1 /* Copyright (c) 2010-2022. 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>
24 SIMGRID_REGISTER_PLUGIN(host_dvfs, "Dvfs support", &sg_host_dvfs_plugin_init)
26 static simgrid::config::Flag<double>
27 cfg_sampling_rate("plugin/dvfs/sampling-rate",
28 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
31 sg_host_dvfs_plugin_init();
34 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
35 "Which Governor should be used that adapts the CPU frequency?",
38 std::map<std::string, std::string, std::less<>>({
40 {"adagio", "TODO: Doc"},
42 {"conservative", "TODO: Doc"}, {"ondemand", "TODO: Doc"},
43 {"performance", "TODO: Doc"}, {"powersave", "TODO: Doc"},
46 [](const std::string& val) {
47 if (val != "performance")
48 sg_host_dvfs_plugin_init();
51 static simgrid::config::Flag<int>
52 cfg_min_pstate("plugin/dvfs/min-pstate",
53 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0);
55 static constexpr int MAX_PSTATE_NOT_LIMITED = -1;
56 static simgrid::config::Flag<int>
57 cfg_max_pstate("plugin/dvfs/max-pstate",
58 "Which pstate is the maximum (and hence slowest) pstate for this governor?", MAX_PSTATE_NOT_LIMITED);
60 /** @addtogroup SURF_plugin_load
62 This plugin makes it very simple for users to obtain the current load for each host.
66 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_dvfs, kernel, "Logging specific to the HostDvfs plugin");
74 * Add this to your host tag:
75 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
77 * Valid values as of now are: performance, powersave, ondemand, conservative
78 * It doesn't matter if you use uppercase or lowercase.
80 * For the sampling rate, use this:
82 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
84 * This will run the update() method of the specified governor every 2 seconds
87 * These properties can also be used within the \<config\> tag to configure
88 * these values globally. Using them within the \<host\> will overwrite this
89 * global configuration
92 simgrid::s4u::Host* const host_;
93 double sampling_rate_;
94 unsigned long min_pstate = cfg_min_pstate; //< Never use a pstate less than this one
95 unsigned long max_pstate = cfg_max_pstate; //< Never use a pstate larger than this one
98 explicit Governor(simgrid::s4u::Host* ptr)
101 if (cfg_max_pstate == MAX_PSTATE_NOT_LIMITED)
102 max_pstate = host_->get_pstate_count() - 1;
105 virtual ~Governor() = default;
106 virtual std::string get_name() const = 0;
107 simgrid::s4u::Host* get_host() const { return host_; }
108 unsigned long get_min_pstate() const { return min_pstate; }
109 unsigned long get_max_pstate() const { return max_pstate; }
113 if (const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name())) {
114 sampling_rate_ = std::stod(local_sampling_rate_config);
116 sampling_rate_ = cfg_sampling_rate;
118 if (const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name())) {
119 min_pstate = std::stoul(local_min_pstate_config);
122 if (const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name())) {
123 max_pstate = std::stoul(local_max_pstate_config);
125 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
126 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
129 virtual void update() = 0;
130 double get_sampling_rate() const { return sampling_rate_; }
134 * The linux kernel doc describes this governor as follows:
135 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
137 * > The CPUfreq governor "performance" sets the CPU statically to the
138 * > highest frequency within the borders of scaling_min_freq and
139 * > scaling_max_freq.
141 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
143 class Performance : public Governor {
145 using Governor::Governor;
146 std::string get_name() const override { return "Performance"; }
148 void update() override { get_host()->set_pstate(get_min_pstate()); }
152 * The linux kernel doc describes this governor as follows:
153 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
155 * > The CPUfreq governor "powersave" sets the CPU statically to the
156 * > lowest frequency within the borders of scaling_min_freq and
157 * > scaling_max_freq.
159 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
161 class Powersave : public Governor {
163 using Governor::Governor;
164 std::string get_name() const override { return "Powersave"; }
166 void update() override { get_host()->set_pstate(get_max_pstate()); }
170 * The linux kernel doc describes this governor as follows:
171 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
173 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
174 * > current system load. [...] when triggered, cpufreq checks
175 * > the CPU-usage statistics over the last period and the governor sets the
178 class OnDemand : public Governor {
180 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
181 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
183 double freq_up_threshold_ = 0.80;
186 using Governor::Governor;
187 std::string get_name() const override { return "OnDemand"; }
189 void update() override
191 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
192 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
194 if (load > freq_up_threshold_) {
195 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
196 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, get_min_pstate());
198 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
200 * freq_next = min_f + load * (max_f - min_f) / 100
202 * So they assume that frequency increases by 100 MHz. We will just use
203 * lowest_pstate - load*pstatesCount()
205 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
206 // because pstate 0 can only be selected if load > freq_up_threshold_
207 auto new_pstate = get_max_pstate() - static_cast<unsigned long>(load) * (get_max_pstate() + 1);
208 if (new_pstate < get_min_pstate())
209 new_pstate = get_min_pstate();
210 get_host()->set_pstate(new_pstate);
212 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, new_pstate);
218 * This is the conservative governor, which is very similar to the
219 * OnDemand governor. The Linux Kernel Documentation describes it
220 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
222 * > The CPUfreq governor "conservative", much like the "ondemand"
223 * > governor, sets the CPU frequency depending on the current usage. It
224 * > differs in behavior in that it gracefully increases and decreases the
225 * > CPU speed rather than jumping to max speed the moment there is any load
226 * > on the CPU. This behavior is more suitable in a battery powered
229 class Conservative : public Governor {
230 double freq_up_threshold_ = .8;
231 double freq_down_threshold_ = .2;
234 using Governor::Governor;
235 std::string get_name() const override { return "Conservative"; }
237 void update() override
239 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
240 unsigned long pstate = get_host()->get_pstate();
241 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
243 if (load > freq_up_threshold_) {
244 if (pstate != get_min_pstate()) {
245 get_host()->set_pstate(pstate - 1);
246 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %lu", load, freq_up_threshold_,
249 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %lu", load,
250 freq_up_threshold_, pstate);
252 } else if (load < freq_down_threshold_) {
253 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
254 get_host()->set_pstate(pstate + 1);
255 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %lu", load, freq_down_threshold_, pstate + 1);
257 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %lu", load,
258 freq_down_threshold_, pstate);
265 class Adagio : public Governor {
266 unsigned long best_pstate = 0;
267 double start_time = 0;
268 double comp_counter = 0;
269 double comp_timer = 0;
271 std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
273 unsigned int task_id = 0;
274 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
277 explicit Adagio(simgrid::s4u::Host* ptr)
278 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
280 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::Actor const& actor) {
281 // Every instance of this class subscribes to this event, so one per host
282 // This means that for any actor, all 'hosts' are normally notified of these
283 // changes, even those who don't currently run the actor 'proc_id'.
284 // -> Let's check if this signal call is for us!
285 if (get_host() == actor.get_host()) {
286 iteration_running = true;
289 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::Actor const& actor) {
290 if (get_host() == actor.get_host()) {
291 iteration_running = false;
295 simgrid::s4u::Exec::on_start_cb([this](simgrid::s4u::Exec const& activity) {
296 if (activity.get_host() == get_host())
299 simgrid::s4u::Activity::on_completion_cb([this](simgrid::s4u::Activity const& activity) {
300 const auto* exec = dynamic_cast<simgrid::s4u::Exec const*>(&activity);
301 if (exec == nullptr) // Only Execs are concerned here
303 // For more than one host (not yet supported), we can access the host via
304 // simcalls_.front()->issuer->get_iface()->get_host()
305 if (exec->get_host() == get_host() && iteration_running) {
306 comp_timer += exec->get_finish_time() - exec->get_start_time();
309 // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
310 // the dst will be notified even though it didn't even arrive at the recv yet
311 kernel::activity::CommImpl::on_start.connect([this](const kernel::activity::CommImpl& comm) {
312 const auto* act = static_cast<kernel::resource::NetworkAction*>(comm.surf_action_);
313 if ((get_host() == &act->get_src() || get_host() == &act->get_dst()) && iteration_running) {
319 std::string get_name() const override { return "Adagio"; }
323 sg_host_load_reset(get_host());
324 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
326 start_time = simgrid::s4u::Engine::get_clock();
327 if (rates.size() <= task_id)
328 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
329 if (rates[task_id][best_pstate] == 0)
331 get_host()->set_pstate(best_pstate); // Load our schedule
332 XBT_DEBUG("Set pstate to %lu", best_pstate);
337 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
338 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
339 target_time = target_time * 99.0 / 100.0; // FIXME We account for t_copy arbitrarily with 1%
340 // -- this needs to be fixed
342 bool is_initialized = rates[task_id][best_pstate] != 0;
343 rates[task_id][best_pstate] = computed_flops / comp_timer;
344 if (not is_initialized) {
345 for (unsigned long i = 1; i < get_host()->get_pstate_count(); i++) {
346 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
350 for (unsigned long pstate = get_host()->get_pstate_count() - 1; pstate != 0; pstate--) {
351 if (computed_flops / rates[task_id][pstate] <= target_time) {
352 // We just found the pstate we want to use!
353 best_pstate = pstate;
360 void update() override {}
364 } // namespace plugin
365 } // namespace simgrid
367 /* **************************** events callback *************************** */
368 static void on_host_added(simgrid::s4u::Host& host)
370 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
373 std::string name = std::string("dvfs-daemon-") + host.get_cname();
374 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
376 * This lambda function is the function the actor (daemon) will execute
377 * all the time - in the case of the dvfs plugin, this controls when to
378 * lower/raise the frequency.
380 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
382 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
384 std::string dvfs_governor;
385 if (const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor")) {
386 dvfs_governor = std::string(host_conf);
387 boost::algorithm::to_lower(dvfs_governor);
389 dvfs_governor = cfg_governor;
390 boost::algorithm::to_lower(dvfs_governor);
393 auto governor = [&dvfs_governor, &daemon_proc]() -> std::unique_ptr<simgrid::plugin::dvfs::Governor> {
394 if (dvfs_governor == "conservative")
395 return std::make_unique<simgrid::plugin::dvfs::Conservative>(daemon_proc->get_host());
396 if (dvfs_governor == "ondemand")
397 return std::make_unique<simgrid::plugin::dvfs::OnDemand>(daemon_proc->get_host());
399 if (dvfs_governor == "adagio")
400 return std::make_unique<simgrid::plugin::dvfs::Adagio>(daemon_proc->get_host());
402 if (dvfs_governor == "powersave")
403 return std::make_unique<simgrid::plugin::dvfs::Powersave>(daemon_proc->get_host());
404 if (dvfs_governor != "performance")
405 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
406 daemon_proc->get_host()->get_cname());
407 return std::make_unique<simgrid::plugin::dvfs::Performance>(daemon_proc->get_host());
411 // Sleep *before* updating; important for startup (i.e., t = 0).
412 // In the beginning, we want to go with the pstates specified in the platform file
413 // (so we sleep first)
414 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
416 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
419 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
423 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
424 // SMPI will take this process as an MPI process!
428 /* **************************** Public interface *************************** */
431 * @brief Initializes the HostDvfs plugin
432 * @details The HostDvfs plugin provides an API to get the current load of each host.
434 void sg_host_dvfs_plugin_init()
436 static bool inited = false;
441 sg_host_load_plugin_init();
443 simgrid::s4u::Host::on_creation_cb(&on_host_added);