-/* Copyright (c) 2009-2011. The SimGrid Team.
+/* Copyright (c) 2009-2013. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
#include "surf/surf_resource.h"
#include "maxmin_private.h"
#include "simgrid/sg_config.h"
-#include "surf/cpu_cas01_private.h"
+#include "cpu_cas01_private.h"
+
+#include "string.h"
+#include "stdlib.h"
/* the model objects for physical machines and virtual machines */
surf_model_t surf_cpu_model_pm = NULL;
surf_model_t surf_cpu_model_vm = NULL;
-#undef GENERIC_LMM_ACTION
-#undef GENERIC_ACTION
-#undef ACTION_GET_CPU
-#define GENERIC_LMM_ACTION(action) action->generic_lmm_action
-#define GENERIC_ACTION(action) GENERIC_LMM_ACTION(action).generic_action
-#define ACTION_GET_CPU(action) ((surf_action_cpu_Cas01_t) action)->cpu
-
-typedef struct surf_action_cpu_cas01 {
- s_surf_action_lmm_t generic_lmm_action;
-} s_surf_action_cpu_Cas01_t, *surf_action_cpu_Cas01_t;
-
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_cpu, surf,
"Logging specific to the SURF CPU IMPROVED module");
static xbt_swag_t
cpu_running_action_set_that_does_not_need_being_checked = NULL;
+/* Additionnal callback function to cleanup some data, called from surf_resource_free */
+
+static void cpu_cas01_cleanup(void* r){
+ cpu_Cas01_t cpu = (cpu_Cas01_t)r;
+ unsigned int iter;
+ xbt_dynar_t power_tuple = NULL;
+ xbt_dynar_foreach(cpu->energy->power_range_watts_list, iter, power_tuple)
+ xbt_dynar_free(&power_tuple);
+ xbt_dynar_free(&cpu->energy->power_range_watts_list);
+ xbt_dynar_free(&cpu->power_peak_list);
+ xbt_free(cpu->energy);
+ return;
+}
-void *cpu_cas01_create_resource(const char *name, double power_peak,
+/* This function is registered as a callback to sg_platf_new_host() and never called directly */
+void *cpu_cas01_create_resource(const char *name, xbt_dynar_t power_peak,
+ int pstate,
double power_scale,
tmgr_trace_t power_trace,
int core,
name);
cpu = (cpu_Cas01_t) surf_resource_new(sizeof(s_cpu_Cas01_t),
cpu_model, name,
- cpu_properties);
- cpu->power_peak = power_peak;
+ cpu_properties, &cpu_cas01_cleanup);
+ cpu->power_peak = xbt_dynar_get_as(power_peak, pstate, double);
+ cpu->power_peak_list = power_peak;
+ cpu->pstate = pstate;
+
+ cpu->energy = xbt_new(s_energy_cpu_cas01_t, 1);
+ cpu->energy->total_energy = 0;
+ cpu->energy->power_range_watts_list = cpu_get_watts_range_list(cpu);
+ cpu->energy->last_updated = surf_get_clock();
+
+ XBT_DEBUG("CPU create: peak=%f, pstate=%d",cpu->power_peak, cpu->pstate);
+
xbt_assert(cpu->power_peak > 0, "Power has to be >0");
cpu->power_scale = power_scale;
cpu->core = core;
lmm_constraint_new(cpu_model->model_private->maxmin_system, cpu,
cpu->core * cpu->power_scale * cpu->power_peak);
+ /* Note (hypervisor): we create a constraint object for each CPU core, which
+ * is used for making a contraint problem of CPU affinity.
+ **/
+ {
+ /* At now, we assume that a VM does not have a multicore CPU. */
+ if (core > 1)
+ xbt_assert(cpu_model == surf_cpu_model_pm);
+
+ cpu->constraint_core = xbt_new(lmm_constraint_t, core);
+
+ unsigned long i;
+ for (i = 0; i < core; i++) {
+ /* just for a unique id, never used as a string. */
+ void *cnst_id = bprintf("%s:%lu", name, i);
+ cpu->constraint_core[i] =
+ lmm_constraint_new(cpu_model->model_private->maxmin_system, cnst_id,
+ cpu->power_scale * cpu->power_peak);
+ }
+ }
+
xbt_lib_set(host_lib, name, SURF_CPU_LEVEL, cpu);
return xbt_lib_get_elm_or_null(host_lib, name);;
* machine layer. */
cpu_cas01_create_resource(host->id,
host->power_peak,
+ host->pstate,
host->power_scale,
host->power_trace,
host->core_amount,
{
surf_model_t cpu_model = ((surf_resource_t) resource)->model;
+ /* Note (hypervisor): we do not need to look up constraint_core[i] here. Even
+ * when a task is pinned or not, its variable object is always linked to the
+ * basic contraint object.
+ **/
+
return lmm_constraint_used(cpu_model->model_private->maxmin_system,
((cpu_Cas01_t) resource)->constraint);
}
generic_update_actions_state_full(now, delta, cpu_model);
}
+xbt_dynar_t cpu_get_watts_range_list(cpu_Cas01_t cpu_model)
+{
+ xbt_dynar_t power_range_list;
+ xbt_dynar_t power_tuple;
+ int i = 0, pstate_nb=0;
+ xbt_dynar_t current_power_values;
+ double min_power, max_power;
+ xbt_dict_t props = cpu_model->generic_resource.properties;
+
+ if (props == NULL)
+ return NULL;
+
+ char* all_power_values_str = xbt_dict_get_or_null(props, "power_per_state");
+
+ if (all_power_values_str == NULL)
+ return NULL;
+
+
+ power_range_list = xbt_dynar_new(sizeof(xbt_dynar_t), NULL);
+ xbt_dynar_t all_power_values = xbt_str_split(all_power_values_str, ",");
+
+ pstate_nb = xbt_dynar_length(all_power_values);
+ for (i=0; i< pstate_nb; i++)
+ {
+ /* retrieve the power values associated with the current pstate */
+ current_power_values = xbt_str_split(xbt_dynar_get_as(all_power_values, i, char*), ":");
+ xbt_assert(xbt_dynar_length(current_power_values) > 1,
+ "Power properties incorrectly defined - could not retrieve min and max power values for host %s",
+ cpu_model->generic_resource.name);
+
+ /* min_power corresponds to the idle power (cpu load = 0) */
+ /* max_power is the power consumed at 100% cpu load */
+ min_power = atof(xbt_dynar_get_as(current_power_values, 0, char*));
+ max_power = atof(xbt_dynar_get_as(current_power_values, 1, char*));
+
+ power_tuple = xbt_dynar_new(sizeof(double), NULL);
+ xbt_dynar_push_as(power_tuple, double, min_power);
+ xbt_dynar_push_as(power_tuple, double, max_power);
+
+ xbt_dynar_push_as(power_range_list, xbt_dynar_t, power_tuple);
+ xbt_dynar_free(¤t_power_values);
+ }
+ xbt_dynar_free(&all_power_values);
+ return power_range_list;
+
+}
+
+/**
+ * Computes the power consumed by the host according to the current pstate and processor load
+ *
+ */
+static double cpu_get_current_watts_value(cpu_Cas01_t cpu_model, double cpu_load)
+{
+ xbt_dynar_t power_range_list = cpu_model->energy->power_range_watts_list;
+
+ if (power_range_list == NULL)
+ {
+ XBT_DEBUG("No power range properties specified for host %s", cpu_model->generic_resource.name);
+ return 0;
+ }
+ xbt_assert(xbt_dynar_length(power_range_list) == xbt_dynar_length(cpu_model->power_peak_list),
+ "The number of power ranges in the properties does not match the number of pstates for host %s",
+ cpu_model->generic_resource.name);
+
+ /* retrieve the power values associated with the current pstate */
+ xbt_dynar_t current_power_values = xbt_dynar_get_as(power_range_list, cpu_model->pstate, xbt_dynar_t);
+
+ /* min_power corresponds to the idle power (cpu load = 0) */
+ /* max_power is the power consumed at 100% cpu load */
+ double min_power = xbt_dynar_get_as(current_power_values, 0, double);
+ double max_power = xbt_dynar_get_as(current_power_values, 1, double);
+ double power_slope = max_power - min_power;
+
+ double current_power = min_power + cpu_load * power_slope;
+
+ XBT_DEBUG("[get_current_watts] min_power=%f, max_power=%f, slope=%f", min_power, max_power, power_slope);
+ XBT_DEBUG("[get_current_watts] Current power (watts) = %f, load = %f", current_power, cpu_load);
+
+ return current_power;
+
+}
+
+/**
+ * Updates the total energy consumed as the sum of the current energy and
+ * the energy consumed by the current action
+ */
+void cpu_update_energy(cpu_Cas01_t cpu_model, double cpu_load)
+{
+
+ double start_time = cpu_model->energy->last_updated;
+ double finish_time = surf_get_clock();
+
+ XBT_DEBUG("[cpu_update_energy] action time interval=(%f-%f), current power peak=%f, current pstate=%d",
+ start_time, finish_time, cpu_model->power_peak, cpu_model->pstate);
+ double current_energy = cpu_model->energy->total_energy;
+ double action_energy = cpu_get_current_watts_value(cpu_model, cpu_load)*(finish_time-start_time);
+
+ cpu_model->energy->total_energy = current_energy + action_energy;
+ cpu_model->energy->last_updated = finish_time;
+
+ XBT_DEBUG("[cpu_update_energy] old_energy_value=%f, action_energy_value=%f", current_energy, action_energy);
+
+}
+
static void cpu_update_resource_state(void *id,
tmgr_trace_event_t event_type,
double value, double date)
lmm_element_t elem = NULL;
surf_model_t cpu_model = ((surf_resource_t) cpu)->model;
- surf_watched_hosts();
-
if (event_type == cpu->power_event) {
+ /* TODO (Hypervisor): do the same thing for constraint_core[i] */
+ xbt_assert(cpu->core == 1, "FIXME: add power scaling code also for constraint_core[i]");
+
cpu->power_scale = value;
lmm_update_constraint_bound(cpu_model->model_private->maxmin_system, cpu->constraint,
cpu->core * cpu->power_scale *
if (tmgr_trace_event_free(event_type))
cpu->power_event = NULL;
} else if (event_type == cpu->state_event) {
- if (value > 0)
+ /* TODO (Hypervisor): do the same thing for constraint_core[i] */
+ xbt_assert(cpu->core == 1, "FIXME: add state change code also for constraint_core[i]");
+
+ if (value > 0) {
+ if(cpu->state_current == SURF_RESOURCE_OFF)
+ xbt_dynar_push_as(host_that_restart, char*, (cpu->generic_resource.name));
cpu->state_current = SURF_RESOURCE_ON;
- else {
+ } else {
lmm_constraint_t cnst = cpu->constraint;
cpu->state_current = SURF_RESOURCE_OFF;
return;
}
+
+/*
+ *
+ * This function formulates a constraint problem that pins a given task to
+ * particular cores. Currently, it is possible to pin a task to an exactly one
+ * specific core. The system links the variable object of the task to the
+ * per-core constraint object.
+ *
+ * But, the taskset command on Linux takes a mask value specifying a CPU
+ * affinity setting of a given task. If the mask value is 0x03, the given task
+ * will be executed on the first core (CPU0) or the second core (CPU1) on the
+ * given PM. The schedular will determine appropriate placements of tasks,
+ * considering given CPU affinities and task activities.
+ *
+ * How should the system formulate constraint problems for an affinity to
+ * multiple cores?
+ *
+ * The cpu argument must be the host where the task is being executed. The
+ * action object does not have the information about the location where the
+ * action is being executed.
+ */
+static void cpu_action_set_affinity(surf_action_t action, void *cpu, unsigned long mask)
+{
+ lmm_variable_t var_obj = ((surf_action_lmm_t) action)->variable;
+
+ surf_model_t cpu_model = action->model_obj;
+ xbt_assert(cpu_model->type == SURF_MODEL_TYPE_CPU);
+ cpu_Cas01_t CPU = surf_cpu_resource_priv(cpu);
+
+ XBT_IN("(%p,%lx)", action, mask);
+
+ {
+ unsigned long nbits = 0;
+
+ /* FIXME: There is much faster algorithms doing this. */
+ unsigned long i;
+ for (i = 0; i < CPU->core; i++) {
+ unsigned long has_affinity = (1UL << i) & mask;
+ if (has_affinity)
+ nbits += 1;
+ }
+
+ if (nbits > 1) {
+ XBT_CRITICAL("Do not specify multiple cores for an affinity mask.");
+ XBT_CRITICAL("See the comment in cpu_action_set_affinity().");
+ DIE_IMPOSSIBLE;
+ }
+ }
+
+
+
+ unsigned long i;
+ for (i = 0; i < CPU->core; i++) {
+ XBT_DEBUG("clear affinity %p to cpu-%lu@%s", action, i, CPU->generic_resource.name);
+ lmm_shrink(cpu_model->model_private->maxmin_system, CPU->constraint_core[i], var_obj);
+
+ unsigned long has_affinity = (1UL << i) & mask;
+ if (has_affinity) {
+ /* This function only accepts an affinity setting on the host where the
+ * task is now running. In future, a task might move to another host.
+ * But, at this moment, this function cannot take an affinity setting on
+ * that future host.
+ *
+ * It might be possible to extend the code to allow this function to
+ * accept affinity settings on a future host. We might be able to assign
+ * zero to elem->value to maintain such inactive affinity settings in the
+ * system. But, this will make the system complex. */
+ XBT_DEBUG("set affinity %p to cpu-%lu@%s", action, i, CPU->generic_resource.name);
+ lmm_expand(cpu_model->model_private->maxmin_system, CPU->constraint_core[i], var_obj, 1.0);
+ }
+ }
+
+ if (cpu_model->model_private->update_mechanism == UM_LAZY) {
+ /* FIXME (hypervisor): Do we need to do something for the LAZY mode? */
+ }
+
+ XBT_OUT();
+}
+
static surf_action_t cpu_execute(void *cpu, double size)
{
surf_action_cpu_Cas01_t action = NULL;
+ //xbt_dict_cursor_t cursor = NULL;
cpu_Cas01_t CPU = surf_cpu_resource_priv(cpu);
surf_model_t cpu_model = ((surf_resource_t) CPU)->model;
GENERIC_LMM_ACTION(action).suspended = 0; /* Should be useless because of the
calloc but it seems to help valgrind... */
+ /* Note (hypervisor): here, the bound value of the variable is set to the
+ * capacity of a CPU core. But, after MSG_{task/vm}_set_bound() were added to
+ * the hypervisor branch, this bound value is overwritten in
+ * SIMIX_host_execute().
+ * TODO: cleanup this.
+ */
GENERIC_LMM_ACTION(action).variable =
lmm_variable_new(cpu_model->model_private->maxmin_system, action,
GENERIC_ACTION(action).priority,
- CPU->power_scale * CPU->power_peak, 1);
+ CPU->power_scale * CPU->power_peak, 1 + CPU->core); // the basic constraint plus core-specific constraints
if (cpu_model->model_private->update_mechanism == UM_LAZY) {
GENERIC_LMM_ACTION(action).index_heap = -1;
GENERIC_LMM_ACTION(action).last_update = surf_get_clock();
return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->state_current;
}
+static void cpu_set_state(void *cpu, e_surf_resource_state_t state)
+{
+ ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->state_current = state;
+}
+
static double cpu_get_speed(void *cpu, double load)
{
return load * ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_peak;
}
+static int cpu_get_core(void *cpu)
+{
+ return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->core;
+}
static double cpu_get_available_speed(void *cpu)
{
/* number between 0 and 1 */
return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_scale;
}
+static double cpu_get_current_power_peak(void *cpu)
+{
+ return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_peak;
+}
+
+static double cpu_get_power_peak_at(void *cpu, int pstate_index)
+{
+ xbt_dynar_t plist = ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_peak_list;
+ xbt_assert((pstate_index <= xbt_dynar_length(plist)), "Invalid parameters (pstate index out of bounds)");
+
+ return xbt_dynar_get_as(plist, pstate_index, double);
+}
+
+static int cpu_get_nb_pstates(void *cpu)
+{
+ return xbt_dynar_length(((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_peak_list);
+}
+
+static void cpu_set_power_peak_at(void *cpu, int pstate_index)
+{
+ cpu_Cas01_t cpu_implem = (cpu_Cas01_t)surf_cpu_resource_priv(cpu);
+ xbt_dynar_t plist = cpu_implem->power_peak_list;
+ xbt_assert((pstate_index <= xbt_dynar_length(plist)), "Invalid parameters (pstate index out of bounds)");
+
+ double new_power_peak = xbt_dynar_get_as(plist, pstate_index, double);
+ cpu_implem->pstate = pstate_index;
+ cpu_implem->power_peak = new_power_peak;
+}
+
+static double cpu_get_consumed_energy(void *cpu)
+{
+ return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->energy->total_energy;
+}
+
static void cpu_finalize(surf_model_t cpu_model)
{
lmm_system_free(cpu_model->model_private->maxmin_system);
char *optim = xbt_cfg_get_string(_sg_cfg_set, "cpu/optim");
int select =
- xbt_cfg_get_int(_sg_cfg_set, "cpu/maxmin_selective_update");
+ xbt_cfg_get_boolean(_sg_cfg_set, "cpu/maxmin_selective_update");
surf_model_t cpu_model = surf_model_init();
cpu_model->set_max_duration = surf_action_set_max_duration;
cpu_model->set_priority = surf_action_set_priority;
cpu_model->set_bound = surf_action_set_bound;
+ cpu_model->set_affinity = cpu_action_set_affinity;
#ifdef HAVE_TRACING
cpu_model->set_category = surf_action_set_category;
#endif
cpu_model->extension.cpu.sleep = cpu_action_sleep;
cpu_model->extension.cpu.get_state = cpu_get_state;
+ cpu_model->extension.cpu.set_state = cpu_set_state;
+ cpu_model->extension.cpu.get_core = cpu_get_core;
cpu_model->extension.cpu.get_speed = cpu_get_speed;
cpu_model->extension.cpu.get_available_speed =
cpu_get_available_speed;
+ cpu_model->extension.cpu.get_current_power_peak = cpu_get_current_power_peak;
+ cpu_model->extension.cpu.get_power_peak_at = cpu_get_power_peak_at;
+ cpu_model->extension.cpu.get_nb_pstates = cpu_get_nb_pstates;
+ cpu_model->extension.cpu.set_power_peak_at = cpu_set_power_peak_at;
+ cpu_model->extension.cpu.get_consumed_energy = cpu_get_consumed_energy;
+
cpu_model->extension.cpu.add_traces = cpu_add_traces_cpu;
if (!cpu_model->model_private->maxmin_system) {