force | --hwp-enable | --turbo-enable) value)
value: # | default | performance | balance-performance | balance-power | power
Policy in MSR_IA32_ENERGY_PERF_BIAS (EPB) may affect a wide range of hardware decisions, such as how aggressively the hardware enters and exits CPU idle states (C-states) and Processor Performance States (P-states). This policy hint does not replace explicit OS C-state and P-state selection. Rather, it tells the hardware how aggressively to implement those selections. Further, it allows the OS to influence energy/performance trade-offs where there is no software interface, such as in the opportunistic "turbo-mode" P-state range. Note that MSR_IA32_ENERGY_PERF_BIAS is defined per CPU, but some implementations share a single MSR among all CPUs in each processor package. On those systems, a write to EPB on one processor will be visible, and will have an effect, on all CPUs in the same processor package.
Hardware P-States (HWP) are effectively an expansion of hardware P-state control from the opportunistic turbo-mode P-state range to include the entire range of available P-states. On Broadwell Xeon, the initial HWP implementation, EPB influenced HWP. That influence was removed in subsequent generations, where it was moved to the Energy_Performance_Preference (EPP) field in a pair of dedicated MSRs -- MSR_IA32_HWP_REQUEST and MSR_IA32_HWP_REQUEST_PKG.
EPP is the most commonly managed knob in HWP mode, but MSR_IA32_HWP_REQUEST also allows the user to specify minimum-frequency for Quality-of-Service, and maximum-frequency for power-capping. MSR_IA32_HWP_REQUEST is defined per-CPU.
MSR_IA32_HWP_REQUEST_PKG has the same capability as MSR_IA32_HWP_REQUEST, but it can simultaneously set the default policy for all CPUs within a package. A bit in per-CPU MSR_IA32_HWP_REQUEST indicates whether it is over-ruled-by or exempt-from MSR_IA32_HWP_REQUEST_PKG.
MSR_HWP_CAPABILITIES shows the default values for the fields in MSR_IA32_HWP_REQUEST. It is displayed when no values are being written.
-p, --pkg Operate on the MSR_IA32_HWP_REQUEST_PKG for each package in the package-list. The list is a string of individual package numbers separated by commas, and or ranges of package numbers separated by a dash, or the string "all". For example '--pkg 1,3' or '--pkg all'
performance Set a policy for maximum performance, accepting no performance sacrifice for the benefit of energy efficiency.
balance-performance Set a policy with a high priority on performance, but allowing some performance loss to benefit energy efficiency.
balance-power Set a policy where the performance and power are balanced. This is the default.
power Set a policy where the processor can accept a measurable performance impact to maximize energy efficiency.
The following table shows the mapping from the value strings above to actual MSR values. This mapping is defined in the Linux-kernel header, msr-index.h.
VALUE STRING EPB EPP performance 0 0 balance-performance 4 128 normal, default 6 128 balance-power 8 192 power 15 255
For MSR_IA32_HWP_REQUEST performance fields (--hwp-min, --hwp-max, --hwp-desired), the value option is in units of 100 MHz, Eg. 12 signifies 1200 MHz.
-B, --epb set EPB per-core or per-package. See value strings in the table above.
-d, --debug debug increases verbosity. By default x86_energy_perf_policy is silent for updates, and verbose for read-only mode.
-P, --hwp-epp set HWP.EPP per-core or per-package. See value strings in the table above.
-m, --hwp-min request HWP to not go below the specified core/bus ratio. The "default" is the value found in IA32_HWP_CAPABILITIES.min.
-M, --hwp-max request HWP not exceed a the specified core/bus ratio. The "default" is the value found in IA32_HWP_CAPABILITIES.max.
-D, --hwp-desired request HWP 'desired' frequency. The "normal" setting is 0, which corresponds to 'full autonomous' HWP control. Non-zero performance values request a specific performance level on this processor, specified in multiples of 100 MHz.
-w, --hwp-window specify integer number of microsec in the sliding window that HWP uses to maintain average frequency. This parameter is meaningful only when the "desired" field above is non-zero. Default is 0, allowing the HW to choose.
-U, --hwp-use-pkg (0 | 1), when used in conjunction with --cpu, indicates whether the per-CPU MSR_IA32_HWP_REQUEST should be overruled (1) or exempt (0) from per-Package MSR_IA32_HWP_REQUEST_PKG settings. The default is exempt.
-H, --hwp-enable enable HardWare-P-state (HWP) mode. Once enabled, system RESET is required to disable HWP mode.
-t, --turbo-enable enable (1) or disable (0) turbo mode.
-v, --version print version and exit.
If no request to change policy is made, the default behavior is to read and display the current system state, including the default capabilities.
Note that --hwp-desired and --hwp-window are considered experimental. Future versions of Linux reserve the right to access these fields internally -- potentially conflicting with user-space access.
# sudo x86_energy_perf_policy cpu0: EPB 6 cpu0: HWP_REQ: min 6 max 35 des 0 epp 128 window 0x0 (0*10^0us) use_pkg 0 cpu0: HWP_CAP: low 1 eff 8 guar 27 high 35 cpu1: EPB 6 cpu1: HWP_REQ: min 6 max 35 des 0 epp 128 window 0x0 (0*10^0us) use_pkg 0 cpu1: HWP_CAP: low 1 eff 8 guar 27 high 35 cpu2: EPB 6 cpu2: HWP_REQ: min 6 max 35 des 0 epp 128 window 0x0 (0*10^0us) use_pkg 0 cpu2: HWP_CAP: low 1 eff 8 guar 27 high 35 cpu3: EPB 6 cpu3: HWP_REQ: min 6 max 35 des 0 epp 128 window 0x0 (0*10^0us) use_pkg 0 cpu3: HWP_CAP: low 1 eff 8 guar 27 high 35
msr(4) Intel(R) 64 and IA-32 Architectures Software Developer's Manual