1 /* Copyright (c) 2010-2021. 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>
25 cfg_sampling_rate("plugin/dvfs/sampling-rate",
26 "How often should the dvfs plugin check whether the frequency needs to be changed?", 0.1,
29 sg_host_dvfs_plugin_init();
32 static simgrid::config::Flag<std::string> cfg_governor("plugin/dvfs/governor",
33 "Which Governor should be used that adapts the CPU frequency?",
36 std::map<std::string, std::string, std::less<>>({
38 {"adagio", "TODO: Doc"},
40 {"conservative", "TODO: Doc"}, {"ondemand", "TODO: Doc"},
41 {"performance", "TODO: Doc"}, {"powersave", "TODO: Doc"},
44 [](const std::string& val) {
45 if (val != "performance")
46 sg_host_dvfs_plugin_init();
49 static simgrid::config::Flag<int>
50 cfg_min_pstate("plugin/dvfs/min-pstate",
51 "Which pstate is the minimum (and hence fastest) pstate for this governor?", 0);
53 static const int max_pstate_not_limited = -1;
54 static simgrid::config::Flag<int>
55 cfg_max_pstate("plugin/dvfs/max-pstate",
56 "Which pstate is the maximum (and hence slowest) pstate for this governor?", max_pstate_not_limited);
58 /** @addtogroup SURF_plugin_load
60 This plugin makes it very simple for users to obtain the current load for each host.
64 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_plugin_dvfs, surf, "Logging specific to the SURF HostDvfs plugin");
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 const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name());
112 if (local_sampling_rate_config != nullptr) {
113 sampling_rate_ = std::stod(local_sampling_rate_config);
115 sampling_rate_ = cfg_sampling_rate;
117 const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name());
118 if (local_min_pstate_config != nullptr) {
119 min_pstate = std::stoul(local_min_pstate_config);
122 const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name());
123 if (local_max_pstate_config != nullptr) {
124 max_pstate = std::stoul(local_max_pstate_config);
126 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
127 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
130 virtual void update() = 0;
131 double get_sampling_rate() const { return sampling_rate_; }
135 * The linux kernel doc describes this governor as follows:
136 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
138 * > The CPUfreq governor "performance" sets the CPU statically to the
139 * > highest frequency within the borders of scaling_min_freq and
140 * > scaling_max_freq.
142 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
144 class Performance : public Governor {
146 using Governor::Governor;
147 std::string get_name() const override { return "Performance"; }
149 void update() override { get_host()->set_pstate(get_min_pstate()); }
153 * The linux kernel doc describes this governor as follows:
154 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
156 * > The CPUfreq governor "powersave" sets the CPU statically to the
157 * > lowest frequency within the borders of scaling_min_freq and
158 * > scaling_max_freq.
160 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
162 class Powersave : public Governor {
164 using Governor::Governor;
165 std::string get_name() const override { return "Powersave"; }
167 void update() override { get_host()->set_pstate(get_max_pstate()); }
171 * The linux kernel doc describes this governor as follows:
172 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
174 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
175 * > current system load. [...] when triggered, cpufreq checks
176 * > the CPU-usage statistics over the last period and the governor sets the
179 class OnDemand : public Governor {
181 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
182 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
184 double freq_up_threshold_ = 0.80;
187 using Governor::Governor;
188 std::string get_name() const override { return "OnDemand"; }
190 void update() override
192 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
193 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
195 if (load > freq_up_threshold_) {
196 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
197 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, get_min_pstate());
199 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
201 * freq_next = min_f + load * (max_f - min_f) / 100
203 * So they assume that frequency increases by 100 MHz. We will just use
204 * lowest_pstate - load*pstatesCount()
206 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
207 // because pstate 0 can only be selected if load > freq_up_threshold_
208 unsigned long new_pstate = static_cast<int>(get_max_pstate() - load * (get_max_pstate() + 1));
209 if (new_pstate < get_min_pstate())
210 new_pstate = get_min_pstate();
211 get_host()->set_pstate(new_pstate);
213 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, new_pstate);
219 * This is the conservative governor, which is very similar to the
220 * OnDemand governor. The Linux Kernel Documentation describes it
221 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
223 * > The CPUfreq governor "conservative", much like the "ondemand"
224 * > governor, sets the CPU frequency depending on the current usage. It
225 * > differs in behavior in that it gracefully increases and decreases the
226 * > CPU speed rather than jumping to max speed the moment there is any load
227 * > on the CPU. This behavior is more suitable in a battery powered
230 class Conservative : public Governor {
231 double freq_up_threshold_ = .8;
232 double freq_down_threshold_ = .2;
235 using Governor::Governor;
236 std::string get_name() const override { return "Conservative"; }
238 void update() override
240 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
241 unsigned long pstate = get_host()->get_pstate();
242 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
244 if (load > freq_up_threshold_) {
245 if (pstate != get_min_pstate()) {
246 get_host()->set_pstate(pstate - 1);
247 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %lu", load, freq_up_threshold_,
250 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %lu", load,
251 freq_up_threshold_, pstate);
253 } else if (load < freq_down_threshold_) {
254 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
255 get_host()->set_pstate(pstate + 1);
256 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %lu", load, freq_down_threshold_, pstate + 1);
258 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %lu", load,
259 freq_down_threshold_, pstate);
266 class Adagio : public Governor {
269 double start_time = 0;
270 double comp_counter = 0;
271 double comp_timer = 0;
273 std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
275 unsigned int task_id = 0;
276 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
279 explicit Adagio(simgrid::s4u::Host* ptr)
280 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
282 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::Actor const& actor) {
283 // Every instance of this class subscribes to this event, so one per host
284 // This means that for any actor, all 'hosts' are normally notified of these
285 // changes, even those who don't currently run the actor 'proc_id'.
286 // -> Let's check if this signal call is for us!
287 if (get_host() == actor.get_host()) {
288 iteration_running = true;
291 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::Actor const& actor) {
292 if (get_host() == actor.get_host()) {
293 iteration_running = false;
297 simgrid::s4u::Exec::on_start.connect([this](simgrid::s4u::Exec const& activity) {
298 if (activity.get_host() == get_host())
301 simgrid::s4u::Exec::on_completion.connect([this](simgrid::s4u::Exec const& activity) {
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 (activity.get_host() == get_host() && iteration_running) {
305 comp_timer += activity.get_finish_time() - activity.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 simgrid::s4u::Link::on_communicate.connect([this](const kernel::resource::NetworkAction& act) {
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 %i", 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 {}
362 } // namespace plugin
363 } // namespace simgrid
365 /* **************************** events callback *************************** */
366 static void on_host_added(simgrid::s4u::Host& host)
368 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
371 std::string name = std::string("dvfs-daemon-") + host.get_cname();
372 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
374 * This lambda function is the function the actor (daemon) will execute
375 * all the time - in the case of the dvfs plugin, this controls when to
376 * lower/raise the frequency.
378 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
380 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
382 std::string dvfs_governor;
383 const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor");
384 if (host_conf != nullptr) {
385 dvfs_governor = std::string(host_conf);
386 boost::algorithm::to_lower(dvfs_governor);
388 dvfs_governor = cfg_governor;
389 boost::algorithm::to_lower(dvfs_governor);
392 auto governor = [&dvfs_governor, &daemon_proc]() -> std::unique_ptr<simgrid::plugin::dvfs::Governor> {
393 if (dvfs_governor == "conservative")
394 return std::make_unique<simgrid::plugin::dvfs::Conservative>(daemon_proc->get_host());
395 if (dvfs_governor == "ondemand")
396 return std::make_unique<simgrid::plugin::dvfs::OnDemand>(daemon_proc->get_host());
398 if (dvfs_governor == "adagio")
399 return std::make_unique<simgrid::plugin::dvfs::Adagio>(daemon_proc->get_host());
401 if (dvfs_governor == "powersave")
402 return std::make_unique<simgrid::plugin::dvfs::Powersave>(daemon_proc->get_host());
403 if (dvfs_governor != "performance")
404 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
405 daemon_proc->get_host()->get_cname());
406 return std::make_unique<simgrid::plugin::dvfs::Performance>(daemon_proc->get_host());
410 // Sleep *before* updating; important for startup (i.e., t = 0).
411 // In the beginning, we want to go with the pstates specified in the platform file
412 // (so we sleep first)
413 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
415 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
418 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
422 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
423 // SMPI will take this process as an MPI process!
427 /* **************************** Public interface *************************** */
430 * @brief Initializes the HostDvfs plugin
431 * @details The HostDvfs plugin provides an API to get the current load of each host.
433 void sg_host_dvfs_plugin_init()
435 static bool inited = false;
440 sg_host_load_plugin_init();
442 simgrid::s4u::Host::on_creation.connect(&on_host_added);