1 /* Copyright (c) 2010-2023. 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/Actor.hpp>
9 #include <simgrid/s4u/Host.hpp>
10 #include <simgrid/s4u/VirtualMachine.hpp>
11 #include <xbt/asserts.h>
12 #include <xbt/config.hpp>
14 #include "src/internal_config.h" // HAVE_SMPI
15 #include "src/kernel/activity/CommImpl.hpp"
16 #include "src/kernel/resource/NetworkModel.hpp"
18 #include "src/smpi/include/smpi_request.hpp"
19 #include "src/smpi/plugins/ampi/ampi.hpp"
22 #include <boost/algorithm/string.hpp>
23 #include <string_view>
25 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
27 static simgrid::config::Flag<double>
28 cfg_sampling_rate("plugin/dvfs/sampling-rate",
29 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
32 sg_host_dvfs_plugin_init();
35 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
36 "Which Governor should be used that adapts the CPU frequency?",
39 std::map<std::string, std::string, std::less<>>({
41 {"adagio", "TODO: Doc"},
43 {"conservative", "TODO: Doc"}, {"ondemand", "TODO: Doc"},
44 {"performance", "TODO: Doc"}, {"powersave", "TODO: Doc"},
47 [](std::string_view val) {
48 if (val != "performance")
49 sg_host_dvfs_plugin_init();
52 static simgrid::config::Flag<int>
53 cfg_min_pstate("plugin/dvfs/min-pstate",
54 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0);
56 static constexpr int MAX_PSTATE_NOT_LIMITED = -1;
57 static simgrid::config::Flag<int>
58 cfg_max_pstate("plugin/dvfs/max-pstate",
59 "Which pstate is the maximum (and hence slowest) pstate for this governor?", MAX_PSTATE_NOT_LIMITED);
61 /** @addtogroup SURF_plugin_load
63 This plugin makes it very simple for users to obtain the current load for each host.
67 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_dvfs, kernel, "Logging specific to the HostDvfs plugin");
69 namespace simgrid::plugin::dvfs {
72 * Add this to your host tag:
73 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
75 * Valid values as of now are: performance, powersave, ondemand, conservative
76 * It doesn't matter if you use uppercase or lowercase.
78 * For the sampling rate, use this:
80 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
82 * This will run the update() method of the specified governor every 2 seconds
85 * These properties can also be used within the \<config\> tag to configure
86 * these values globally. Using them within the \<host\> will overwrite this
87 * global configuration
90 simgrid::s4u::Host* const host_;
91 double sampling_rate_;
92 unsigned long min_pstate = cfg_min_pstate; //< Never use a pstate less than this one
93 unsigned long max_pstate = cfg_max_pstate; //< Never use a pstate larger than this one
96 explicit Governor(simgrid::s4u::Host* ptr)
99 if (cfg_max_pstate == MAX_PSTATE_NOT_LIMITED)
100 max_pstate = host_->get_pstate_count() - 1;
103 virtual ~Governor() = default;
104 virtual std::string get_name() const = 0;
105 simgrid::s4u::Host* get_host() const { return host_; }
106 unsigned long get_min_pstate() const { return min_pstate; }
107 unsigned long get_max_pstate() const { return max_pstate; }
111 if (const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name())) {
112 sampling_rate_ = std::stod(local_sampling_rate_config);
114 sampling_rate_ = cfg_sampling_rate;
116 if (const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name())) {
117 min_pstate = std::stoul(local_min_pstate_config);
120 if (const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name())) {
121 max_pstate = std::stoul(local_max_pstate_config);
123 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
124 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
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 using Governor::Governor;
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 using Governor::Governor;
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 using Governor::Governor;
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 %lu", 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 auto new_pstate = get_max_pstate() - static_cast<unsigned long>(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 %lu", 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 behavior 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 behavior 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 using Governor::Governor;
233 std::string get_name() const override { return "Conservative"; }
235 void update() override
237 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
238 unsigned long 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 %lu", load, freq_up_threshold_,
247 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %lu", 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 %lu", load, freq_down_threshold_, pstate + 1);
255 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %lu", load,
256 freq_down_threshold_, pstate);
263 class Adagio : public Governor {
264 unsigned long best_pstate = 0;
265 double start_time = 0;
266 double comp_counter = 0;
267 double comp_timer = 0;
269 std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
271 unsigned int task_id = 0;
272 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
275 explicit Adagio(simgrid::s4u::Host* ptr)
276 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
278 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::Actor const& actor) {
279 // Every instance of this class subscribes to this event, so one per host
280 // This means that for any actor, all 'hosts' are normally notified of these
281 // changes, even those who don't currently run the actor 'proc_id'.
282 // -> Let's check if this signal call is for us!
283 if (get_host() == actor.get_host()) {
284 iteration_running = true;
287 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::Actor const& actor) {
288 if (get_host() == actor.get_host()) {
289 iteration_running = false;
293 simgrid::s4u::Exec::on_start_cb([this](simgrid::s4u::Exec const& activity) {
294 if (activity.get_host() == get_host())
297 simgrid::s4u::Activity::on_completion_cb([this](simgrid::s4u::Activity const& activity) {
298 const auto* exec = dynamic_cast<simgrid::s4u::Exec const*>(&activity);
299 if (exec == nullptr) // Only Execs are concerned here
301 // For more than one host (not yet supported), we can access the host via
302 // simcalls_.front()->issuer->get_iface()->get_host()
303 if (exec->get_host() == get_host() && iteration_running) {
304 comp_timer += exec->get_finish_time() - exec->get_start_time();
307 // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
308 // the dst will be notified even though it didn't even arrive at the recv yet
309 kernel::activity::CommImpl::on_start.connect([this](const kernel::activity::CommImpl& comm) {
310 const auto* act = static_cast<kernel::resource::NetworkAction*>(comm.model_action_);
311 if ((get_host() == &act->get_src() || get_host() == &act->get_dst()) && iteration_running) {
317 std::string get_name() const override { return "Adagio"; }
321 sg_host_load_reset(get_host());
322 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
324 start_time = simgrid::s4u::Engine::get_clock();
325 if (rates.size() <= task_id)
326 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
327 if (rates[task_id][best_pstate] == 0)
329 get_host()->set_pstate(best_pstate); // Load our schedule
330 XBT_DEBUG("Set pstate to %lu", best_pstate);
335 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
336 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
337 target_time = target_time * 99.0 / 100.0; // FIXME We account for t_copy arbitrarily with 1%
338 // -- this needs to be fixed
340 bool is_initialized = rates[task_id][best_pstate] != 0;
341 rates[task_id][best_pstate] = computed_flops / comp_timer;
342 if (not is_initialized) {
343 for (unsigned long i = 1; i < get_host()->get_pstate_count(); i++) {
344 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
348 for (unsigned long pstate = get_host()->get_pstate_count() - 1; pstate != 0; pstate--) {
349 if (computed_flops / rates[task_id][pstate] <= target_time) {
350 // We just found the pstate we want to use!
351 best_pstate = pstate;
358 void update() override {}
361 } // namespace simgrid::plugin::dvfs
363 /* **************************** events callback *************************** */
364 static void on_host_added(simgrid::s4u::Host& host)
366 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
369 std::string name = "dvfs-daemon-" + host.get_name();
370 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
372 * This lambda function is the function the actor (daemon) will execute
373 * all the time - in the case of the dvfs plugin, this controls when to
374 * lower/raise the frequency.
376 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
378 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
380 std::string dvfs_governor;
381 if (const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor")) {
382 dvfs_governor = host_conf;
383 boost::algorithm::to_lower(dvfs_governor);
385 dvfs_governor = cfg_governor;
386 boost::algorithm::to_lower(dvfs_governor);
389 auto governor = [&dvfs_governor, &daemon_proc]() -> std::unique_ptr<simgrid::plugin::dvfs::Governor> {
390 if (dvfs_governor == "conservative")
391 return std::make_unique<simgrid::plugin::dvfs::Conservative>(daemon_proc->get_host());
392 if (dvfs_governor == "ondemand")
393 return std::make_unique<simgrid::plugin::dvfs::OnDemand>(daemon_proc->get_host());
395 if (dvfs_governor == "adagio")
396 return std::make_unique<simgrid::plugin::dvfs::Adagio>(daemon_proc->get_host());
398 if (dvfs_governor == "powersave")
399 return std::make_unique<simgrid::plugin::dvfs::Powersave>(daemon_proc->get_host());
400 if (dvfs_governor != "performance")
401 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
402 daemon_proc->get_host()->get_cname());
403 return std::make_unique<simgrid::plugin::dvfs::Performance>(daemon_proc->get_host());
407 // Sleep *before* updating; important for startup (i.e., t = 0).
408 // In the beginning, we want to go with the pstates specified in the platform file
409 // (so we sleep first)
410 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
412 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
415 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
419 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
420 // SMPI will take this process as an MPI process!
424 /* **************************** Public interface *************************** */
427 * @brief Initializes the HostDvfs plugin
428 * @details The HostDvfs plugin provides an API to get the current load of each host.
430 void sg_host_dvfs_plugin_init()
432 static bool inited = false;
437 sg_host_load_plugin_init();
439 simgrid::s4u::Host::on_creation_cb(&on_host_added);