1 /* Copyright (c) 2010-2018. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "simgrid/plugins/dvfs.h"
7 #include "simgrid/plugins/load.h"
8 #include "simgrid/s4u/Engine.hpp"
9 #include "src/kernel/activity/ExecImpl.hpp"
10 #include "src/plugins/vm/VirtualMachineImpl.hpp"
11 #include "src/smpi/plugins/ampi/ampi.hpp"
12 #include <xbt/config.hpp>
14 #include <boost/algorithm/string.hpp>
16 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
18 static simgrid::config::Flag<double> cfg_sampling_rate("plugin/dvfs/sampling-rate", {"plugin/dvfs/sampling_rate"},
19 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
20 [](double val){if (val != 0.1) sg_host_dvfs_plugin_init();});
22 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
23 "Which Governor should be used that adapts the CPU frequency?", "performance",
25 std::map<std::string, std::string>({
26 {"adagio", "TODO: Doc"},
27 {"conservative", "TODO: Doc"},
28 {"ondemand", "TODO: Doc"},
29 {"performance", "TODO: Doc"},
30 {"powersave", "TODO: Doc"},
33 [](std::string val) { if (val != "performance") sg_host_dvfs_plugin_init(); });
35 static simgrid::config::Flag<int> cfg_min_pstate("plugin/dvfs/min-pstate", {"plugin/dvfs/min_pstate"},
36 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0,
39 static const int max_pstate_not_limited = -1;
40 static simgrid::config::Flag<int> cfg_max_pstate("plugin/dvfs/max-pstate", {"plugin/dvfs/max_pstate"},
41 "Which pstate is the maximum (and hence slowest) pstate for this governor?", max_pstate_not_limited,
44 /** @addtogroup SURF_plugin_load
46 This plugin makes it very simple for users to obtain the current load for each host.
50 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_plugin_dvfs, surf, "Logging specific to the SURF HostDvfs plugin");
58 * Add this to your host tag:
59 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
61 * Valid values as of now are: performance, powersave, ondemand, conservative
62 * It doesn't matter if you use uppercase or lowercase.
64 * For the sampling rate, use this:
66 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
68 * This will run the update() method of the specified governor every 2 seconds
71 * These properties can also be used within the \<config\> tag to configure
72 * these values globally. Using them within the \<host\> will overwrite this
73 * global configuration
78 simgrid::s4u::Host* const host_;
79 double sampling_rate_;
80 int min_pstate; //< Never use a pstate less than this one
81 int max_pstate; //< Never use a pstate larger than this one
84 explicit Governor(simgrid::s4u::Host* ptr)
86 , min_pstate(cfg_min_pstate)
87 , max_pstate(cfg_max_pstate == max_pstate_not_limited ? host_->get_pstate_count() - 1 : cfg_max_pstate)
91 virtual ~Governor() = default;
92 virtual std::string get_name() const = 0;
93 simgrid::s4u::Host* get_host() const { return host_; }
94 int get_min_pstate() const { return min_pstate; }
95 int get_max_pstate() const { return max_pstate; }
99 const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name());
100 if (local_sampling_rate_config != nullptr) {
101 sampling_rate_ = std::stod(local_sampling_rate_config);
103 sampling_rate_ = cfg_sampling_rate;
105 const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name());
106 if (local_min_pstate_config != nullptr) {
107 min_pstate = std::stoi(local_min_pstate_config);
110 const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name());
111 if (local_max_pstate_config != nullptr) {
112 max_pstate = std::stod(local_max_pstate_config);
114 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
115 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
116 xbt_assert(0 <= min_pstate, "min_pstate is negative!");
119 virtual void update() = 0;
120 double get_sampling_rate() const { return sampling_rate_; }
124 * The linux kernel doc describes this governor as follows:
125 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
127 * > The CPUfreq governor "performance" sets the CPU statically to the
128 * > highest frequency within the borders of scaling_min_freq and
129 * > scaling_max_freq.
131 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
133 class Performance : public Governor {
135 explicit Performance(simgrid::s4u::Host* ptr) : Governor(ptr) {}
136 std::string get_name() const override { return "Performance"; }
138 void update() override { get_host()->set_pstate(get_min_pstate()); }
142 * The linux kernel doc describes this governor as follows:
143 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
145 * > The CPUfreq governor "powersave" sets the CPU statically to the
146 * > lowest frequency within the borders of scaling_min_freq and
147 * > scaling_max_freq.
149 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
151 class Powersave : public Governor {
153 explicit Powersave(simgrid::s4u::Host* ptr) : Governor(ptr) {}
154 std::string get_name() const override { return "Powersave"; }
156 void update() override { get_host()->set_pstate(get_max_pstate()); }
160 * The linux kernel doc describes this governor as follows:
161 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
163 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
164 * > current system load. [...] when triggered, cpufreq checks
165 * > the CPU-usage statistics over the last period and the governor sets the
168 class OnDemand : public Governor {
170 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
171 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
173 double freq_up_threshold_ = 0.80;
176 explicit OnDemand(simgrid::s4u::Host* ptr) : Governor(ptr) {}
177 std::string get_name() const override { return "OnDemand"; }
179 void update() override
181 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
182 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
184 if (load > freq_up_threshold_) {
185 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
186 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %i", load, freq_up_threshold_, get_min_pstate());
188 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
190 * freq_next = min_f + load * (max_f - min_f) / 100
192 * So they assume that frequency increases by 100 MHz. We will just use
193 * lowest_pstate - load*pstatesCount()
195 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
196 // because pstate 0 can only be selected if load > freq_up_threshold_
197 int new_pstate = get_max_pstate() - load * (get_max_pstate() + 1);
198 if (new_pstate < get_min_pstate())
199 new_pstate = get_min_pstate();
200 get_host()->set_pstate(new_pstate);
202 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %i", load, freq_up_threshold_, new_pstate);
208 * This is the conservative governor, which is very similar to the
209 * OnDemand governor. The Linux Kernel Documentation describes it
210 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
212 * > The CPUfreq governor "conservative", much like the "ondemand"
213 * > governor, sets the CPU frequency depending on the current usage. It
214 * > differs in behaviour in that it gracefully increases and decreases the
215 * > CPU speed rather than jumping to max speed the moment there is any load
216 * > on the CPU. This behaviour is more suitable in a battery powered
219 class Conservative : public Governor {
220 double freq_up_threshold_ = .8;
221 double freq_down_threshold_ = .2;
224 explicit Conservative(simgrid::s4u::Host* ptr) : Governor(ptr) {}
225 virtual std::string get_name() const override { return "Conservative"; }
227 virtual void update() override
229 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
230 int pstate = get_host()->get_pstate();
231 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
233 if (load > freq_up_threshold_) {
234 if (pstate != get_min_pstate()) {
235 get_host()->set_pstate(pstate - 1);
236 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %d", load, freq_up_threshold_,
239 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %d", load,
240 freq_up_threshold_, pstate);
242 } else if (load < freq_down_threshold_) {
243 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
244 get_host()->set_pstate(pstate + 1);
245 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %d", load, freq_down_threshold_, pstate + 1);
247 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %d", load,
248 freq_down_threshold_, pstate);
254 class Adagio : public Governor {
257 double start_time = 0;
258 double comp_counter = 0;
259 double comp_timer = 0;
261 std::vector<std::vector<double>> rates;
263 unsigned int task_id = 0;
264 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
267 explicit Adagio(simgrid::s4u::Host* ptr)
268 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
270 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::ActorPtr actor) {
271 // Every instance of this class subscribes to this event, so one per host
272 // This means that for any actor, all 'hosts' are normally notified of these
273 // changes, even those who don't currently run the actor 'proc_id'.
274 // -> Let's check if this signal call is for us!
275 if (get_host() == actor->get_host()) {
276 iteration_running = true;
279 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::ActorPtr actor) {
280 if (get_host() == actor->get_host()) {
281 iteration_running = false;
285 simgrid::kernel::activity::ExecImpl::on_creation.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
286 if (activity->host_ == get_host())
289 simgrid::kernel::activity::ExecImpl::on_completion.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
290 // For more than one host (not yet supported), we can access the host via
291 // simcalls_.front()->issuer->iface()->get_host()
292 if (activity->host_ == get_host() && iteration_running) {
293 comp_timer += activity->surf_action_->get_finish_time() - activity->surf_action_->get_start_time();
296 simgrid::s4u::Link::on_communicate.connect(
297 [this](kernel::resource::NetworkAction* action, s4u::Host* src, s4u::Host* dst) {
298 if ((get_host() == src || get_host() == dst) && iteration_running) {
304 virtual std::string get_name() const override { return "Adagio"; }
308 sg_host_load_reset(get_host());
309 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
311 start_time = simgrid::s4u::Engine::get_clock();
312 if (rates.size() <= task_id)
313 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
314 if (rates[task_id][best_pstate] == 0)
316 get_host()->set_pstate(best_pstate); // Load our schedule
317 XBT_DEBUG("Set pstate to %i", best_pstate);
322 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
323 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
326 static_cast<double>(99.0 / 100.0); // FIXME We account for t_copy arbitrarily with 1% -- this needs to be fixed
328 bool is_initialized = rates[task_id][best_pstate] != 0;
329 rates[task_id][best_pstate] = computed_flops / comp_timer;
330 if (not is_initialized) {
331 for (int i = 1; i < get_host()->get_pstate_count(); i++) {
332 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
334 is_initialized = true;
337 for (int pstate = get_host()->get_pstate_count() - 1; pstate >= 0; pstate--) {
338 if (computed_flops / rates[task_id][pstate] <= target_time) {
339 // We just found the pstate we want to use!
340 best_pstate = pstate;
347 virtual void update() override {}
350 } // namespace plugin
351 } // namespace simgrid
353 /* **************************** events callback *************************** */
354 static void on_host_added(simgrid::s4u::Host& host)
356 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
359 std::string name = std::string("dvfs-daemon-") + host.get_cname();
360 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
362 * This lambda function is the function the actor (daemon) will execute
363 * all the time - in the case of the dvfs plugin, this controls when to
364 * lower/raise the frequency.
366 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
368 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
370 std::string dvfs_governor;
371 const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor");
372 if (host_conf != nullptr) {
373 dvfs_governor = std::string(host_conf);
374 boost::algorithm::to_lower(dvfs_governor);
376 dvfs_governor = cfg_governor;
377 boost::algorithm::to_lower(dvfs_governor);
380 auto governor = [&dvfs_governor, &daemon_proc]() {
381 if (dvfs_governor == "conservative") {
382 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
383 new simgrid::plugin::dvfs::Conservative(daemon_proc->get_host()));
384 } else if (dvfs_governor == "ondemand") {
385 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
386 new simgrid::plugin::dvfs::OnDemand(daemon_proc->get_host()));
387 } else if (dvfs_governor == "adagio") {
388 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
389 new simgrid::plugin::dvfs::Adagio(daemon_proc->get_host()));
390 } else if (dvfs_governor == "performance") {
391 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
392 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
393 } else if (dvfs_governor == "powersave") {
394 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
395 new simgrid::plugin::dvfs::Powersave(daemon_proc->get_host()));
397 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
398 daemon_proc->get_host()->get_cname());
399 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
400 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
405 // Sleep *before* updating; important for startup (i.e., t = 0).
406 // In the beginning, we want to go with the pstates specified in the platform file
407 // (so we sleep first)
408 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
410 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
413 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
417 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
418 // SMPI will take this process as an MPI process!
422 /* **************************** Public interface *************************** */
424 /** @ingroup SURF_plugin_load
425 * @brief Initializes the HostDvfs plugin
426 * @details The HostDvfs plugin provides an API to get the current load of each host.
428 void sg_host_dvfs_plugin_init()
430 static bool inited = false;
435 sg_host_load_plugin_init();
437 simgrid::s4u::Host::on_creation.connect(&on_host_added);