1.. SPDX-License-Identifier: GPL-2.0 2.. include:: <isonum.txt> 3 4.. |intel_pstate| replace:: :doc:`intel_pstate <intel_pstate>` 5 6======================= 7CPU Performance Scaling 8======================= 9 10:Copyright: |copy| 2017 Intel Corporation 11 12:Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com> 13 14 15The Concept of CPU Performance Scaling 16====================================== 17 18The majority of modern processors are capable of operating in a number of 19different clock frequency and voltage configurations, often referred to as 20Operating Performance Points or P-states (in ACPI terminology). As a rule, 21the higher the clock frequency and the higher the voltage, the more instructions 22can be retired by the CPU over a unit of time, but also the higher the clock 23frequency and the higher the voltage, the more energy is consumed over a unit of 24time (or the more power is drawn) by the CPU in the given P-state. Therefore 25there is a natural tradeoff between the CPU capacity (the number of instructions 26that can be executed over a unit of time) and the power drawn by the CPU. 27 28In some situations it is desirable or even necessary to run the program as fast 29as possible and then there is no reason to use any P-states different from the 30highest one (i.e. the highest-performance frequency/voltage configuration 31available). In some other cases, however, it may not be necessary to execute 32instructions so quickly and maintaining the highest available CPU capacity for a 33relatively long time without utilizing it entirely may be regarded as wasteful. 34It also may not be physically possible to maintain maximum CPU capacity for too 35long for thermal or power supply capacity reasons or similar. To cover those 36cases, there are hardware interfaces allowing CPUs to be switched between 37different frequency/voltage configurations or (in the ACPI terminology) to be 38put into different P-states. 39 40Typically, they are used along with algorithms to estimate the required CPU 41capacity, so as to decide which P-states to put the CPUs into. Of course, since 42the utilization of the system generally changes over time, that has to be done 43repeatedly on a regular basis. The activity by which this happens is referred 44to as CPU performance scaling or CPU frequency scaling (because it involves 45adjusting the CPU clock frequency). 46 47 48CPU Performance Scaling in Linux 49================================ 50 51The Linux kernel supports CPU performance scaling by means of the ``CPUFreq`` 52(CPU Frequency scaling) subsystem that consists of three layers of code: the 53core, scaling governors and scaling drivers. 54 55The ``CPUFreq`` core provides the common code infrastructure and user space 56interfaces for all platforms that support CPU performance scaling. It defines 57the basic framework in which the other components operate. 58 59Scaling governors implement algorithms to estimate the required CPU capacity. 60As a rule, each governor implements one, possibly parametrized, scaling 61algorithm. 62 63Scaling drivers talk to the hardware. They provide scaling governors with 64information on the available P-states (or P-state ranges in some cases) and 65access platform-specific hardware interfaces to change CPU P-states as requested 66by scaling governors. 67 68In principle, all available scaling governors can be used with every scaling 69driver. That design is based on the observation that the information used by 70performance scaling algorithms for P-state selection can be represented in a 71platform-independent form in the majority of cases, so it should be possible 72to use the same performance scaling algorithm implemented in exactly the same 73way regardless of which scaling driver is used. Consequently, the same set of 74scaling governors should be suitable for every supported platform. 75 76However, that observation may not hold for performance scaling algorithms 77based on information provided by the hardware itself, for example through 78feedback registers, as that information is typically specific to the hardware 79interface it comes from and may not be easily represented in an abstract, 80platform-independent way. For this reason, ``CPUFreq`` allows scaling drivers 81to bypass the governor layer and implement their own performance scaling 82algorithms. That is done by the |intel_pstate| scaling driver. 83 84 85``CPUFreq`` Policy Objects 86========================== 87 88In some cases the hardware interface for P-state control is shared by multiple 89CPUs. That is, for example, the same register (or set of registers) is used to 90control the P-state of multiple CPUs at the same time and writing to it affects 91all of those CPUs simultaneously. 92 93Sets of CPUs sharing hardware P-state control interfaces are represented by 94``CPUFreq`` as struct cpufreq_policy objects. For consistency, 95struct cpufreq_policy is also used when there is only one CPU in the given 96set. 97 98The ``CPUFreq`` core maintains a pointer to a struct cpufreq_policy object for 99every CPU in the system, including CPUs that are currently offline. If multiple 100CPUs share the same hardware P-state control interface, all of the pointers 101corresponding to them point to the same struct cpufreq_policy object. 102 103``CPUFreq`` uses struct cpufreq_policy as its basic data type and the design 104of its user space interface is based on the policy concept. 105 106 107CPU Initialization 108================== 109 110First of all, a scaling driver has to be registered for ``CPUFreq`` to work. 111It is only possible to register one scaling driver at a time, so the scaling 112driver is expected to be able to handle all CPUs in the system. 113 114The scaling driver may be registered before or after CPU registration. If 115CPUs are registered earlier, the driver core invokes the ``CPUFreq`` core to 116take a note of all of the already registered CPUs during the registration of the 117scaling driver. In turn, if any CPUs are registered after the registration of 118the scaling driver, the ``CPUFreq`` core will be invoked to take note of them 119at their registration time. 120 121In any case, the ``CPUFreq`` core is invoked to take note of any logical CPU it 122has not seen so far as soon as it is ready to handle that CPU. [Note that the 123logical CPU may be a physical single-core processor, or a single core in a 124multicore processor, or a hardware thread in a physical processor or processor 125core. In what follows "CPU" always means "logical CPU" unless explicitly stated 126otherwise and the word "processor" is used to refer to the physical part 127possibly including multiple logical CPUs.] 128 129Once invoked, the ``CPUFreq`` core checks if the policy pointer is already set 130for the given CPU and if so, it skips the policy object creation. Otherwise, 131a new policy object is created and initialized, which involves the creation of 132a new policy directory in ``sysfs``, and the policy pointer corresponding to 133the given CPU is set to the new policy object's address in memory. 134 135Next, the scaling driver's ``->init()`` callback is invoked with the policy 136pointer of the new CPU passed to it as the argument. That callback is expected 137to initialize the performance scaling hardware interface for the given CPU (or, 138more precisely, for the set of CPUs sharing the hardware interface it belongs 139to, represented by its policy object) and, if the policy object it has been 140called for is new, to set parameters of the policy, like the minimum and maximum 141frequencies supported by the hardware, the table of available frequencies (if 142the set of supported P-states is not a continuous range), and the mask of CPUs 143that belong to the same policy (including both online and offline CPUs). That 144mask is then used by the core to populate the policy pointers for all of the 145CPUs in it. 146 147The next major initialization step for a new policy object is to attach a 148scaling governor to it (to begin with, that is the default scaling governor 149determined by the kernel command line or configuration, but it may be changed 150later via ``sysfs``). First, a pointer to the new policy object is passed to 151the governor's ``->init()`` callback which is expected to initialize all of the 152data structures necessary to handle the given policy and, possibly, to add 153a governor ``sysfs`` interface to it. Next, the governor is started by 154invoking its ``->start()`` callback. 155 156That callback is expected to register per-CPU utilization update callbacks for 157all of the online CPUs belonging to the given policy with the CPU scheduler. 158The utilization update callbacks will be invoked by the CPU scheduler on 159important events, like task enqueue and dequeue, on every iteration of the 160scheduler tick or generally whenever the CPU utilization may change (from the 161scheduler's perspective). They are expected to carry out computations needed 162to determine the P-state to use for the given policy going forward and to 163invoke the scaling driver to make changes to the hardware in accordance with 164the P-state selection. The scaling driver may be invoked directly from 165scheduler context or asynchronously, via a kernel thread or workqueue, depending 166on the configuration and capabilities of the scaling driver and the governor. 167 168Similar steps are taken for policy objects that are not new, but were "inactive" 169previously, meaning that all of the CPUs belonging to them were offline. The 170only practical difference in that case is that the ``CPUFreq`` core will attempt 171to use the scaling governor previously used with the policy that became 172"inactive" (and is re-initialized now) instead of the default governor. 173 174In turn, if a previously offline CPU is being brought back online, but some 175other CPUs sharing the policy object with it are online already, there is no 176need to re-initialize the policy object at all. In that case, it only is 177necessary to restart the scaling governor so that it can take the new online CPU 178into account. That is achieved by invoking the governor's ``->stop`` and 179``->start()`` callbacks, in this order, for the entire policy. 180 181As mentioned before, the |intel_pstate| scaling driver bypasses the scaling 182governor layer of ``CPUFreq`` and provides its own P-state selection algorithms. 183Consequently, if |intel_pstate| is used, scaling governors are not attached to 184new policy objects. Instead, the driver's ``->setpolicy()`` callback is invoked 185to register per-CPU utilization update callbacks for each policy. These 186callbacks are invoked by the CPU scheduler in the same way as for scaling 187governors, but in the |intel_pstate| case they both determine the P-state to 188use and change the hardware configuration accordingly in one go from scheduler 189context. 190 191The policy objects created during CPU initialization and other data structures 192associated with them are torn down when the scaling driver is unregistered 193(which happens when the kernel module containing it is unloaded, for example) or 194when the last CPU belonging to the given policy in unregistered. 195 196 197Policy Interface in ``sysfs`` 198============================= 199 200During the initialization of the kernel, the ``CPUFreq`` core creates a 201``sysfs`` directory (kobject) called ``cpufreq`` under 202:file:`/sys/devices/system/cpu/`. 203 204That directory contains a ``policyX`` subdirectory (where ``X`` represents an 205integer number) for every policy object maintained by the ``CPUFreq`` core. 206Each ``policyX`` directory is pointed to by ``cpufreq`` symbolic links 207under :file:`/sys/devices/system/cpu/cpuY/` (where ``Y`` represents an integer 208that may be different from the one represented by ``X``) for all of the CPUs 209associated with (or belonging to) the given policy. The ``policyX`` directories 210in :file:`/sys/devices/system/cpu/cpufreq` each contain policy-specific 211attributes (files) to control ``CPUFreq`` behavior for the corresponding policy 212objects (that is, for all of the CPUs associated with them). 213 214Some of those attributes are generic. They are created by the ``CPUFreq`` core 215and their behavior generally does not depend on what scaling driver is in use 216and what scaling governor is attached to the given policy. Some scaling drivers 217also add driver-specific attributes to the policy directories in ``sysfs`` to 218control policy-specific aspects of driver behavior. 219 220The generic attributes under :file:`/sys/devices/system/cpu/cpufreq/policyX/` 221are the following: 222 223``affected_cpus`` 224 List of online CPUs belonging to this policy (i.e. sharing the hardware 225 performance scaling interface represented by the ``policyX`` policy 226 object). 227 228``bios_limit`` 229 If the platform firmware (BIOS) tells the OS to apply an upper limit to 230 CPU frequencies, that limit will be reported through this attribute (if 231 present). 232 233 The existence of the limit may be a result of some (often unintentional) 234 BIOS settings, restrictions coming from a service processor or another 235 BIOS/HW-based mechanisms. 236 237 This does not cover ACPI thermal limitations which can be discovered 238 through a generic thermal driver. 239 240 This attribute is not present if the scaling driver in use does not 241 support it. 242 243``cpuinfo_cur_freq`` 244 Current frequency of the CPUs belonging to this policy as obtained from 245 the hardware (in KHz). 246 247 This is expected to be the frequency the hardware actually runs at. 248 If that frequency cannot be determined, this attribute should not 249 be present. 250 251``cpuinfo_max_freq`` 252 Maximum possible operating frequency the CPUs belonging to this policy 253 can run at (in kHz). 254 255``cpuinfo_min_freq`` 256 Minimum possible operating frequency the CPUs belonging to this policy 257 can run at (in kHz). 258 259``cpuinfo_transition_latency`` 260 The time it takes to switch the CPUs belonging to this policy from one 261 P-state to another, in nanoseconds. 262 263 If unknown or if known to be so high that the scaling driver does not 264 work with the `ondemand`_ governor, -1 (:c:macro:`CPUFREQ_ETERNAL`) 265 will be returned by reads from this attribute. 266 267``related_cpus`` 268 List of all (online and offline) CPUs belonging to this policy. 269 270``scaling_available_governors`` 271 List of ``CPUFreq`` scaling governors present in the kernel that can 272 be attached to this policy or (if the |intel_pstate| scaling driver is 273 in use) list of scaling algorithms provided by the driver that can be 274 applied to this policy. 275 276 [Note that some governors are modular and it may be necessary to load a 277 kernel module for the governor held by it to become available and be 278 listed by this attribute.] 279 280``scaling_cur_freq`` 281 Current frequency of all of the CPUs belonging to this policy (in kHz). 282 283 In the majority of cases, this is the frequency of the last P-state 284 requested by the scaling driver from the hardware using the scaling 285 interface provided by it, which may or may not reflect the frequency 286 the CPU is actually running at (due to hardware design and other 287 limitations). 288 289 Some architectures (e.g. ``x86``) may attempt to provide information 290 more precisely reflecting the current CPU frequency through this 291 attribute, but that still may not be the exact current CPU frequency as 292 seen by the hardware at the moment. 293 294``scaling_driver`` 295 The scaling driver currently in use. 296 297``scaling_governor`` 298 The scaling governor currently attached to this policy or (if the 299 |intel_pstate| scaling driver is in use) the scaling algorithm 300 provided by the driver that is currently applied to this policy. 301 302 This attribute is read-write and writing to it will cause a new scaling 303 governor to be attached to this policy or a new scaling algorithm 304 provided by the scaling driver to be applied to it (in the 305 |intel_pstate| case), as indicated by the string written to this 306 attribute (which must be one of the names listed by the 307 ``scaling_available_governors`` attribute described above). 308 309``scaling_max_freq`` 310 Maximum frequency the CPUs belonging to this policy are allowed to be 311 running at (in kHz). 312 313 This attribute is read-write and writing a string representing an 314 integer to it will cause a new limit to be set (it must not be lower 315 than the value of the ``scaling_min_freq`` attribute). 316 317``scaling_min_freq`` 318 Minimum frequency the CPUs belonging to this policy are allowed to be 319 running at (in kHz). 320 321 This attribute is read-write and writing a string representing a 322 non-negative integer to it will cause a new limit to be set (it must not 323 be higher than the value of the ``scaling_max_freq`` attribute). 324 325``scaling_setspeed`` 326 This attribute is functional only if the `userspace`_ scaling governor 327 is attached to the given policy. 328 329 It returns the last frequency requested by the governor (in kHz) or can 330 be written to in order to set a new frequency for the policy. 331 332 333Generic Scaling Governors 334========================= 335 336``CPUFreq`` provides generic scaling governors that can be used with all 337scaling drivers. As stated before, each of them implements a single, possibly 338parametrized, performance scaling algorithm. 339 340Scaling governors are attached to policy objects and different policy objects 341can be handled by different scaling governors at the same time (although that 342may lead to suboptimal results in some cases). 343 344The scaling governor for a given policy object can be changed at any time with 345the help of the ``scaling_governor`` policy attribute in ``sysfs``. 346 347Some governors expose ``sysfs`` attributes to control or fine-tune the scaling 348algorithms implemented by them. Those attributes, referred to as governor 349tunables, can be either global (system-wide) or per-policy, depending on the 350scaling driver in use. If the driver requires governor tunables to be 351per-policy, they are located in a subdirectory of each policy directory. 352Otherwise, they are located in a subdirectory under 353:file:`/sys/devices/system/cpu/cpufreq/`. In either case the name of the 354subdirectory containing the governor tunables is the name of the governor 355providing them. 356 357``performance`` 358--------------- 359 360When attached to a policy object, this governor causes the highest frequency, 361within the ``scaling_max_freq`` policy limit, to be requested for that policy. 362 363The request is made once at that time the governor for the policy is set to 364``performance`` and whenever the ``scaling_max_freq`` or ``scaling_min_freq`` 365policy limits change after that. 366 367``powersave`` 368------------- 369 370When attached to a policy object, this governor causes the lowest frequency, 371within the ``scaling_min_freq`` policy limit, to be requested for that policy. 372 373The request is made once at that time the governor for the policy is set to 374``powersave`` and whenever the ``scaling_max_freq`` or ``scaling_min_freq`` 375policy limits change after that. 376 377``userspace`` 378------------- 379 380This governor does not do anything by itself. Instead, it allows user space 381to set the CPU frequency for the policy it is attached to by writing to the 382``scaling_setspeed`` attribute of that policy. 383 384``schedutil`` 385------------- 386 387This governor uses CPU utilization data available from the CPU scheduler. It 388generally is regarded as a part of the CPU scheduler, so it can access the 389scheduler's internal data structures directly. 390 391It runs entirely in scheduler context, although in some cases it may need to 392invoke the scaling driver asynchronously when it decides that the CPU frequency 393should be changed for a given policy (that depends on whether or not the driver 394is capable of changing the CPU frequency from scheduler context). 395 396The actions of this governor for a particular CPU depend on the scheduling class 397invoking its utilization update callback for that CPU. If it is invoked by the 398RT or deadline scheduling classes, the governor will increase the frequency to 399the allowed maximum (that is, the ``scaling_max_freq`` policy limit). In turn, 400if it is invoked by the CFS scheduling class, the governor will use the 401Per-Entity Load Tracking (PELT) metric for the root control group of the 402given CPU as the CPU utilization estimate (see the *Per-entity load tracking* 403LWN.net article [1]_ for a description of the PELT mechanism). Then, the new 404CPU frequency to apply is computed in accordance with the formula 405 406 f = 1.25 * ``f_0`` * ``util`` / ``max`` 407 408where ``util`` is the PELT number, ``max`` is the theoretical maximum of 409``util``, and ``f_0`` is either the maximum possible CPU frequency for the given 410policy (if the PELT number is frequency-invariant), or the current CPU frequency 411(otherwise). 412 413This governor also employs a mechanism allowing it to temporarily bump up the 414CPU frequency for tasks that have been waiting on I/O most recently, called 415"IO-wait boosting". That happens when the :c:macro:`SCHED_CPUFREQ_IOWAIT` flag 416is passed by the scheduler to the governor callback which causes the frequency 417to go up to the allowed maximum immediately and then draw back to the value 418returned by the above formula over time. 419 420This governor exposes only one tunable: 421 422``rate_limit_us`` 423 Minimum time (in microseconds) that has to pass between two consecutive 424 runs of governor computations (default: 1000 times the scaling driver's 425 transition latency). 426 427 The purpose of this tunable is to reduce the scheduler context overhead 428 of the governor which might be excessive without it. 429 430This governor generally is regarded as a replacement for the older `ondemand`_ 431and `conservative`_ governors (described below), as it is simpler and more 432tightly integrated with the CPU scheduler, its overhead in terms of CPU context 433switches and similar is less significant, and it uses the scheduler's own CPU 434utilization metric, so in principle its decisions should not contradict the 435decisions made by the other parts of the scheduler. 436 437``ondemand`` 438------------ 439 440This governor uses CPU load as a CPU frequency selection metric. 441 442In order to estimate the current CPU load, it measures the time elapsed between 443consecutive invocations of its worker routine and computes the fraction of that 444time in which the given CPU was not idle. The ratio of the non-idle (active) 445time to the total CPU time is taken as an estimate of the load. 446 447If this governor is attached to a policy shared by multiple CPUs, the load is 448estimated for all of them and the greatest result is taken as the load estimate 449for the entire policy. 450 451The worker routine of this governor has to run in process context, so it is 452invoked asynchronously (via a workqueue) and CPU P-states are updated from 453there if necessary. As a result, the scheduler context overhead from this 454governor is minimum, but it causes additional CPU context switches to happen 455relatively often and the CPU P-state updates triggered by it can be relatively 456irregular. Also, it affects its own CPU load metric by running code that 457reduces the CPU idle time (even though the CPU idle time is only reduced very 458slightly by it). 459 460It generally selects CPU frequencies proportional to the estimated load, so that 461the value of the ``cpuinfo_max_freq`` policy attribute corresponds to the load of 4621 (or 100%), and the value of the ``cpuinfo_min_freq`` policy attribute 463corresponds to the load of 0, unless when the load exceeds a (configurable) 464speedup threshold, in which case it will go straight for the highest frequency 465it is allowed to use (the ``scaling_max_freq`` policy limit). 466 467This governor exposes the following tunables: 468 469``sampling_rate`` 470 This is how often the governor's worker routine should run, in 471 microseconds. 472 473 Typically, it is set to values of the order of 10000 (10 ms). Its 474 default value is equal to the value of ``cpuinfo_transition_latency`` 475 for each policy this governor is attached to (but since the unit here 476 is greater by 1000, this means that the time represented by 477 ``sampling_rate`` is 1000 times greater than the transition latency by 478 default). 479 480 If this tunable is per-policy, the following shell command sets the time 481 represented by it to be 750 times as high as the transition latency:: 482 483 # echo `$(($(cat cpuinfo_transition_latency) * 750 / 1000)) > ondemand/sampling_rate 484 485``up_threshold`` 486 If the estimated CPU load is above this value (in percent), the governor 487 will set the frequency to the maximum value allowed for the policy. 488 Otherwise, the selected frequency will be proportional to the estimated 489 CPU load. 490 491``ignore_nice_load`` 492 If set to 1 (default 0), it will cause the CPU load estimation code to 493 treat the CPU time spent on executing tasks with "nice" levels greater 494 than 0 as CPU idle time. 495 496 This may be useful if there are tasks in the system that should not be 497 taken into account when deciding what frequency to run the CPUs at. 498 Then, to make that happen it is sufficient to increase the "nice" level 499 of those tasks above 0 and set this attribute to 1. 500 501``sampling_down_factor`` 502 Temporary multiplier, between 1 (default) and 100 inclusive, to apply to 503 the ``sampling_rate`` value if the CPU load goes above ``up_threshold``. 504 505 This causes the next execution of the governor's worker routine (after 506 setting the frequency to the allowed maximum) to be delayed, so the 507 frequency stays at the maximum level for a longer time. 508 509 Frequency fluctuations in some bursty workloads may be avoided this way 510 at the cost of additional energy spent on maintaining the maximum CPU 511 capacity. 512 513``powersave_bias`` 514 Reduction factor to apply to the original frequency target of the 515 governor (including the maximum value used when the ``up_threshold`` 516 value is exceeded by the estimated CPU load) or sensitivity threshold 517 for the AMD frequency sensitivity powersave bias driver 518 (:file:`drivers/cpufreq/amd_freq_sensitivity.c`), between 0 and 1000 519 inclusive. 520 521 If the AMD frequency sensitivity powersave bias driver is not loaded, 522 the effective frequency to apply is given by 523 524 f * (1 - ``powersave_bias`` / 1000) 525 526 where f is the governor's original frequency target. The default value 527 of this attribute is 0 in that case. 528 529 If the AMD frequency sensitivity powersave bias driver is loaded, the 530 value of this attribute is 400 by default and it is used in a different 531 way. 532 533 On Family 16h (and later) AMD processors there is a mechanism to get a 534 measured workload sensitivity, between 0 and 100% inclusive, from the 535 hardware. That value can be used to estimate how the performance of the 536 workload running on a CPU will change in response to frequency changes. 537 538 The performance of a workload with the sensitivity of 0 (memory-bound or 539 IO-bound) is not expected to increase at all as a result of increasing 540 the CPU frequency, whereas workloads with the sensitivity of 100% 541 (CPU-bound) are expected to perform much better if the CPU frequency is 542 increased. 543 544 If the workload sensitivity is less than the threshold represented by 545 the ``powersave_bias`` value, the sensitivity powersave bias driver 546 will cause the governor to select a frequency lower than its original 547 target, so as to avoid over-provisioning workloads that will not benefit 548 from running at higher CPU frequencies. 549 550``conservative`` 551---------------- 552 553This governor uses CPU load as a CPU frequency selection metric. 554 555It estimates the CPU load in the same way as the `ondemand`_ governor described 556above, but the CPU frequency selection algorithm implemented by it is different. 557 558Namely, it avoids changing the frequency significantly over short time intervals 559which may not be suitable for systems with limited power supply capacity (e.g. 560battery-powered). To achieve that, it changes the frequency in relatively 561small steps, one step at a time, up or down - depending on whether or not a 562(configurable) threshold has been exceeded by the estimated CPU load. 563 564This governor exposes the following tunables: 565 566``freq_step`` 567 Frequency step in percent of the maximum frequency the governor is 568 allowed to set (the ``scaling_max_freq`` policy limit), between 0 and 569 100 (5 by default). 570 571 This is how much the frequency is allowed to change in one go. Setting 572 it to 0 will cause the default frequency step (5 percent) to be used 573 and setting it to 100 effectively causes the governor to periodically 574 switch the frequency between the ``scaling_min_freq`` and 575 ``scaling_max_freq`` policy limits. 576 577``down_threshold`` 578 Threshold value (in percent, 20 by default) used to determine the 579 frequency change direction. 580 581 If the estimated CPU load is greater than this value, the frequency will 582 go up (by ``freq_step``). If the load is less than this value (and the 583 ``sampling_down_factor`` mechanism is not in effect), the frequency will 584 go down. Otherwise, the frequency will not be changed. 585 586``sampling_down_factor`` 587 Frequency decrease deferral factor, between 1 (default) and 10 588 inclusive. 589 590 It effectively causes the frequency to go down ``sampling_down_factor`` 591 times slower than it ramps up. 592 593 594Frequency Boost Support 595======================= 596 597Background 598---------- 599 600Some processors support a mechanism to raise the operating frequency of some 601cores in a multicore package temporarily (and above the sustainable frequency 602threshold for the whole package) under certain conditions, for example if the 603whole chip is not fully utilized and below its intended thermal or power budget. 604 605Different names are used by different vendors to refer to this functionality. 606For Intel processors it is referred to as "Turbo Boost", AMD calls it 607"Turbo-Core" or (in technical documentation) "Core Performance Boost" and so on. 608As a rule, it also is implemented differently by different vendors. The simple 609term "frequency boost" is used here for brevity to refer to all of those 610implementations. 611 612The frequency boost mechanism may be either hardware-based or software-based. 613If it is hardware-based (e.g. on x86), the decision to trigger the boosting is 614made by the hardware (although in general it requires the hardware to be put 615into a special state in which it can control the CPU frequency within certain 616limits). If it is software-based (e.g. on ARM), the scaling driver decides 617whether or not to trigger boosting and when to do that. 618 619The ``boost`` File in ``sysfs`` 620------------------------------- 621 622This file is located under :file:`/sys/devices/system/cpu/cpufreq/` and controls 623the "boost" setting for the whole system. It is not present if the underlying 624scaling driver does not support the frequency boost mechanism (or supports it, 625but provides a driver-specific interface for controlling it, like 626|intel_pstate|). 627 628If the value in this file is 1, the frequency boost mechanism is enabled. This 629means that either the hardware can be put into states in which it is able to 630trigger boosting (in the hardware-based case), or the software is allowed to 631trigger boosting (in the software-based case). It does not mean that boosting 632is actually in use at the moment on any CPUs in the system. It only means a 633permission to use the frequency boost mechanism (which still may never be used 634for other reasons). 635 636If the value in this file is 0, the frequency boost mechanism is disabled and 637cannot be used at all. 638 639The only values that can be written to this file are 0 and 1. 640 641Rationale for Boost Control Knob 642-------------------------------- 643 644The frequency boost mechanism is generally intended to help to achieve optimum 645CPU performance on time scales below software resolution (e.g. below the 646scheduler tick interval) and it is demonstrably suitable for many workloads, but 647it may lead to problems in certain situations. 648 649For this reason, many systems make it possible to disable the frequency boost 650mechanism in the platform firmware (BIOS) setup, but that requires the system to 651be restarted for the setting to be adjusted as desired, which may not be 652practical at least in some cases. For example: 653 654 1. Boosting means overclocking the processor, although under controlled 655 conditions. Generally, the processor's energy consumption increases 656 as a result of increasing its frequency and voltage, even temporarily. 657 That may not be desirable on systems that switch to power sources of 658 limited capacity, such as batteries, so the ability to disable the boost 659 mechanism while the system is running may help there (but that depends on 660 the workload too). 661 662 2. In some situations deterministic behavior is more important than 663 performance or energy consumption (or both) and the ability to disable 664 boosting while the system is running may be useful then. 665 666 3. To examine the impact of the frequency boost mechanism itself, it is useful 667 to be able to run tests with and without boosting, preferably without 668 restarting the system in the meantime. 669 670 4. Reproducible results are important when running benchmarks. Since 671 the boosting functionality depends on the load of the whole package, 672 single-thread performance may vary because of it which may lead to 673 unreproducible results sometimes. That can be avoided by disabling the 674 frequency boost mechanism before running benchmarks sensitive to that 675 issue. 676 677Legacy AMD ``cpb`` Knob 678----------------------- 679 680The AMD powernow-k8 scaling driver supports a ``sysfs`` knob very similar to 681the global ``boost`` one. It is used for disabling/enabling the "Core 682Performance Boost" feature of some AMD processors. 683 684If present, that knob is located in every ``CPUFreq`` policy directory in 685``sysfs`` (:file:`/sys/devices/system/cpu/cpufreq/policyX/`) and is called 686``cpb``, which indicates a more fine grained control interface. The actual 687implementation, however, works on the system-wide basis and setting that knob 688for one policy causes the same value of it to be set for all of the other 689policies at the same time. 690 691That knob is still supported on AMD processors that support its underlying 692hardware feature, but it may be configured out of the kernel (via the 693:c:macro:`CONFIG_X86_ACPI_CPUFREQ_CPB` configuration option) and the global 694``boost`` knob is present regardless. Thus it is always possible use the 695``boost`` knob instead of the ``cpb`` one which is highly recommended, as that 696is more consistent with what all of the other systems do (and the ``cpb`` knob 697may not be supported any more in the future). 698 699The ``cpb`` knob is never present for any processors without the underlying 700hardware feature (e.g. all Intel ones), even if the 701:c:macro:`CONFIG_X86_ACPI_CPUFREQ_CPB` configuration option is set. 702 703 704References 705========== 706 707.. [1] Jonathan Corbet, *Per-entity load tracking*, 708 https://lwn.net/Articles/531853/ 709