1======================= 2Energy Aware Scheduling 3======================= 4 51. Introduction 6--------------- 7 8Energy Aware Scheduling (or EAS) gives the scheduler the ability to predict 9the impact of its decisions on the energy consumed by CPUs. EAS relies on an 10Energy Model (EM) of the CPUs to select an energy efficient CPU for each task, 11with a minimal impact on throughput. This document aims at providing an 12introduction on how EAS works, what are the main design decisions behind it, and 13details what is needed to get it to run. 14 15Before going any further, please note that at the time of writing:: 16 17 /!\ EAS does not support platforms with symmetric CPU topologies /!\ 18 19EAS operates only on heterogeneous CPU topologies (such as Arm big.LITTLE) 20because this is where the potential for saving energy through scheduling is 21the highest. 22 23The actual EM used by EAS is _not_ maintained by the scheduler, but by a 24dedicated framework. For details about this framework and what it provides, 25please refer to its documentation (see Documentation/power/energy-model.rst). 26 27 282. Background and Terminology 29----------------------------- 30 31To make it clear from the start: 32 - energy = [joule] (resource like a battery on powered devices) 33 - power = energy/time = [joule/second] = [watt] 34 35The goal of EAS is to minimize energy, while still getting the job done. That 36is, we want to maximize:: 37 38 performance [inst/s] 39 -------------------- 40 power [W] 41 42which is equivalent to minimizing:: 43 44 energy [J] 45 ----------- 46 instruction 47 48while still getting 'good' performance. It is essentially an alternative 49optimization objective to the current performance-only objective for the 50scheduler. This alternative considers two objectives: energy-efficiency and 51performance. 52 53The idea behind introducing an EM is to allow the scheduler to evaluate the 54implications of its decisions rather than blindly applying energy-saving 55techniques that may have positive effects only on some platforms. At the same 56time, the EM must be as simple as possible to minimize the scheduler latency 57impact. 58 59In short, EAS changes the way CFS tasks are assigned to CPUs. When it is time 60for the scheduler to decide where a task should run (during wake-up), the EM 61is used to break the tie between several good CPU candidates and pick the one 62that is predicted to yield the best energy consumption without harming the 63system's throughput. The predictions made by EAS rely on specific elements of 64knowledge about the platform's topology, which include the 'capacity' of CPUs, 65and their respective energy costs. 66 67 683. Topology information 69----------------------- 70 71EAS (as well as the rest of the scheduler) uses the notion of 'capacity' to 72differentiate CPUs with different computing throughput. The 'capacity' of a CPU 73represents the amount of work it can absorb when running at its highest 74frequency compared to the most capable CPU of the system. Capacity values are 75normalized in a 1024 range, and are comparable with the utilization signals of 76tasks and CPUs computed by the Per-Entity Load Tracking (PELT) mechanism. Thanks 77to capacity and utilization values, EAS is able to estimate how big/busy a 78task/CPU is, and to take this into consideration when evaluating performance vs 79energy trade-offs. The capacity of CPUs is provided via arch-specific code 80through the arch_scale_cpu_capacity() callback. 81 82The rest of platform knowledge used by EAS is directly read from the Energy 83Model (EM) framework. The EM of a platform is composed of a power cost table 84per 'performance domain' in the system (see Documentation/power/energy-model.rst 85for futher details about performance domains). 86 87The scheduler manages references to the EM objects in the topology code when the 88scheduling domains are built, or re-built. For each root domain (rd), the 89scheduler maintains a singly linked list of all performance domains intersecting 90the current rd->span. Each node in the list contains a pointer to a struct 91em_perf_domain as provided by the EM framework. 92 93The lists are attached to the root domains in order to cope with exclusive 94cpuset configurations. Since the boundaries of exclusive cpusets do not 95necessarily match those of performance domains, the lists of different root 96domains can contain duplicate elements. 97 98Example 1. 99 Let us consider a platform with 12 CPUs, split in 3 performance domains 100 (pd0, pd4 and pd8), organized as follows:: 101 102 CPUs: 0 1 2 3 4 5 6 7 8 9 10 11 103 PDs: |--pd0--|--pd4--|---pd8---| 104 RDs: |----rd1----|-----rd2-----| 105 106 Now, consider that userspace decided to split the system with two 107 exclusive cpusets, hence creating two independent root domains, each 108 containing 6 CPUs. The two root domains are denoted rd1 and rd2 in the 109 above figure. Since pd4 intersects with both rd1 and rd2, it will be 110 present in the linked list '->pd' attached to each of them: 111 112 * rd1->pd: pd0 -> pd4 113 * rd2->pd: pd4 -> pd8 114 115 Please note that the scheduler will create two duplicate list nodes for 116 pd4 (one for each list). However, both just hold a pointer to the same 117 shared data structure of the EM framework. 118 119Since the access to these lists can happen concurrently with hotplug and other 120things, they are protected by RCU, like the rest of topology structures 121manipulated by the scheduler. 122 123EAS also maintains a static key (sched_energy_present) which is enabled when at 124least one root domain meets all conditions for EAS to start. Those conditions 125are summarized in Section 6. 126 127 1284. Energy-Aware task placement 129------------------------------ 130 131EAS overrides the CFS task wake-up balancing code. It uses the EM of the 132platform and the PELT signals to choose an energy-efficient target CPU during 133wake-up balance. When EAS is enabled, select_task_rq_fair() calls 134find_energy_efficient_cpu() to do the placement decision. This function looks 135for the CPU with the highest spare capacity (CPU capacity - CPU utilization) in 136each performance domain since it is the one which will allow us to keep the 137frequency the lowest. Then, the function checks if placing the task there could 138save energy compared to leaving it on prev_cpu, i.e. the CPU where the task ran 139in its previous activation. 140 141find_energy_efficient_cpu() uses compute_energy() to estimate what will be the 142energy consumed by the system if the waking task was migrated. compute_energy() 143looks at the current utilization landscape of the CPUs and adjusts it to 144'simulate' the task migration. The EM framework provides the em_pd_energy() API 145which computes the expected energy consumption of each performance domain for 146the given utilization landscape. 147 148An example of energy-optimized task placement decision is detailed below. 149 150Example 2. 151 Let us consider a (fake) platform with 2 independent performance domains 152 composed of two CPUs each. CPU0 and CPU1 are little CPUs; CPU2 and CPU3 153 are big. 154 155 The scheduler must decide where to place a task P whose util_avg = 200 156 and prev_cpu = 0. 157 158 The current utilization landscape of the CPUs is depicted on the graph 159 below. CPUs 0-3 have a util_avg of 400, 100, 600 and 500 respectively 160 Each performance domain has three Operating Performance Points (OPPs). 161 The CPU capacity and power cost associated with each OPP is listed in 162 the Energy Model table. The util_avg of P is shown on the figures 163 below as 'PP':: 164 165 CPU util. 166 1024 - - - - - - - Energy Model 167 +-----------+-------------+ 168 | Little | Big | 169 768 ============= +-----+-----+------+------+ 170 | Cap | Pwr | Cap | Pwr | 171 +-----+-----+------+------+ 172 512 =========== - ##- - - - - | 170 | 50 | 512 | 400 | 173 ## ## | 341 | 150 | 768 | 800 | 174 341 -PP - - - - ## ## | 512 | 300 | 1024 | 1700 | 175 PP ## ## +-----+-----+------+------+ 176 170 -## - - - - ## ## 177 ## ## ## ## 178 ------------ ------------- 179 CPU0 CPU1 CPU2 CPU3 180 181 Current OPP: ===== Other OPP: - - - util_avg (100 each): ## 182 183 184 find_energy_efficient_cpu() will first look for the CPUs with the 185 maximum spare capacity in the two performance domains. In this example, 186 CPU1 and CPU3. Then it will estimate the energy of the system if P was 187 placed on either of them, and check if that would save some energy 188 compared to leaving P on CPU0. EAS assumes that OPPs follow utilization 189 (which is coherent with the behaviour of the schedutil CPUFreq 190 governor, see Section 6. for more details on this topic). 191 192 **Case 1. P is migrated to CPU1**:: 193 194 1024 - - - - - - - 195 196 Energy calculation: 197 768 ============= * CPU0: 200 / 341 * 150 = 88 198 * CPU1: 300 / 341 * 150 = 131 199 * CPU2: 600 / 768 * 800 = 625 200 512 - - - - - - - ##- - - - - * CPU3: 500 / 768 * 800 = 520 201 ## ## => total_energy = 1364 202 341 =========== ## ## 203 PP ## ## 204 170 -## - - PP- ## ## 205 ## ## ## ## 206 ------------ ------------- 207 CPU0 CPU1 CPU2 CPU3 208 209 210 **Case 2. P is migrated to CPU3**:: 211 212 1024 - - - - - - - 213 214 Energy calculation: 215 768 ============= * CPU0: 200 / 341 * 150 = 88 216 * CPU1: 100 / 341 * 150 = 43 217 PP * CPU2: 600 / 768 * 800 = 625 218 512 - - - - - - - ##- - -PP - * CPU3: 700 / 768 * 800 = 729 219 ## ## => total_energy = 1485 220 341 =========== ## ## 221 ## ## 222 170 -## - - - - ## ## 223 ## ## ## ## 224 ------------ ------------- 225 CPU0 CPU1 CPU2 CPU3 226 227 228 **Case 3. P stays on prev_cpu / CPU 0**:: 229 230 1024 - - - - - - - 231 232 Energy calculation: 233 768 ============= * CPU0: 400 / 512 * 300 = 234 234 * CPU1: 100 / 512 * 300 = 58 235 * CPU2: 600 / 768 * 800 = 625 236 512 =========== - ##- - - - - * CPU3: 500 / 768 * 800 = 520 237 ## ## => total_energy = 1437 238 341 -PP - - - - ## ## 239 PP ## ## 240 170 -## - - - - ## ## 241 ## ## ## ## 242 ------------ ------------- 243 CPU0 CPU1 CPU2 CPU3 244 245 246 From these calculations, the Case 1 has the lowest total energy. So CPU 1 247 is be the best candidate from an energy-efficiency standpoint. 248 249Big CPUs are generally more power hungry than the little ones and are thus used 250mainly when a task doesn't fit the littles. However, little CPUs aren't always 251necessarily more energy-efficient than big CPUs. For some systems, the high OPPs 252of the little CPUs can be less energy-efficient than the lowest OPPs of the 253bigs, for example. So, if the little CPUs happen to have enough utilization at 254a specific point in time, a small task waking up at that moment could be better 255of executing on the big side in order to save energy, even though it would fit 256on the little side. 257 258And even in the case where all OPPs of the big CPUs are less energy-efficient 259than those of the little, using the big CPUs for a small task might still, under 260specific conditions, save energy. Indeed, placing a task on a little CPU can 261result in raising the OPP of the entire performance domain, and that will 262increase the cost of the tasks already running there. If the waking task is 263placed on a big CPU, its own execution cost might be higher than if it was 264running on a little, but it won't impact the other tasks of the little CPUs 265which will keep running at a lower OPP. So, when considering the total energy 266consumed by CPUs, the extra cost of running that one task on a big core can be 267smaller than the cost of raising the OPP on the little CPUs for all the other 268tasks. 269 270The examples above would be nearly impossible to get right in a generic way, and 271for all platforms, without knowing the cost of running at different OPPs on all 272CPUs of the system. Thanks to its EM-based design, EAS should cope with them 273correctly without too many troubles. However, in order to ensure a minimal 274impact on throughput for high-utilization scenarios, EAS also implements another 275mechanism called 'over-utilization'. 276 277 2785. Over-utilization 279------------------- 280 281From a general standpoint, the use-cases where EAS can help the most are those 282involving a light/medium CPU utilization. Whenever long CPU-bound tasks are 283being run, they will require all of the available CPU capacity, and there isn't 284much that can be done by the scheduler to save energy without severly harming 285throughput. In order to avoid hurting performance with EAS, CPUs are flagged as 286'over-utilized' as soon as they are used at more than 80% of their compute 287capacity. As long as no CPUs are over-utilized in a root domain, load balancing 288is disabled and EAS overridess the wake-up balancing code. EAS is likely to load 289the most energy efficient CPUs of the system more than the others if that can be 290done without harming throughput. So, the load-balancer is disabled to prevent 291it from breaking the energy-efficient task placement found by EAS. It is safe to 292do so when the system isn't overutilized since being below the 80% tipping point 293implies that: 294 295 a. there is some idle time on all CPUs, so the utilization signals used by 296 EAS are likely to accurately represent the 'size' of the various tasks 297 in the system; 298 b. all tasks should already be provided with enough CPU capacity, 299 regardless of their nice values; 300 c. since there is spare capacity all tasks must be blocking/sleeping 301 regularly and balancing at wake-up is sufficient. 302 303As soon as one CPU goes above the 80% tipping point, at least one of the three 304assumptions above becomes incorrect. In this scenario, the 'overutilized' flag 305is raised for the entire root domain, EAS is disabled, and the load-balancer is 306re-enabled. By doing so, the scheduler falls back onto load-based algorithms for 307wake-up and load balance under CPU-bound conditions. This provides a better 308respect of the nice values of tasks. 309 310Since the notion of overutilization largely relies on detecting whether or not 311there is some idle time in the system, the CPU capacity 'stolen' by higher 312(than CFS) scheduling classes (as well as IRQ) must be taken into account. As 313such, the detection of overutilization accounts for the capacity used not only 314by CFS tasks, but also by the other scheduling classes and IRQ. 315 316 3176. Dependencies and requirements for EAS 318---------------------------------------- 319 320Energy Aware Scheduling depends on the CPUs of the system having specific 321hardware properties and on other features of the kernel being enabled. This 322section lists these dependencies and provides hints as to how they can be met. 323 324 3256.1 - Asymmetric CPU topology 326^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 327 328 329As mentioned in the introduction, EAS is only supported on platforms with 330asymmetric CPU topologies for now. This requirement is checked at run-time by 331looking for the presence of the SD_ASYM_CPUCAPACITY flag when the scheduling 332domains are built. 333 334See Documentation/sched/sched-capacity.rst for requirements to be met for this 335flag to be set in the sched_domain hierarchy. 336 337Please note that EAS is not fundamentally incompatible with SMP, but no 338significant savings on SMP platforms have been observed yet. This restriction 339could be amended in the future if proven otherwise. 340 341 3426.2 - Energy Model presence 343^^^^^^^^^^^^^^^^^^^^^^^^^^^ 344 345EAS uses the EM of a platform to estimate the impact of scheduling decisions on 346energy. So, your platform must provide power cost tables to the EM framework in 347order to make EAS start. To do so, please refer to documentation of the 348independent EM framework in Documentation/power/energy-model.rst. 349 350Please also note that the scheduling domains need to be re-built after the 351EM has been registered in order to start EAS. 352 353 3546.3 - Energy Model complexity 355^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 356 357The task wake-up path is very latency-sensitive. When the EM of a platform is 358too complex (too many CPUs, too many performance domains, too many performance 359states, ...), the cost of using it in the wake-up path can become prohibitive. 360The energy-aware wake-up algorithm has a complexity of: 361 362 C = Nd * (Nc + Ns) 363 364with: Nd the number of performance domains; Nc the number of CPUs; and Ns the 365total number of OPPs (ex: for two perf. domains with 4 OPPs each, Ns = 8). 366 367A complexity check is performed at the root domain level, when scheduling 368domains are built. EAS will not start on a root domain if its C happens to be 369higher than the completely arbitrary EM_MAX_COMPLEXITY threshold (2048 at the 370time of writing). 371 372If you really want to use EAS but the complexity of your platform's Energy 373Model is too high to be used with a single root domain, you're left with only 374two possible options: 375 376 1. split your system into separate, smaller, root domains using exclusive 377 cpusets and enable EAS locally on each of them. This option has the 378 benefit to work out of the box but the drawback of preventing load 379 balance between root domains, which can result in an unbalanced system 380 overall; 381 2. submit patches to reduce the complexity of the EAS wake-up algorithm, 382 hence enabling it to cope with larger EMs in reasonable time. 383 384 3856.4 - Schedutil governor 386^^^^^^^^^^^^^^^^^^^^^^^^ 387 388EAS tries to predict at which OPP will the CPUs be running in the close future 389in order to estimate their energy consumption. To do so, it is assumed that OPPs 390of CPUs follow their utilization. 391 392Although it is very difficult to provide hard guarantees regarding the accuracy 393of this assumption in practice (because the hardware might not do what it is 394told to do, for example), schedutil as opposed to other CPUFreq governors at 395least _requests_ frequencies calculated using the utilization signals. 396Consequently, the only sane governor to use together with EAS is schedutil, 397because it is the only one providing some degree of consistency between 398frequency requests and energy predictions. 399 400Using EAS with any other governor than schedutil is not supported. 401 402 4036.5 Scale-invariant utilization signals 404^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 405 406In order to make accurate prediction across CPUs and for all performance 407states, EAS needs frequency-invariant and CPU-invariant PELT signals. These can 408be obtained using the architecture-defined arch_scale{cpu,freq}_capacity() 409callbacks. 410 411Using EAS on a platform that doesn't implement these two callbacks is not 412supported. 413 414 4156.6 Multithreading (SMT) 416^^^^^^^^^^^^^^^^^^^^^^^^ 417 418EAS in its current form is SMT unaware and is not able to leverage 419multithreaded hardware to save energy. EAS considers threads as independent 420CPUs, which can actually be counter-productive for both performance and energy. 421 422EAS on SMT is not supported. 423