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 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(host_dvfs, kernel, "Logging specific to the HostDvfs plugin");
64 namespace simgrid::plugin::dvfs {
67 * Add this to your host tag:
68 * - \<prop id="plugin/dvfs/governor" value="performance" /\>
70 * Valid values as of now are: performance, powersave, ondemand, conservative
71 * It doesn't matter if you use uppercase or lowercase.
73 * For the sampling rate, use this:
75 * - \<prop id="plugin/dvfs/sampling-rate" value="2" /\>
77 * This will run the update() method of the specified governor every 2 seconds
80 * These properties can also be used within the \<config\> tag to configure
81 * these values globally. Using them within the \<host\> will overwrite this
82 * global configuration
85 simgrid::s4u::Host* const host_;
86 double sampling_rate_;
87 unsigned long min_pstate = cfg_min_pstate; //< Never use a pstate less than this one
88 unsigned long max_pstate = cfg_max_pstate; //< Never use a pstate larger than this one
91 explicit Governor(simgrid::s4u::Host* ptr)
94 if (cfg_max_pstate == MAX_PSTATE_NOT_LIMITED)
95 max_pstate = host_->get_pstate_count() - 1;
98 virtual ~Governor() = default;
99 virtual std::string get_name() const = 0;
100 simgrid::s4u::Host* get_host() const { return host_; }
101 unsigned long get_min_pstate() const { return min_pstate; }
102 unsigned long get_max_pstate() const { return max_pstate; }
106 if (const char* local_sampling_rate_config = host_->get_property(cfg_sampling_rate.get_name())) {
107 sampling_rate_ = std::stod(local_sampling_rate_config);
109 sampling_rate_ = cfg_sampling_rate;
111 if (const char* local_min_pstate_config = host_->get_property(cfg_min_pstate.get_name())) {
112 min_pstate = std::stoul(local_min_pstate_config);
115 if (const char* local_max_pstate_config = host_->get_property(cfg_max_pstate.get_name())) {
116 max_pstate = std::stoul(local_max_pstate_config);
118 xbt_assert(max_pstate <= host_->get_pstate_count() - 1, "Value for max_pstate too large!");
119 xbt_assert(min_pstate <= max_pstate, "min_pstate is larger than max_pstate!");
122 virtual void update() = 0;
123 double get_sampling_rate() const { return sampling_rate_; }
127 * The linux kernel doc describes this governor as follows:
128 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
130 * > The CPUfreq governor "performance" sets the CPU statically to the
131 * > highest frequency within the borders of scaling_min_freq and
132 * > scaling_max_freq.
134 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
136 class Performance : public Governor {
138 using Governor::Governor;
139 std::string get_name() const override { return "Performance"; }
141 void update() override { get_host()->set_pstate(get_min_pstate()); }
145 * The linux kernel doc describes this governor as follows:
146 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
148 * > The CPUfreq governor "powersave" sets the CPU statically to the
149 * > lowest frequency within the borders of scaling_min_freq and
150 * > scaling_max_freq.
152 * We do not support scaling_min_freq/scaling_max_freq -- we just pick the lowest frequency.
154 class Powersave : public Governor {
156 using Governor::Governor;
157 std::string get_name() const override { return "Powersave"; }
159 void update() override { get_host()->set_pstate(get_max_pstate()); }
163 * The linux kernel doc describes this governor as follows:
164 * https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt
166 * > The CPUfreq governor "ondemand" sets the CPU frequency depending on the
167 * > current system load. [...] when triggered, cpufreq checks
168 * > the CPU-usage statistics over the last period and the governor sets the
171 class OnDemand : public Governor {
173 * See https://elixir.bootlin.com/linux/v4.15.4/source/drivers/cpufreq/cpufreq_ondemand.c
174 * DEF_FREQUENCY_UP_THRESHOLD and od_update()
176 double freq_up_threshold_ = 0.80;
179 using Governor::Governor;
180 std::string get_name() const override { return "OnDemand"; }
182 void update() override
184 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
185 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
187 if (load > freq_up_threshold_) {
188 get_host()->set_pstate(get_min_pstate()); /* Run at max. performance! */
189 XBT_INFO("Load: %f > threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, get_min_pstate());
191 /* The actual implementation uses a formula here: (See Kernel file cpufreq_ondemand.c:158)
193 * freq_next = min_f + load * (max_f - min_f) / 100
195 * So they assume that frequency increases by 100 MHz. We will just use
196 * lowest_pstate - load*pstatesCount()
198 // Load is now < freq_up_threshold; exclude pstate 0 (the fastest)
199 // because pstate 0 can only be selected if load > freq_up_threshold_
200 auto new_pstate = get_max_pstate() - static_cast<unsigned long>(load) * (get_max_pstate() + 1);
201 if (new_pstate < get_min_pstate())
202 new_pstate = get_min_pstate();
203 get_host()->set_pstate(new_pstate);
205 XBT_DEBUG("Load: %f < threshold: %f --> changed to pstate %lu", load, freq_up_threshold_, new_pstate);
211 * This is the conservative governor, which is very similar to the
212 * OnDemand governor. The Linux Kernel Documentation describes it
213 * very well, see https://www.kernel.org/doc/Documentation/cpu-freq/governors.txt:
215 * > The CPUfreq governor "conservative", much like the "ondemand"
216 * > governor, sets the CPU frequency depending on the current usage. It
217 * > differs in behavior in that it gracefully increases and decreases the
218 * > CPU speed rather than jumping to max speed the moment there is any load
219 * > on the CPU. This behavior is more suitable in a battery powered
222 class Conservative : public Governor {
223 double freq_up_threshold_ = .8;
224 double freq_down_threshold_ = .2;
227 using Governor::Governor;
228 std::string get_name() const override { return "Conservative"; }
230 void update() override
232 double load = get_host()->get_core_count() * sg_host_get_avg_load(get_host());
233 unsigned long pstate = get_host()->get_pstate();
234 sg_host_load_reset(get_host()); // Only consider the period between two calls to this method!
236 if (load > freq_up_threshold_) {
237 if (pstate != get_min_pstate()) {
238 get_host()->set_pstate(pstate - 1);
239 XBT_INFO("Load: %f > threshold: %f -> increasing performance to pstate %lu", load, freq_up_threshold_,
242 XBT_DEBUG("Load: %f > threshold: %f -> but cannot speed up even more, already in highest pstate %lu", load,
243 freq_up_threshold_, pstate);
245 } else if (load < freq_down_threshold_) {
246 if (pstate != get_max_pstate()) { // Are we in the slowest pstate already?
247 get_host()->set_pstate(pstate + 1);
248 XBT_INFO("Load: %f < threshold: %f -> slowing down to pstate %lu", load, freq_down_threshold_, pstate + 1);
250 XBT_DEBUG("Load: %f < threshold: %f -> cannot slow down even more, already in slowest pstate %lu", load,
251 freq_down_threshold_, pstate);
258 class Adagio : public Governor {
259 unsigned long best_pstate = 0;
260 double start_time = 0;
261 double comp_counter = 0;
262 double comp_timer = 0;
264 std::vector<std::vector<double>> rates; // Each host + all frequencies of that host
266 unsigned int task_id = 0;
267 bool iteration_running = false; /*< Are we currently between iteration_in and iteration_out calls? */
270 explicit Adagio(simgrid::s4u::Host* ptr)
271 : Governor(ptr), rates(100, std::vector<double>(ptr->get_pstate_count(), 0.0))
273 simgrid::smpi::plugin::ampi::on_iteration_in.connect([this](simgrid::s4u::Actor const& actor) {
274 // Every instance of this class subscribes to this event, so one per host
275 // This means that for any actor, all 'hosts' are normally notified of these
276 // changes, even those who don't currently run the actor 'proc_id'.
277 // -> Let's check if this signal call is for us!
278 if (get_host() == actor.get_host()) {
279 iteration_running = true;
282 simgrid::smpi::plugin::ampi::on_iteration_out.connect([this](simgrid::s4u::Actor const& actor) {
283 if (get_host() == actor.get_host()) {
284 iteration_running = false;
288 simgrid::s4u::Exec::on_start_cb([this](simgrid::s4u::Exec const& activity) {
289 if (activity.get_host() == get_host())
292 simgrid::s4u::Exec::on_completion_cb([this](simgrid::s4u::Exec const& exec) {
293 // For more than one host (not yet supported), we can access the host via
294 // simcalls_.front()->issuer->get_iface()->get_host()
295 if (exec.get_host() == get_host() && iteration_running) {
296 comp_timer += exec.get_finish_time() - exec.get_start_time();
299 // FIXME I think that this fires at the same time for all hosts, so when the src sends something,
300 // the dst will be notified even though it didn't even arrive at the recv yet
301 simgrid::s4u::Comm::on_start_cb([this](const s4u::Comm& comm) {
302 if ((get_host() == comm.get_sender()->get_host() || get_host() == comm.get_receiver()->get_host()) &&
309 std::string get_name() const override { return "Adagio"; }
313 sg_host_load_reset(get_host());
314 comp_counter = sg_host_get_computed_flops(get_host()); // Should be 0 because of the reset
316 start_time = simgrid::s4u::Engine::get_clock();
317 if (rates.size() <= task_id)
318 rates.resize(task_id + 5, std::vector<double>(get_host()->get_pstate_count(), 0.0));
319 if (rates[task_id][best_pstate] == 0)
321 get_host()->set_pstate(best_pstate); // Load our schedule
322 XBT_DEBUG("Set pstate to %lu", best_pstate);
327 double computed_flops = sg_host_get_computed_flops(get_host()) - comp_counter;
328 double target_time = (simgrid::s4u::Engine::get_clock() - start_time);
329 target_time = target_time * 99.0 / 100.0; // FIXME We account for t_copy arbitrarily with 1%
330 // -- this needs to be fixed
332 bool is_initialized = rates[task_id][best_pstate] != 0;
333 rates[task_id][best_pstate] = computed_flops / comp_timer;
334 if (not is_initialized) {
335 for (unsigned long i = 1; i < get_host()->get_pstate_count(); i++) {
336 rates[task_id][i] = rates[task_id][0] * (get_host()->get_pstate_speed(i) / get_host()->get_speed());
340 for (unsigned long pstate = get_host()->get_pstate_count() - 1; pstate != 0; pstate--) {
341 if (computed_flops / rates[task_id][pstate] <= target_time) {
342 // We just found the pstate we want to use!
343 best_pstate = pstate;
350 void update() override {}
353 } // namespace simgrid::plugin::dvfs
355 /* **************************** events callback *************************** */
356 static void on_host_added(simgrid::s4u::Host& host)
358 if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host)) // Ignore virtual machines
361 std::string name = "dvfs-daemon-" + host.get_name();
362 simgrid::s4u::ActorPtr daemon = simgrid::s4u::Actor::create(name.c_str(), &host, []() {
364 * This lambda function is the function the actor (daemon) will execute
365 * all the time - in the case of the dvfs plugin, this controls when to
366 * lower/raise the frequency.
368 simgrid::s4u::ActorPtr daemon_proc = simgrid::s4u::Actor::self();
370 XBT_DEBUG("DVFS process on %s is a daemon: %d", daemon_proc->get_host()->get_cname(), daemon_proc->is_daemon());
372 std::string dvfs_governor;
373 if (const char* host_conf = daemon_proc->get_host()->get_property("plugin/dvfs/governor")) {
374 dvfs_governor = host_conf;
375 boost::algorithm::to_lower(dvfs_governor);
377 dvfs_governor = cfg_governor;
378 boost::algorithm::to_lower(dvfs_governor);
381 auto governor = [&dvfs_governor, &daemon_proc]() -> std::unique_ptr<simgrid::plugin::dvfs::Governor> {
382 if (dvfs_governor == "conservative")
383 return std::make_unique<simgrid::plugin::dvfs::Conservative>(daemon_proc->get_host());
384 if (dvfs_governor == "ondemand")
385 return std::make_unique<simgrid::plugin::dvfs::OnDemand>(daemon_proc->get_host());
387 if (dvfs_governor == "adagio")
388 return std::make_unique<simgrid::plugin::dvfs::Adagio>(daemon_proc->get_host());
390 if (dvfs_governor == "powersave")
391 return std::make_unique<simgrid::plugin::dvfs::Powersave>(daemon_proc->get_host());
392 if (dvfs_governor != "performance")
393 XBT_CRITICAL("No governor specified for host %s, falling back to Performance",
394 daemon_proc->get_host()->get_cname());
395 return std::make_unique<simgrid::plugin::dvfs::Performance>(daemon_proc->get_host());
399 // Sleep *before* updating; important for startup (i.e., t = 0).
400 // In the beginning, we want to go with the pstates specified in the platform file
401 // (so we sleep first)
402 simgrid::s4u::this_actor::sleep_for(governor->get_sampling_rate());
404 XBT_DEBUG("Governor (%s) just updated!", governor->get_name().c_str());
407 XBT_WARN("I should have never reached this point: daemons should be killed when all regular processes are done");
411 // This call must be placed in this function. Otherwise, the daemonize() call comes too late and
412 // SMPI will take this process as an MPI process!
416 /* **************************** Public interface *************************** */
419 * @brief Initializes the HostDvfs plugin
420 * @details The HostDvfs plugin provides an API to get the current load of each host.
422 void sg_host_dvfs_plugin_init()
424 static bool inited = false;
429 sg_host_load_plugin_init();
431 simgrid::s4u::Host::on_creation_cb(&on_host_added);