#include <utility>
#include <vector>
-/** @addtogroup SURF_plugin_energy
+/** @addtogroup plugin_energy
This is the energy plugin, enabling to account not only for computation time,
but also for the dissipated energy in the simulated platform.
+To activate this plugin, first call sg_host_energy_plugin_init() before your #MSG_init(),
+and then use MSG_host_get_consumed_energy() to retrieve the consumption of a given host.
-The energy consumption of a CPU depends directly of its current load. Specify that consumption in your platform file as
-follows:
+When the host is on, this energy consumption naturally depends on both the
+current CPU load and the host energy profile. According to our measurements,
+the consumption is somehow linear in the amount of cores at full speed,
+with an abnormality when all the cores are idle. The full details are in
+<a href="https://hal.inria.fr/hal-01523608">our scientific paper</a> on that topic.
-\verbatim
-<host id="HostA" power="100.0Mf" cores="8">
+As a result, our energy model takes 4 parameters:
+
+ - \b Idle: instantaneous consumption (in Watt) when your host is up and running, but without anything to do.
+ - \b OneCore: instantaneous consumption (in Watt) when only one core is active, at 100%.
+ - \b AllCores: instantaneous consumption (in Watt) when all cores of the host are at 100%.
+ - \b Off: instantaneous consumption (in Watt) when the host is turned off.
+
+Here is an example of XML declaration:
+
+\code{.xml}
+<host id="HostA" power="100.0Mf" cores="4">
<prop id="watt_per_state" value="100.0:120.0:200.0" />
<prop id="watt_off" value="10" />
</host>
-\endverbatim
+\endcode
+
+This example gives the following parameters: \b Off is 10 Watts; \b Idle is 100 Watts; \b OneCore is 120 Watts and \b
+AllCores is 200 Watts.
+This is enough to compute the consumption as a function of the amount of loaded cores:
+
+<table>
+<tr><th>#Cores loaded</th><th>Consumption</th><th>Explanation</th></tr>
+<tr><td>0</td><td> 100 Watts</td><td>Idle value</td></tr>
+<tr><td>1</td><td> 120 Watts</td><td>OneCore value</td></tr>
+<tr><td>2</td><td> 147 Watts</td><td>linear extrapolation between OneCore and AllCores</td></tr>
+<tr><td>3</td><td> 173 Watts</td><td>linear extrapolation between OneCore and AllCores</td></tr>
+<tr><td>4</td><td> 200 Watts</td><td>AllCores value</td></tr>
+</table>
+
+### What if a given core is only at load 50%?
+
+This is impossible in SimGrid because we recompute everything each time
+that the CPU starts or stops doing something. So if a core is at load 50% over
+a period, it means that it is at load 100% half of the time and at load 0% the
+rest of the time, and our model holds.
+
+### What if the host has only one core?
+
+In this case, the parameters \b OneCore and \b AllCores are obviously the same.
+Actually, SimGrid expect an energetic profile formated as 'Idle:Running' for mono-cores hosts.
+If you insist on passing 3 parameters in this case, then you must have the same value for \b OneCore and \b AllCores.
+
+\code{.xml}
+<host id="HostC" power="100.0Mf" cores="1">
+ <prop id="watt_per_state" value="95.0:200.0" /> <!-- we may have used '95:200:200' instead -->
+ <prop id="watt_off" value="10" />
+</host>
+\endcode
-The first property means that when your host is up and running, but without anything to do, it will dissipate 100 Watts.
-If only one care is active, it will dissipate 120 Watts. If it's fully loaded, it will dissipate 200 Watts. If its load is at 50%, then it will dissipate 153.33 Watts.
-The second property means that when your host is turned off, it will dissipate only 10 Watts (please note that these
-values are arbitrary).
+### How does DVFS interact with the host energy model?
-If your CPU is using pstates, then you can provide one consumption interval per pstate.
+If your host has several DVFS levels (several pstates), then you should
+give the energetic profile of each pstate level:
-\verbatim
-<host id="HostB" power="100.0Mf,50.0Mf,20.0Mf" pstate="0" >
+\code{.xml}
+<host id="HostC" power="100.0Mf,50.0Mf,20.0Mf" cores="4">
<prop id="watt_per_state" value="95.0:120.0:200.0, 93.0:115.0:170.0, 90.0:110.0:150.0" />
<prop id="watt_off" value="10" />
</host>
-\endverbatim
+\endcode
-That host has 3 levels of performance with the following performance: 100 Mflop/s, 50 Mflop/s or 20 Mflop/s.
-It starts at pstate 0 (ie, at 100 Mflop/s). In this case, you have to specify one interval per pstate in the
-watt_per_state property.
-In this example, the idle consumption is 95 Watts, 93 Watts and 90 Watts in each pstate while the CPU burn consumption
-are at 200 Watts, 170 Watts, and 150 Watts respectively. If only one core is active, this machine consumes 120 / 115 / 110 watts.
+This encodes the following values
+<table>
+<tr><th>pstate</th><th>Performance</th><th>Idle</th><th>OneCore</th><th>AllCores</th></tr>
+<tr><td>0</td><td>100 Mflop/s</td><td>95 Watts</td><td>120 Watts</td><td>200 Watts</td></tr>
+<tr><td>1</td><td>50 Mflop/s</td><td>93 Watts</td><td>115 Watts</td><td>170 Watts</td></tr>
+<tr><td>2</td><td>20 Mflop/s</td><td>90 Watts</td><td>110 Watts</td><td>150 Watts</td></tr>
+</table>
To change the pstate of a given CPU, use the following functions:
#MSG_host_get_nb_pstates(), simgrid#s4u#Host#setPstate(), #MSG_host_get_power_peak_at().
-To simulate the energy-related elements, first call the simgrid#energy#sg_energy_plugin_init() before your #MSG_init(),
-and then use the following function to retrieve the consumption of a given host: MSG_host_get_consumed_energy().
+### How accurate are these models?
+
+This model cannot be more accurate than your instantiation:
+with the default values, your result will not be accurate at all. You can still get
+accurate energy prediction, provided that you carefully instantiate the model.
+The first step is to ensure that your timing prediction match perfectly. But this
+is only the first step of the path, and you really want to read
+<a href="https://hal.inria.fr/hal-01523608">this paper</a> to see all what you need
+to do before you can get accurate energy predictions.
+
*/
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_energy, surf, "Logging specific to the SURF energy plugin");
{
double start_time = this->last_updated;
double finish_time = surf_get_clock();
- double cpu_load;
double current_speed = host->speed();
- if (current_speed <= 0)
- // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
- // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
- cpu_load = 1;
- else
- cpu_load = lmm_constraint_get_usage(host->pimpl_cpu->constraint()) / current_speed;
-
- /** Divide by the number of cores here **/
- cpu_load /= host->pimpl_cpu->coreCount();
-
- if (cpu_load > 1) // A machine with a load > 1 consumes as much as a fully loaded machine, not more
- cpu_load = 1;
-
- /* The problem with this model is that the load is always 0 or 1, never something less.
- * Another possibility could be to model the total energy as
- *
- * X/(X+Y)*W_idle + Y/(X+Y)*W_burn
- *
- * where X is the amount of idling cores, and Y the amount of computing cores.
- */
- double previous_energy = this->total_energy;
+ if (start_time < finish_time) {
+ double cpu_load;
+ // We may have start == finish if the past consumption was updated since the simcall was started
+ // for example if 2 actors requested to update the same host's consumption in a given scheduling round.
+ //
+ // Even in this case, we need to save the pstate for the next call (after this big if),
+ // which may have changed since that recent update.
+
+ if (current_speed <= 0)
+ // Some users declare a pstate of speed 0 flops (e.g., to model boot time).
+ // We consider that the machine is then fully loaded. That's arbitrary but it avoids a NaN
+ cpu_load = 1;
+ else
+ cpu_load = lmm_constraint_get_usage(host->pimpl_cpu->constraint()) / current_speed;
- double instantaneous_consumption;
- if (this->pstate == -1) // The host was off at the beginning of this time interval
- instantaneous_consumption = this->watts_off;
- else
- instantaneous_consumption = this->getCurrentWattsValue(cpu_load);
+ /** Divide by the number of cores here **/
+ cpu_load /= host->pimpl_cpu->coreCount();
- double energy_this_step = instantaneous_consumption * (finish_time - start_time);
+ if (cpu_load > 1) // A machine with a load > 1 consumes as much as a fully loaded machine, not more
+ cpu_load = 1;
- //TODO Trace: Trace energy_this_step from start_time to finish_time in host->name()
+ /* The problem with this model is that the load is always 0 or 1, never something less.
+ * Another possibility could be to model the total energy as
+ *
+ * X/(X+Y)*W_idle + Y/(X+Y)*W_burn
+ *
+ * where X is the amount of idling cores, and Y the amount of computing cores.
+ */
- this->total_energy = previous_energy + energy_this_step;
- this->last_updated = finish_time;
+ double previous_energy = this->total_energy;
- XBT_DEBUG(
- "[update_energy of %s] period=[%.2f-%.2f]; current power peak=%.0E flop/s; consumption change: %.2f J -> %.2f J",
- host->cname(), start_time, finish_time, host->pimpl_cpu->speed_.peak, previous_energy, energy_this_step);
+ double instantaneous_consumption;
+ if (this->pstate == -1) // The host was off at the beginning of this time interval
+ instantaneous_consumption = this->watts_off;
+ else
+ instantaneous_consumption = this->getCurrentWattsValue(cpu_load);
+
+ double energy_this_step = instantaneous_consumption * (finish_time - start_time);
+
+ // TODO Trace: Trace energy_this_step from start_time to finish_time in host->name()
+
+ this->total_energy = previous_energy + energy_this_step;
+ this->last_updated = finish_time;
+
+ XBT_DEBUG("[update_energy of %s] period=[%.2f-%.2f]; current power peak=%.0E flop/s; consumption change: %.2f J -> "
+ "%.2f J",
+ host->cname(), start_time, finish_time, host->pimpl_cpu->speed_.peak, previous_energy, energy_this_step);
+ }
/* Save data for the upcoming time interval: whether it's on/off and the pstate if it's on */
this->pstate = host->isOn() ? host->pstate() : -1;
std::vector<std::string> all_power_values;
boost::split(all_power_values, all_power_values_str, boost::is_any_of(","));
+ XBT_DEBUG("%s: profile: %s, cores: %d", host->cname(), all_power_values_str, host->coreCount());
int i = 0;
for (auto current_power_values_str : all_power_values) {
/* retrieve the power values associated with the current pstate */
std::vector<std::string> current_power_values;
boost::split(current_power_values, current_power_values_str, boost::is_any_of(":"));
- xbt_assert(current_power_values.size() == 3, "Power properties incorrectly defined - "
- "could not retrieve idle, min and max power values for host %s",
- host->cname());
+ if (host->coreCount() == 1) {
+ xbt_assert(current_power_values.size() == 2 || current_power_values.size() == 3,
+ "Power properties incorrectly defined for host %s."
+ "It should be 'Idle:FullSpeed' power values because you have one core only.",
+ host->cname());
+ if (current_power_values.size() == 2) {
+ // In this case, 1core == AllCores
+ current_power_values.push_back(current_power_values.at(1));
+ } else { // size == 3
+ xbt_assert((current_power_values.at(1)) == (current_power_values.at(2)),
+ "Power properties incorrectly defined for host %s.\n"
+ "The energy profile of mono-cores should be formated as 'Idle:FullSpeed' only.\n"
+ "If you go for a 'Idle:OneCore:AllCores' power profile on mono-cores, then OneCore and AllCores "
+ "must be equal.",
+ host->cname());
+ }
+ } else {
+ xbt_assert(current_power_values.size() == 3,
+ "Power properties incorrectly defined for host %s."
+ "It should be 'Idle:OneCore:AllCores' power values because you have more than one core.",
+ host->cname());
+ }
/* min_power corresponds to the idle power (cpu load = 0) */
/* max_power is the power consumed at 100% cpu load */
char* msg_idle = bprintf("Invalid idle value for pstate %d on host %s: %%s", i, host->cname());
- char* msg_min = bprintf("Invalid min value for pstate %d on host %s: %%s", i, host->cname());
- char* msg_max = bprintf("Invalid max value for pstate %d on host %s: %%s", i, host->cname());
+ char* msg_min = bprintf("Invalid OneCore value for pstate %d on host %s: %%s", i, host->cname());
+ char* msg_max = bprintf("Invalid AllCores value for pstate %d on host %s: %%s", i, host->cname());
PowerRange range(xbt_str_parse_double((current_power_values.at(0)).c_str(), msg_idle),
xbt_str_parse_double((current_power_values.at(1)).c_str(), msg_min),
xbt_str_parse_double((current_power_values.at(2)).c_str(), msg_max));
/* **************************** Public interface *************************** */
SG_BEGIN_DECL()
-/** \ingroup SURF_plugin_energy
+/** \ingroup plugin_energy
* \brief Enable host energy plugin
* \details Enable energy plugin to get joules consumption of each cpu. Call this function before #MSG_init().
*/
simgrid::surf::CpuAction::onStateChange.connect(&onActionStateChange);
}
-/** @brief Returns the total energy consumed by the host so far (in Joules)
+/** @ingroup plugin_energy
+ * @brief updates the consumption of all hosts
*
- * See also @ref SURF_plugin_energy.
+ * After this call, sg_host_get_consumed_energy() will not interrupt your process
+ * (until after the next clock update).
+ */
+void sg_host_energy_update_all()
+{
+ simgrid::simix::kernelImmediate([]() {
+ std::vector<simgrid::s4u::Host*> list;
+ simgrid::s4u::Engine::instance()->hostList(&list);
+ for (auto host : list)
+ if (dynamic_cast<simgrid::s4u::VirtualMachine*>(host) == nullptr) // Ignore virtual machines
+ host->extension<HostEnergy>()->update();
+ });
+}
+
+/** @ingroup plugin_energy
+ * @brief Returns the total energy consumed by the host so far (in Joules)
+ *
+ * Please note that since the consumption is lazily updated, it may require a simcall to update it.
+ * The result is that the actor requesting this value will be interrupted,
+ * the value will be updated in kernel mode before returning the control to the requesting actor.
*/
double sg_host_get_consumed_energy(sg_host_t host)
{
return host->extension<HostEnergy>()->getConsumedEnergy();
}
-/** @brief Get the amount of watt dissipated at the given pstate when the host is idling */
+/** @ingroup plugin_energy
+ * @brief Get the amount of watt dissipated at the given pstate when the host is idling
+ */
double sg_host_get_wattmin_at(sg_host_t host, int pstate)
{
xbt_assert(HostEnergy::EXTENSION_ID.valid(),
"The Energy plugin is not active. Please call sg_energy_plugin_init() during initialization.");
return host->extension<HostEnergy>()->getWattMinAt(pstate);
}
-/** @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100% */
+/** @ingroup plugin_energy
+ * @brief Returns the amount of watt dissipated at the given pstate when the host burns CPU at 100%
+ */
double sg_host_get_wattmax_at(sg_host_t host, int pstate)
{
xbt_assert(HostEnergy::EXTENSION_ID.valid(),
return host->extension<HostEnergy>()->getWattMaxAt(pstate);
}
+/** @ingroup plugin_energy
+ * @brief Returns the current consumption of the host
+ */
+double sg_host_get_current_consumption(sg_host_t host)
+{
+ xbt_assert(HostEnergy::EXTENSION_ID.valid(),
+ "The Energy plugin is not active. Please call sg_energy_plugin_init() during initialization.");
+ double cpu_load = lmm_constraint_get_usage(host->pimpl_cpu->constraint()) / host->speed();
+ return host->extension<HostEnergy>()->getCurrentWattsValue(cpu_load);
+}
+
SG_END_DECL()