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"
17 #include "src/simgrid/module.hpp"
19 #include "src/smpi/include/smpi_request.hpp"
20 #include "src/smpi/plugins/ampi/ampi.hpp"
23 #include <boost/algorithm/string.hpp>
24 #include <string_view>
26 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
28 static simgrid::config::Flag<double>
29 cfg_sampling_rate("plugin/dvfs/sampling-rate",
30 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
33 sg_host_dvfs_plugin_init();
36 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
37 "Which Governor should be used that adapts the CPU frequency?",
40 std::map<std::string, std::string, std::less<>>({
42 {"adagio", "TODO: Doc"},
44 {"conservative", "TODO: Doc"}, {"ondemand", "TODO: Doc"},
45 {"performance", "TODO: Doc"}, {"powersave", "TODO: Doc"},
48 [](std::string_view val) {
49 if (val != "performance")
50 sg_host_dvfs_plugin_init();
53 static simgrid::config::Flag<int>
54 cfg_min_pstate("plugin/dvfs/min-pstate",
55 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0);
57 static constexpr int MAX_PSTATE_NOT_LIMITED = -1;
58 static simgrid::config::Flag<int>
59 cfg_max_pstate("plugin/dvfs/max-pstate",
60 "Which pstate is the maximum (and hence slowest) pstate for this governor?", MAX_PSTATE_NOT_LIMITED);
62 /** @addtogroup SURF_plugin_load
64 This plugin makes it very simple for users to obtain the current load for each host.
68 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_dvfs, kernel, "Logging specific to the HostDvfs plugin");
70 namespace simgrid::plugin::dvfs {
73 * Add this to your host tag:
74 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
76 * Valid values as of now are: performance, powersave, ondemand, conservative
77 * It doesn't matter if you use uppercase or lowercase.
79 * For the sampling rate, use this:
81 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
83 * This will run the update() method of the specified governor every 2 seconds
86 * These properties can also be used within the \<config\> tag to configure
87 * these values globally. Using them within the \<host\> will overwrite this
88 * global configuration
91 simgrid::s4u::Host* const host_;
92 double sampling_rate_;
93 unsigned long min_pstate = cfg_min_pstate; //< Never use a pstate less than this one
94 unsigned long max_pstate = cfg_max_pstate; //< Never use a pstate larger than this one
97 explicit Governor(simgrid::s4u::Host* ptr)
100 if (cfg_max_pstate == MAX_PSTATE_NOT_LIMITED)
101 max_pstate = host_->get_pstate_count() - 1;
104 virtual ~Governor() = default;
105 virtual std::string get_name() const = 0;
106 simgrid::s4u::Host* get_host() const { return host_; }
107 unsigned long get_min_pstate() const { return min_pstate; }
108 unsigned long get_max_pstate() const { return max_pstate; }
112 if (const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name())) {
113 sampling_rate_ = std::stod(local_sampling_rate_config);
115 sampling_rate_ = cfg_sampling_rate;
117 if (const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name())) {
118 min_pstate = std::stoul(local_min_pstate_config);
121 if (const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name())) {
122 max_pstate = std::stoul(local_max_pstate_config);
124 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
125 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
128 virtual void update() = 0;
129 double get_sampling_rate() const { return sampling_rate_; }
133 * The linux kernel doc describes this governor as follows:
134 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
136 * > The CPUfreq governor "performance" sets the CPU statically to the
137 * > highest frequency within the borders of scaling_min_freq and
138 * > scaling_max_freq.
140 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
142 class Performance : public Governor {
144 using Governor::Governor;
145 std::string get_name() const override { return "Performance"; }
147 void update() override { get_host()->set_pstate(get_min_pstate()); }
151 * The linux kernel doc describes this governor as follows:
152 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
154 * > The CPUfreq governor "powersave" sets the CPU statically to the
155 * > lowest frequency within the borders of scaling_min_freq and
156 * > scaling_max_freq.
158 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
160 class Powersave : public Governor {
162 using Governor::Governor;
163 std::string get_name() const override { return "Powersave"; }
165 void update() override { get_host()->set_pstate(get_max_pstate()); }
169 * The linux kernel doc describes this governor as follows:
170 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
172 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
173 * > current system load. [...] when triggered, cpufreq checks
174 * > the CPU-usage statistics over the last period and the governor sets the
177 class OnDemand : public Governor {
179 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
180 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
182 double freq_up_threshold_ = 0.80;
185 using Governor::Governor;
186 std::string get_name() const override { return "OnDemand"; }
188 void update() override
190 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
191 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
193 if (load > freq_up_threshold_) {
194 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
195 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, get_min_pstate());
197 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
199 * freq_next = min_f + load * (max_f - min_f) / 100
201 * So they assume that frequency increases by 100 MHz. We will just use
202 * lowest_pstate - load*pstatesCount()
204 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
205 // because pstate 0 can only be selected if load > freq_up_threshold_
206 auto new_pstate = get_max_pstate() - static_cast<unsigned long>(load) * (get_max_pstate() + 1);
207 if (new_pstate < get_min_pstate())
208 new_pstate = get_min_pstate();
209 get_host()->set_pstate(new_pstate);
211 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, new_pstate);
217 * This is the conservative governor, which is very similar to the
218 * OnDemand governor. The Linux Kernel Documentation describes it
219 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
221 * > The CPUfreq governor "conservative", much like the "ondemand"
222 * > governor, sets the CPU frequency depending on the current usage. It
223 * > differs in behavior in that it gracefully increases and decreases the
224 * > CPU speed rather than jumping to max speed the moment there is any load
225 * > on the CPU. This behavior is more suitable in a battery powered
228 class Conservative : public Governor {
229 double freq_up_threshold_ = .8;
230 double freq_down_threshold_ = .2;
233 using Governor::Governor;
234 std::string get_name() const override { return "Conservative"; }
236 void update() override
238 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
239 unsigned long pstate = get_host()->get_pstate();
240 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
242 if (load > freq_up_threshold_) {
243 if (pstate != get_min_pstate()) {
244 get_host()->set_pstate(pstate - 1);
245 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %lu", load, freq_up_threshold_,
248 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %lu", load,
249 freq_up_threshold_, pstate);
251 } else if (load < freq_down_threshold_) {
252 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
253 get_host()->set_pstate(pstate + 1);
254 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %lu", load, freq_down_threshold_, pstate + 1);
256 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %lu", load,
257 freq_down_threshold_, pstate);
264 class Adagio : public Governor {
265 unsigned long best_pstate = 0;
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::Actor const& 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::Actor const& actor) {
289 if (get_host() == actor.get_host()) {
290 iteration_running = false;
294 simgrid::s4u::Exec::on_start_cb([this](simgrid::s4u::Exec const& activity) {
295 if (activity.get_host() == get_host())
298 simgrid::s4u::Activity::on_completion_cb([this](simgrid::s4u::Activity const& activity) {
299 const auto* exec = dynamic_cast<simgrid::s4u::Exec const*>(&activity);
300 if (exec == nullptr) // Only Execs are concerned here
302 // For more than one host (not yet supported), we can access the host via
303 // simcalls_.front()->issuer->get_iface()->get_host()
304 if (exec->get_host() == get_host() && iteration_running) {
305 comp_timer += exec->get_finish_time() - exec->get_start_time();
308 // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
309 // the dst will be notified even though it didn't even arrive at the recv yet
310 kernel::activity::CommImpl::on_start.connect([this](const kernel::activity::CommImpl& comm) {
311 const auto* act = static_cast<kernel::resource::NetworkAction*>(comm.model_action_);
312 if ((get_host() == &act->get_src() || get_host() == &act->get_dst()) && iteration_running) {
318 std::string get_name() const override { return "Adagio"; }
322 sg_host_load_reset(get_host());
323 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
325 start_time = simgrid::s4u::Engine::get_clock();
326 if (rates.size() <= task_id)
327 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
328 if (rates[task_id][best_pstate] == 0)
330 get_host()->set_pstate(best_pstate); // Load our schedule
331 XBT_DEBUG("Set pstate to %lu", best_pstate);
336 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
337 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
338 target_time = target_time * 99.0 / 100.0; // FIXME We account for t_copy arbitrarily with 1%
339 // -- this needs to be fixed
341 bool is_initialized = rates[task_id][best_pstate] != 0;
342 rates[task_id][best_pstate] = computed_flops / comp_timer;
343 if (not is_initialized) {
344 for (unsigned long i = 1; i < get_host()->get_pstate_count(); i++) {
345 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
349 for (unsigned long pstate = get_host()->get_pstate_count() - 1; pstate != 0; pstate--) {
350 if (computed_flops / rates[task_id][pstate] <= target_time) {
351 // We just found the pstate we want to use!
352 best_pstate = pstate;
359 void update() override {}
362 } // namespace simgrid::plugin::dvfs
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 = "dvfs-daemon-" + host.get_name();
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 if (const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor")) {
383 dvfs_governor = host_conf;
384 boost::algorithm::to_lower(dvfs_governor);
386 dvfs_governor = cfg_governor;
387 boost::algorithm::to_lower(dvfs_governor);
390 auto governor = [&dvfs_governor, &daemon_proc]() -> std::unique_ptr<simgrid::plugin::dvfs::Governor> {
391 if (dvfs_governor == "conservative")
392 return std::make_unique<simgrid::plugin::dvfs::Conservative>(daemon_proc->get_host());
393 if (dvfs_governor == "ondemand")
394 return std::make_unique<simgrid::plugin::dvfs::OnDemand>(daemon_proc->get_host());
396 if (dvfs_governor == "adagio")
397 return std::make_unique<simgrid::plugin::dvfs::Adagio>(daemon_proc->get_host());
399 if (dvfs_governor == "powersave")
400 return std::make_unique<simgrid::plugin::dvfs::Powersave>(daemon_proc->get_host());
401 if (dvfs_governor != "performance")
402 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
403 daemon_proc->get_host()->get_cname());
404 return std::make_unique<simgrid::plugin::dvfs::Performance>(daemon_proc->get_host());
408 // Sleep *before* updating; important for startup (i.e., t = 0).
409 // In the beginning, we want to go with the pstates specified in the platform file
410 // (so we sleep first)
411 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
413 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
416 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
420 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
421 // SMPI will take this process as an MPI process!
425 /* **************************** Public interface *************************** */
428 * @brief Initializes the HostDvfs plugin
429 * @details The HostDvfs plugin provides an API to get the current load of each host.
431 void sg_host_dvfs_plugin_init()
433 static bool inited = false;
438 sg_host_load_plugin_init();
440 simgrid::s4u::Host::on_creation_cb(&on_host_added);