1 /* Copyright (c) 2010-2019. 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/internal_config.h" // HAVE_SMPI
10 #include "src/kernel/activity/ExecImpl.hpp"
11 #include "src/plugins/vm/VirtualMachineImpl.hpp"
13 #include "src/smpi/plugins/ampi/ampi.hpp"
15 #include <xbt/config.hpp>
17 #include <boost/algorithm/string.hpp>
19 #include "src/smpi/include/smpi_request.hpp"
22 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
24 static simgrid::config::Flag<double> cfg_sampling_rate("plugin/dvfs/sampling-rate", {"plugin/dvfs/sampling_rate"},
25 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
26 [](double val){if (val != 0.1) sg_host_dvfs_plugin_init();});
28 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
29 "Which Governor should be used that adapts the CPU frequency?", "performance",
31 std::map<std::string, std::string>({
33 {"adagio", "TODO: Doc"},
35 {"conservative", "TODO: Doc"},
36 {"ondemand", "TODO: Doc"},
37 {"performance", "TODO: Doc"},
38 {"powersave", "TODO: Doc"},
41 [](std::string val) { if (val != "performance") sg_host_dvfs_plugin_init(); });
43 static simgrid::config::Flag<int> cfg_min_pstate("plugin/dvfs/min-pstate", {"plugin/dvfs/min_pstate"},
44 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0,
47 static const int max_pstate_not_limited = -1;
48 static simgrid::config::Flag<int> cfg_max_pstate("plugin/dvfs/max-pstate", {"plugin/dvfs/max_pstate"},
49 "Which pstate is the maximum (and hence slowest) pstate for this governor?", max_pstate_not_limited,
52 /** @addtogroup SURF_plugin_load
54 This plugin makes it very simple for users to obtain the current load for each host.
58 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_plugin_dvfs, surf, "Logging specific to the SURF HostDvfs plugin");
66 * Add this to your host tag:
67 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
69 * Valid values as of now are: performance, powersave, ondemand, conservative
70 * It doesn't matter if you use uppercase or lowercase.
72 * For the sampling rate, use this:
74 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
76 * This will run the update() method of the specified governor every 2 seconds
79 * These properties can also be used within the \<config\> tag to configure
80 * these values globally. Using them within the \<host\> will overwrite this
81 * global configuration
86 simgrid::s4u::Host* const host_;
87 double sampling_rate_;
88 int min_pstate; //< Never use a pstate less than this one
89 int max_pstate; //< Never use a pstate larger than this one
92 explicit Governor(simgrid::s4u::Host* ptr)
94 , min_pstate(cfg_min_pstate)
95 , max_pstate(cfg_max_pstate == max_pstate_not_limited ? host_->get_pstate_count() - 1 : cfg_max_pstate)
99 virtual ~Governor() = default;
100 virtual std::string get_name() const = 0;
101 simgrid::s4u::Host* get_host() const { return host_; }
102 int get_min_pstate() const { return min_pstate; }
103 int get_max_pstate() const { return max_pstate; }
107 const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name());
108 if (local_sampling_rate_config != nullptr) {
109 sampling_rate_ = std::stod(local_sampling_rate_config);
111 sampling_rate_ = cfg_sampling_rate;
113 const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name());
114 if (local_min_pstate_config != nullptr) {
115 min_pstate = std::stoi(local_min_pstate_config);
118 const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name());
119 if (local_max_pstate_config != nullptr) {
120 max_pstate = std::stod(local_max_pstate_config);
122 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
123 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
124 xbt_assert(0 <= min_pstate, "min_pstate is negative!");
127 virtual void update() = 0;
128 double get_sampling_rate() const { return sampling_rate_; }
132 * The linux kernel doc describes this governor as follows:
133 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
135 * > The CPUfreq governor "performance" sets the CPU statically to the
136 * > highest frequency within the borders of scaling_min_freq and
137 * > scaling_max_freq.
139 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
141 class Performance : public Governor {
143 explicit Performance(simgrid::s4u::Host* ptr) : Governor(ptr) {}
144 std::string get_name() const override { return "Performance"; }
146 void update() override { get_host()->set_pstate(get_min_pstate()); }
150 * The linux kernel doc describes this governor as follows:
151 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
153 * > The CPUfreq governor "powersave" sets the CPU statically to the
154 * > lowest frequency within the borders of scaling_min_freq and
155 * > scaling_max_freq.
157 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
159 class Powersave : public Governor {
161 explicit Powersave(simgrid::s4u::Host* ptr) : Governor(ptr) {}
162 std::string get_name() const override { return "Powersave"; }
164 void update() override { get_host()->set_pstate(get_max_pstate()); }
168 * The linux kernel doc describes this governor as follows:
169 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
171 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
172 * > current system load. [...] when triggered, cpufreq checks
173 * > the CPU-usage statistics over the last period and the governor sets the
176 class OnDemand : public Governor {
178 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
179 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
181 double freq_up_threshold_ = 0.80;
184 explicit OnDemand(simgrid::s4u::Host* ptr) : Governor(ptr) {}
185 std::string get_name() const override { return "OnDemand"; }
187 void update() override
189 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
190 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
192 if (load > freq_up_threshold_) {
193 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
194 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %i", load, freq_up_threshold_, get_min_pstate());
196 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
198 * freq_next = min_f + load * (max_f - min_f) / 100
200 * So they assume that frequency increases by 100 MHz. We will just use
201 * lowest_pstate - load*pstatesCount()
203 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
204 // because pstate 0 can only be selected if load > freq_up_threshold_
205 int new_pstate = get_max_pstate() - load * (get_max_pstate() + 1);
206 if (new_pstate < get_min_pstate())
207 new_pstate = get_min_pstate();
208 get_host()->set_pstate(new_pstate);
210 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %i", load, freq_up_threshold_, new_pstate);
216 * This is the conservative governor, which is very similar to the
217 * OnDemand governor. The Linux Kernel Documentation describes it
218 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
220 * > The CPUfreq governor "conservative", much like the "ondemand"
221 * > governor, sets the CPU frequency depending on the current usage. It
222 * > differs in behaviour in that it gracefully increases and decreases the
223 * > CPU speed rather than jumping to max speed the moment there is any load
224 * > on the CPU. This behaviour is more suitable in a battery powered
227 class Conservative : public Governor {
228 double freq_up_threshold_ = .8;
229 double freq_down_threshold_ = .2;
232 explicit Conservative(simgrid::s4u::Host* ptr) : Governor(ptr) {}
233 virtual std::string get_name() const override { return "Conservative"; }
235 virtual void update() override
237 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
238 int pstate = get_host()->get_pstate();
239 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
241 if (load > freq_up_threshold_) {
242 if (pstate != get_min_pstate()) {
243 get_host()->set_pstate(pstate - 1);
244 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %d", load, freq_up_threshold_,
247 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %d", load,
248 freq_up_threshold_, pstate);
250 } else if (load < freq_down_threshold_) {
251 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
252 get_host()->set_pstate(pstate + 1);
253 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %d", load, freq_down_threshold_, pstate + 1);
255 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %d", load,
256 freq_down_threshold_, pstate);
263 class Adagio : public Governor {
266 double start_time = 0;
267 double comp_counter = 0;
268 double comp_timer = 0;
270 std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
272 unsigned int task_id = 0;
273 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
276 explicit Adagio(simgrid::s4u::Host* ptr)
277 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
279 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::ActorPtr actor) {
280 // Every instance of this class subscribes to this event, so one per host
281 // This means that for any actor, all 'hosts' are normally notified of these
282 // changes, even those who don't currently run the actor 'proc_id'.
283 // -> Let's check if this signal call is for us!
284 if (get_host() == actor->get_host()) {
285 iteration_running = true;
288 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::ActorPtr actor) {
289 if (get_host() == actor->get_host()) {
290 iteration_running = false;
294 simgrid::kernel::activity::ExecImpl::on_creation.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
295 if (activity->host_ == get_host())
298 simgrid::kernel::activity::ExecImpl::on_completion.connect([this](simgrid::kernel::activity::ExecImplPtr activity) {
299 // For more than one host (not yet supported), we can access the host via
300 // simcalls_.front()->issuer->iface()->get_host()
301 if (activity->host_ == get_host() && iteration_running) {
302 comp_timer += activity->surf_action_->get_finish_time() - activity->surf_action_->get_start_time();
305 // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
306 // the dst will be notified even though it didn't even arrive at the recv yet
307 simgrid::s4u::Link::on_communicate.connect(
308 [this](kernel::resource::NetworkAction* action, s4u::Host* src, s4u::Host* dst) {
309 if ((get_host() == src || get_host() == dst) && iteration_running) {
315 virtual std::string get_name() const override { return "Adagio"; }
319 sg_host_load_reset(get_host());
320 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
322 start_time = simgrid::s4u::Engine::get_clock();
323 if (rates.size() <= task_id)
324 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
325 if (rates[task_id][best_pstate] == 0)
327 get_host()->set_pstate(best_pstate); // Load our schedule
328 XBT_DEBUG("Set pstate to %i", best_pstate);
333 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
334 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
337 static_cast<double>(99.0 / 100.0); // FIXME We account for t_copy arbitrarily with 1% -- this needs to be fixed
339 bool is_initialized = rates[task_id][best_pstate] != 0;
340 rates[task_id][best_pstate] = computed_flops / comp_timer;
341 if (not is_initialized) {
342 for (int i = 1; i < get_host()->get_pstate_count(); i++) {
343 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
347 for (int pstate = get_host()->get_pstate_count() - 1; pstate >= 0; pstate--) {
348 if (computed_flops / rates[task_id][pstate] <= target_time) {
349 // We just found the pstate we want to use!
350 best_pstate = pstate;
357 virtual void update() override {}
361 } // namespace plugin
362 } // namespace simgrid
364 /* **************************** events callback *************************** */
365 static void on_host_added(simgrid::s4u::Host& host)
367 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
370 std::string name = std::string("dvfs-daemon-") + host.get_cname();
371 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
373 * This lambda function is the function the actor (daemon) will execute
374 * all the time - in the case of the dvfs plugin, this controls when to
375 * lower/raise the frequency.
377 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
379 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
381 std::string dvfs_governor;
382 const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor");
383 if (host_conf != nullptr) {
384 dvfs_governor = std::string(host_conf);
385 boost::algorithm::to_lower(dvfs_governor);
387 dvfs_governor = cfg_governor;
388 boost::algorithm::to_lower(dvfs_governor);
391 auto governor = [&dvfs_governor, &daemon_proc]() {
392 if (dvfs_governor == "conservative") {
393 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
394 new simgrid::plugin::dvfs::Conservative(daemon_proc->get_host()));
395 } else if (dvfs_governor == "ondemand") {
396 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
397 new simgrid::plugin::dvfs::OnDemand(daemon_proc->get_host()));
400 else if (dvfs_governor == "adagio") {
401 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
402 new simgrid::plugin::dvfs::Adagio(daemon_proc->get_host()));
405 else if (dvfs_governor == "performance") {
406 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
407 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
408 } else if (dvfs_governor == "powersave") {
409 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
410 new simgrid::plugin::dvfs::Powersave(daemon_proc->get_host()));
412 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
413 daemon_proc->get_host()->get_cname());
414 return std::unique_ptr<simgrid::plugin::dvfs::Governor>(
415 new simgrid::plugin::dvfs::Performance(daemon_proc->get_host()));
420 // Sleep *before* updating; important for startup (i.e., t = 0).
421 // In the beginning, we want to go with the pstates specified in the platform file
422 // (so we sleep first)
423 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
425 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
428 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
432 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
433 // SMPI will take this process as an MPI process!
437 /* **************************** Public interface *************************** */
439 /** @ingroup SURF_plugin_load
440 * @brief Initializes the HostDvfs plugin
441 * @details The HostDvfs plugin provides an API to get the current load of each host.
443 void sg_host_dvfs_plugin_init()
445 static bool inited = false;
450 sg_host_load_plugin_init();
452 simgrid::s4u::Host::on_creation.connect(&on_host_added);