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_;
85 explicit Governor(simgrid::s4u::Host* ptr) : host_(ptr), min_pstate(cfg_min_pstate),
86 max_pstate(cfg_max_pstate == max_pstate_not_limited ? host_->get_pstate_count() - 1 : cfg_max_pstate) {
87 xbt_assert(max_pstate <= host_->get_pstate_count() - 1);
88 xbt_assert(min_pstate <= max_pstate);
89 xbt_assert(0 <= min_pstate);
92 virtual ~Governor() = default;
93 virtual std::string get_name() const = 0;
94 simgrid::s4u::Host* get_host() const { return host_; }
98 const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name());
99 double global_sampling_rate_config = cfg_sampling_rate;
100 if (local_sampling_rate_config != nullptr) {
101 sampling_rate_ = std::stod(local_sampling_rate_config);
103 sampling_rate_ = global_sampling_rate_config;
107 virtual void update() = 0;
108 double get_sampling_rate() const { return sampling_rate_; }
112 * The linux kernel doc describes this governor as follows:
113 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
115 * > The CPUfreq governor "performance" sets the CPU statically to the
116 * > highest frequency within the borders of scaling_min_freq and
117 * > scaling_max_freq.
119 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
121 class Performance : public Governor {
123 explicit Performance(simgrid::s4u::Host* ptr) : Governor(ptr) {}
124 std::string get_name() const override { return "Performance"; }
126 void update() override { get_host()->set_pstate(min_pstate); }
130 * The linux kernel doc describes this governor as follows:
131 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
133 * > The CPUfreq governor "powersave" sets the CPU statically to the
134 * > lowest frequency within the borders of scaling_min_freq and
135 * > scaling_max_freq.
137 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
139 class Powersave : public Governor {
141 explicit Powersave(simgrid::s4u::Host* ptr) : Governor(ptr) {}
142 std::string get_name() const override { return "Powersave"; }
144 void update() override { get_host()->set_pstate(max_pstate); }
148 * The linux kernel doc describes this governor as follows:
149 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
151 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
152 * > current system load. [...] when triggered, cpufreq checks
153 * > the CPU-usage statistics over the last period and the governor sets the
156 class OnDemand : public Governor {
158 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
159 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
161 double freq_up_threshold_ = 0.80;
164 explicit OnDemand(simgrid::s4u::Host* ptr) : Governor(ptr) {}
165 std::string get_name() const override { return "OnDemand"; }
167 void update() override
169 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
170 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
172 if (load > freq_up_threshold_) {
173 get_host()->set_pstate(min_pstate); /* Run at max. performance! */
174 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %i", load, freq_up_threshold_, min_pstate);
176 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
178 * freq_next = min_f + load * (max_f - min_f) / 100
180 * So they assume that frequency increases by 100 MHz. We will just use
181 * lowest_pstate - load*pstatesCount()
183 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
184 // because pstate 0 can only be selected if load > freq_up_threshold_
185 int new_pstate = max_pstate - load * (max_pstate + 1);
186 get_host()->set_pstate(new_pstate);
188 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %i", load, freq_up_threshold_, new_pstate);
195 * This is the conservative governor, which is very similar to the
196 * OnDemand governor. The Linux Kernel Documentation describes it
197 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
199 * > The CPUfreq governor "conservative", much like the "ondemand"
200 * > governor, sets the CPU frequency depending on the current usage. It
201 * > differs in behaviour in that it gracefully increases and decreases the
202 * > CPU speed rather than jumping to max speed the moment there is any load
203 * > on the CPU. This behaviour is more suitable in a battery powered
206 class Conservative : public Governor {
207 double freq_up_threshold_ = .8;
208 double freq_down_threshold_ = .2;
211 explicit Conservative(simgrid::s4u::Host* ptr) : Governor(ptr) {}
212 virtual std::string get_name() const override { return "Conservative"; }
214 virtual void update() override
216 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
217 int pstate = get_host()->get_pstate();
218 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
220 if (load > freq_up_threshold_) {
221 if (pstate != min_pstate) {
222 get_host()->set_pstate(pstate - 1);
223 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %d", load, freq_up_threshold_,
226 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %d", load,
227 freq_up_threshold_, pstate);
229 } else if (load < freq_down_threshold_) {
230 if (pstate != max_pstate) { // Are we in the slowest pstate already?
231 get_host()->set_pstate(pstate + 1);
232 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %d", load, freq_down_threshold_, pstate + 1);
234 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %d", load,
235 freq_down_threshold_, pstate);
241 class Adagio : public Governor {
244 double start_time = 0;
245 double comp_counter = 0;
246 double comp_timer = 0;
248 std::vector<std::vector<double>> rates;
250 unsigned int task_id = 0;
251 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
254 explicit Adagio(simgrid::s4u::Host* ptr)
255 : Governor(ptr), rates(100, std::vector<double>(host_->get_pstate_count(), 0.0))
257 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::ActorPtr actor) {
258 // Every instance of this class subscribes to this event, so one per host
259 // This means that for any actor, all 'hosts' are normally notified of these
260 // changes, even those who don't currently run the actor 'proc_id'.
261 // -> Let's check if this signal call is for us!
262 if (get_host() == actor->get_host()) {
263 iteration_running = true;
266 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::ActorPtr actor) {
267 if (get_host() == actor->get_host()) {
268 iteration_running = false;
272 simgrid::kernel::activity::ExecImpl::on_creation.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
273 if (activity->host_ == get_host())
276 simgrid::kernel::activity::ExecImpl::on_completion.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
277 // For more than one host (not yet supported), we can access the host via
278 // simcalls_.front()->issuer->iface()->get_host()
279 if (activity->host_ == get_host() && iteration_running) {
280 comp_timer += activity->surf_action_->get_finish_time() - activity->surf_action_->get_start_time();
283 simgrid::s4u::Link::on_communicate.connect(
284 [this](kernel::resource::NetworkAction* action, s4u::Host* src, s4u::Host* dst) {
285 if ((get_host() == src || get_host() == dst) && iteration_running) {
291 virtual std::string get_name() const override { return "Adagio"; }
295 sg_host_load_reset(host_);
296 comp_counter = sg_host_get_computed_flops(host_); // Should be 0 because of the reset
298 start_time = simgrid::s4u::Engine::get_clock();
299 if (rates.size() <= task_id)
300 rates.resize(task_id + 5, std::vector<double>(host_->get_pstate_count(), 0.0));
301 if (rates[task_id][best_pstate] == 0)
303 host_->set_pstate(best_pstate); // Load our schedule
304 XBT_DEBUG("Set pstate to %i", best_pstate);
309 double computed_flops = sg_host_get_computed_flops(host_) - comp_counter;
310 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
313 static_cast<double>(99.0 / 100.0); // FIXME We account for t_copy arbitrarily with 1% -- this needs to be fixed
315 bool is_initialized = rates[task_id][best_pstate] != 0;
316 rates[task_id][best_pstate] = computed_flops / comp_timer;
317 if (not is_initialized) {
318 for (int i = 1; i < host_->get_pstate_count(); i++) {
319 rates[task_id][i] = rates[task_id][0] * (host_->get_pstate_speed(i) / host_->get_speed());
321 is_initialized = true;
324 for (int pstate = host_->get_pstate_count() - 1; pstate >= 0; pstate--) {
325 if (computed_flops / rates[task_id][pstate] <= target_time) {
326 // We just found the pstate we want to use!
327 best_pstate = pstate;
334 virtual void update() override {}
337 } // namespace plugin
338 } // namespace simgrid
340 /* **************************** events callback *************************** */
341 static void on_host_added(simgrid::s4u::Host& host)
343 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
346 std::string name = std::string("dvfs-daemon-") + host.get_cname();
347 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
349 * This lambda function is the function the actor (daemon) will execute
350 * all the time - in the case of the dvfs plugin, this controls when to
351 * lower/raise the frequency.
353 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
355 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
357 std::string dvfs_governor;
358 const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor");
359 if (host_conf != nullptr) {
360 dvfs_governor = std::string(host_conf);
361 boost::algorithm::to_lower(dvfs_governor);
363 dvfs_governor = cfg_governor;
364 boost::algorithm::to_lower(dvfs_governor);
367 auto governor = [&dvfs_governor, &daemon_proc]() {
368 if (dvfs_governor == "conservative") {
369 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
370 new simgrid::plugin::dvfs::Conservative(daemon_proc->get_host()));
371 } else if (dvfs_governor == "ondemand") {
372 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
373 new simgrid::plugin::dvfs::OnDemand(daemon_proc->get_host()));
374 } else if (dvfs_governor == "adagio") {
375 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
376 new simgrid::plugin::dvfs::Adagio(daemon_proc->get_host()));
377 } else if (dvfs_governor == "performance") {
378 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
379 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
380 } else if (dvfs_governor == "powersave") {
381 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
382 new simgrid::plugin::dvfs::Powersave(daemon_proc->get_host()));
384 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
385 daemon_proc->get_host()->get_cname());
386 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
387 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
392 // Sleep *before* updating; important for startup (i.e., t = 0).
393 // In the beginning, we want to go with the pstates specified in the platform file
394 // (so we sleep first)
395 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
397 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
400 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
404 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
405 // SMPI will take this process as an MPI process!
409 /* **************************** Public interface *************************** */
411 /** @ingroup SURF_plugin_load
412 * @brief Initializes the HostDvfs plugin
413 * @details The HostDvfs plugin provides an API to get the current load of each host.
415 void sg_host_dvfs_plugin_init()
417 static bool inited = false;
422 sg_host_load_plugin_init();
424 simgrid::s4u::Host::on_creation.connect(&on_host_added);