1 /* Copyright (c) 2009-2011. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "surf_private.h"
8 #include "surf/surf_resource.h"
9 #include "maxmin_private.h"
10 #include "simgrid/sg_config.h"
11 #include "surf/cpu_cas01_private.h"
13 /* the model objects for physical machines and virtual machines */
14 surf_model_t surf_cpu_model_pm = NULL;
15 surf_model_t surf_cpu_model_vm = NULL;
17 #undef GENERIC_LMM_ACTION
20 #define GENERIC_LMM_ACTION(action) action->generic_lmm_action
21 #define GENERIC_ACTION(action) GENERIC_LMM_ACTION(action).generic_action
22 #define ACTION_GET_CPU(action) ((surf_action_cpu_Cas01_t) action)->cpu
24 typedef struct surf_action_cpu_cas01 {
25 s_surf_action_lmm_t generic_lmm_action;
26 } s_surf_action_cpu_Cas01_t, *surf_action_cpu_Cas01_t;
28 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_cpu, surf,
29 "Logging specific to the SURF CPU IMPROVED module");
32 cpu_running_action_set_that_does_not_need_being_checked = NULL;
35 void *cpu_cas01_create_resource(const char *name, double power_peak,
37 tmgr_trace_t power_trace,
39 e_surf_resource_state_t state_initial,
40 tmgr_trace_t state_trace,
41 xbt_dict_t cpu_properties,
42 surf_model_t cpu_model)
44 cpu_Cas01_t cpu = NULL;
46 xbt_assert(!surf_cpu_resource_priv(surf_cpu_resource_by_name(name)),
47 "Host '%s' declared several times in the platform file",
49 cpu = (cpu_Cas01_t) surf_resource_new(sizeof(s_cpu_Cas01_t),
52 cpu->power_peak = power_peak;
53 xbt_assert(cpu->power_peak > 0, "Power has to be >0");
54 cpu->power_scale = power_scale;
56 xbt_assert(core > 0, "Invalid number of cores %d", core);
60 tmgr_history_add_trace(history, power_trace, 0.0, 0, cpu);
62 cpu->state_current = state_initial;
65 tmgr_history_add_trace(history, state_trace, 0.0, 0, cpu);
68 lmm_constraint_new(cpu_model->model_private->maxmin_system, cpu,
69 cpu->core * cpu->power_scale * cpu->power_peak);
71 /* Note (hypervisor): we create a constraint object for each CPU core, which
72 * is used for making a contraint problem of CPU affinity.
75 /* At now, we assume that a VM does not have a multicore CPU. */
77 xbt_assert(cpu_model == surf_cpu_model_pm);
79 cpu->constraint_core = xbt_new(lmm_constraint_t, core);
82 for (i = 0; i < core; i++) {
83 /* just for a unique id, never used as a string. */
84 void *cnst_id = bprintf("%s:%lu", name, i);
85 cpu->constraint_core[i] =
86 lmm_constraint_new(cpu_model->model_private->maxmin_system, cnst_id,
87 cpu->power_scale * cpu->power_peak);
91 xbt_lib_set(host_lib, name, SURF_CPU_LEVEL, cpu);
93 return xbt_lib_get_elm_or_null(host_lib, name);;
97 static void parse_cpu_init(sg_platf_host_cbarg_t host)
99 /* This function is called when a platform file is parsed. Physical machines
100 * are defined there. Thus, we use the cpu model object for the physical
102 cpu_cas01_create_resource(host->id,
108 host->state_trace, host->properties,
112 static void cpu_add_traces_cpu(void)
114 xbt_dict_cursor_t cursor = NULL;
115 char *trace_name, *elm;
116 static int called = 0;
121 /* connect all traces relative to hosts */
122 xbt_dict_foreach(trace_connect_list_host_avail, cursor, trace_name, elm) {
123 tmgr_trace_t trace = xbt_dict_get_or_null(traces_set_list, trace_name);
124 cpu_Cas01_t host = surf_cpu_resource_by_name(elm);
126 xbt_assert(host, "Host %s undefined", elm);
127 xbt_assert(trace, "Trace %s undefined", trace_name);
130 tmgr_history_add_trace(history, trace, 0.0, 0, host);
133 xbt_dict_foreach(trace_connect_list_power, cursor, trace_name, elm) {
134 tmgr_trace_t trace = xbt_dict_get_or_null(traces_set_list, trace_name);
135 cpu_Cas01_t host = surf_cpu_resource_by_name(elm);
137 xbt_assert(host, "Host %s undefined", elm);
138 xbt_assert(trace, "Trace %s undefined", trace_name);
141 tmgr_history_add_trace(history, trace, 0.0, 0, host);
145 static void cpu_define_callbacks_cas01()
147 sg_platf_host_add_cb(parse_cpu_init);
148 sg_platf_postparse_add_cb(cpu_add_traces_cpu);
151 static int cpu_resource_used(void *resource)
153 surf_model_t cpu_model = ((surf_resource_t) resource)->model;
155 /* Note (hypervisor): we do not need to look up constraint_core[i] here. Even
156 * when a task is pinned or not, its variable object is always linked to the
157 * basic contraint object.
160 return lmm_constraint_used(cpu_model->model_private->maxmin_system,
161 ((cpu_Cas01_t) resource)->constraint);
164 static double cpu_share_resources_lazy(surf_model_t cpu_model, double now)
166 return generic_share_resources_lazy(now, cpu_model);
169 static double cpu_share_resources_full(surf_model_t cpu_model, double now)
171 s_surf_action_cpu_Cas01_t action;
172 return generic_maxmin_share_resources(cpu_model->states.
174 xbt_swag_offset(action,
177 cpu_model->model_private->maxmin_system, lmm_solve);
180 static void cpu_update_actions_state_lazy(surf_model_t cpu_model, double now, double delta)
182 generic_update_actions_state_lazy(now, delta, cpu_model);
185 static void cpu_update_actions_state_full(surf_model_t cpu_model, double now, double delta)
187 generic_update_actions_state_full(now, delta, cpu_model);
190 static void cpu_update_resource_state(void *id,
191 tmgr_trace_event_t event_type,
192 double value, double date)
194 cpu_Cas01_t cpu = id;
195 lmm_variable_t var = NULL;
196 lmm_element_t elem = NULL;
197 surf_model_t cpu_model = ((surf_resource_t) cpu)->model;
199 surf_watched_hosts();
201 if (event_type == cpu->power_event) {
202 /* TODO (Hypervisor): do the same thing for constraint_core[i] */
203 xbt_assert(cpu->core == 1, "FIXME: add power scaling code also for constraint_core[i]");
205 cpu->power_scale = value;
206 lmm_update_constraint_bound(cpu_model->model_private->maxmin_system, cpu->constraint,
207 cpu->core * cpu->power_scale *
210 TRACE_surf_host_set_power(date, cpu->generic_resource.name,
211 cpu->core * cpu->power_scale *
214 while ((var = lmm_get_var_from_cnst
215 (cpu_model->model_private->maxmin_system, cpu->constraint, &elem))) {
216 surf_action_cpu_Cas01_t action = lmm_variable_id(var);
217 lmm_update_variable_bound(cpu_model->model_private->maxmin_system,
218 GENERIC_LMM_ACTION(action).variable,
219 cpu->power_scale * cpu->power_peak);
221 if (tmgr_trace_event_free(event_type))
222 cpu->power_event = NULL;
223 } else if (event_type == cpu->state_event) {
224 /* TODO (Hypervisor): do the same thing for constraint_core[i] */
225 xbt_assert(cpu->core == 1, "FIXME: add state change code also for constraint_core[i]");
228 cpu->state_current = SURF_RESOURCE_ON;
230 lmm_constraint_t cnst = cpu->constraint;
232 cpu->state_current = SURF_RESOURCE_OFF;
234 while ((var = lmm_get_var_from_cnst(cpu_model->model_private->maxmin_system, cnst, &elem))) {
235 surf_action_t action = lmm_variable_id(var);
237 if (surf_action_state_get(action) == SURF_ACTION_RUNNING ||
238 surf_action_state_get(action) == SURF_ACTION_READY ||
239 surf_action_state_get(action) ==
240 SURF_ACTION_NOT_IN_THE_SYSTEM) {
241 action->finish = date;
242 surf_action_state_set(action, SURF_ACTION_FAILED);
246 if (tmgr_trace_event_free(event_type))
247 cpu->state_event = NULL;
249 XBT_CRITICAL("Unknown event ! \n");
259 * This function formulates a constraint problem that pins a given task to
260 * particular cores. Currently, it is possible to pin a task to an exactly one
261 * specific core. The system links the variable object of the task to the
262 * per-core constraint object.
264 * But, the taskset command on Linux takes a mask value specifying a CPU
265 * affinity setting of a given task. If the mask value is 0x03, the given task
266 * will be executed on the first core (CPU0) or the second core (CPU1) on the
267 * given PM. The schedular will determine appropriate placements of tasks,
268 * considering given CPU affinities and task activities.
270 * How should the system formulate constraint problems for an affinity to
273 * The cpu argument must be the host where the task is being executed. The
274 * action object does not have the information about the location where the
275 * action is being executed.
277 static void cpu_action_set_affinity(surf_action_t action, void *cpu, unsigned long mask)
279 lmm_variable_t var_obj = ((surf_action_lmm_t) action)->variable;
281 surf_model_t cpu_model = action->model_obj;
282 xbt_assert(cpu_model->type == SURF_MODEL_TYPE_CPU);
283 cpu_Cas01_t CPU = surf_cpu_resource_priv(cpu);
285 XBT_IN("(%p,%lx)", action, mask);
288 unsigned long nbits = 0;
290 /* FIXME: There is much faster algorithms doing this. */
292 for (i = 0; i < CPU->core; i++) {
293 unsigned long has_affinity = (1UL << i) & mask;
299 XBT_CRITICAL("Do not specify multiple cores for an affinity mask.");
300 XBT_CRITICAL("See the comment in cpu_action_set_affinity().");
308 for (i = 0; i < CPU->core; i++) {
309 XBT_DEBUG("clear affinity %p to cpu-%lu@%s", action, i, CPU->generic_resource.name);
310 lmm_shrink(cpu_model->model_private->maxmin_system, CPU->constraint_core[i], var_obj);
312 unsigned long has_affinity = (1UL << i) & mask;
314 /* This function only accepts an affinity setting on the host where the
315 * task is now running. In future, a task might move to another host.
316 * But, at this moment, this function cannot take an affinity setting on
319 * It might be possible to extend the code to allow this function to
320 * accept affinity settings on a future host. We might be able to assign
321 * zero to elem->value to maintain such inactive affinity settings in the
322 * system. But, this will make the system complex. */
323 XBT_DEBUG("set affinity %p to cpu-%lu@%s", action, i, CPU->generic_resource.name);
324 lmm_expand(cpu_model->model_private->maxmin_system, CPU->constraint_core[i], var_obj, 1.0);
328 if (cpu_model->model_private->update_mechanism == UM_LAZY) {
329 /* FIXME (hypervisor): Do we need to do something for the LAZY mode? */
335 static surf_action_t cpu_execute(void *cpu, double size)
337 surf_action_cpu_Cas01_t action = NULL;
338 cpu_Cas01_t CPU = surf_cpu_resource_priv(cpu);
339 surf_model_t cpu_model = ((surf_resource_t) CPU)->model;
341 XBT_IN("(%s,%g)", surf_resource_name(CPU), size);
343 surf_action_new(sizeof(s_surf_action_cpu_Cas01_t), size,
345 CPU->state_current != SURF_RESOURCE_ON);
347 GENERIC_LMM_ACTION(action).suspended = 0; /* Should be useless because of the
348 calloc but it seems to help valgrind... */
350 /* Note (hypervisor): here, the bound value of the variable is set to the
351 * capacity of a CPU core. But, after MSG_{task/vm}_set_bound() were added to
352 * the hypervisor branch, this bound value is overwritten in
353 * SIMIX_host_execute().
354 * TODO: cleanup this.
356 GENERIC_LMM_ACTION(action).variable =
357 lmm_variable_new(cpu_model->model_private->maxmin_system, action,
358 GENERIC_ACTION(action).priority,
359 CPU->power_scale * CPU->power_peak, 1 + CPU->core); // the basic constraint plus core-specific constraints
360 if (cpu_model->model_private->update_mechanism == UM_LAZY) {
361 GENERIC_LMM_ACTION(action).index_heap = -1;
362 GENERIC_LMM_ACTION(action).last_update = surf_get_clock();
363 GENERIC_LMM_ACTION(action).last_value = 0.0;
365 lmm_expand(cpu_model->model_private->maxmin_system, CPU->constraint,
366 GENERIC_LMM_ACTION(action).variable, 1.0);
368 return (surf_action_t) action;
371 static surf_action_t cpu_action_sleep(void *cpu, double duration)
373 surf_action_cpu_Cas01_t action = NULL;
374 cpu_Cas01_t CPU = surf_cpu_resource_priv(cpu);
375 surf_model_t cpu_model = ((surf_resource_t) CPU)->model;
378 duration = MAX(duration, MAXMIN_PRECISION);
380 XBT_IN("(%s,%g)", surf_resource_name(surf_cpu_resource_priv(cpu)), duration);
381 action = (surf_action_cpu_Cas01_t) cpu_execute(cpu, 1.0);
382 // FIXME: sleep variables should not consume 1.0 in lmm_expand
383 GENERIC_ACTION(action).max_duration = duration;
384 GENERIC_LMM_ACTION(action).suspended = 2;
385 if (duration == NO_MAX_DURATION) {
386 /* Move to the *end* of the corresponding action set. This convention
387 is used to speed up update_resource_state */
388 xbt_swag_remove(action, ((surf_action_t) action)->state_set);
389 ((surf_action_t) action)->state_set =
390 cpu_running_action_set_that_does_not_need_being_checked;
391 xbt_swag_insert(action, ((surf_action_t) action)->state_set);
394 lmm_update_variable_weight(cpu_model->model_private->maxmin_system,
395 GENERIC_LMM_ACTION(action).variable, 0.0);
396 if (cpu_model->model_private->update_mechanism == UM_LAZY) { // remove action from the heap
397 surf_action_lmm_heap_remove(cpu_model->model_private->action_heap,(surf_action_lmm_t)action);
398 // this is necessary for a variable with weight 0 since such
399 // variables are ignored in lmm and we need to set its max_duration
400 // correctly at the next call to share_resources
401 xbt_swag_insert_at_head(action, cpu_model->model_private->modified_set);
405 return (surf_action_t) action;
408 static e_surf_resource_state_t cpu_get_state(void *cpu)
410 return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->state_current;
413 static void cpu_set_state(void *cpu, e_surf_resource_state_t state)
415 ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->state_current = state;
418 static double cpu_get_speed(void *cpu, double load)
420 return load * ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_peak;
423 static int cpu_get_core(void *cpu)
425 return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->core;
427 static double cpu_get_available_speed(void *cpu)
429 /* number between 0 and 1 */
430 return ((cpu_Cas01_t)surf_cpu_resource_priv(cpu))->power_scale;
433 static void cpu_finalize(surf_model_t cpu_model)
435 lmm_system_free(cpu_model->model_private->maxmin_system);
436 cpu_model->model_private->maxmin_system = NULL;
438 if (cpu_model->model_private->action_heap)
439 xbt_heap_free(cpu_model->model_private->action_heap);
440 xbt_swag_free(cpu_model->model_private->modified_set);
442 surf_model_exit(cpu_model);
445 xbt_swag_free(cpu_running_action_set_that_does_not_need_being_checked);
446 cpu_running_action_set_that_does_not_need_being_checked = NULL;
449 static surf_model_t surf_cpu_model_init_cas01(void)
451 s_surf_action_t action;
452 s_surf_action_cpu_Cas01_t comp;
454 char *optim = xbt_cfg_get_string(_sg_cfg_set, "cpu/optim");
456 xbt_cfg_get_boolean(_sg_cfg_set, "cpu/maxmin_selective_update");
458 surf_model_t cpu_model = surf_model_init();
460 if (!strcmp(optim, "Full")) {
461 cpu_model->model_private->update_mechanism = UM_FULL;
462 cpu_model->model_private->selective_update = select;
463 } else if (!strcmp(optim, "Lazy")) {
464 cpu_model->model_private->update_mechanism = UM_LAZY;
465 cpu_model->model_private->selective_update = 1;
466 xbt_assert((select == 1)
468 (xbt_cfg_is_default_value
469 (_sg_cfg_set, "cpu/maxmin_selective_update")),
470 "Disabling selective update while using the lazy update mechanism is dumb!");
472 xbt_die("Unsupported optimization (%s) for this model", optim);
475 cpu_running_action_set_that_does_not_need_being_checked =
476 xbt_swag_new(xbt_swag_offset(action, state_hookup));
478 cpu_model->name = "cpu";
479 cpu_model->type = SURF_MODEL_TYPE_CPU;
481 cpu_model->action_unref = surf_action_unref;
482 cpu_model->action_cancel = surf_action_cancel;
483 cpu_model->action_state_set = surf_action_state_set;
485 cpu_model->model_private->resource_used = cpu_resource_used;
487 if (cpu_model->model_private->update_mechanism == UM_LAZY) {
488 cpu_model->model_private->share_resources =
489 cpu_share_resources_lazy;
490 cpu_model->model_private->update_actions_state =
491 cpu_update_actions_state_lazy;
492 } else if (cpu_model->model_private->update_mechanism == UM_FULL) {
493 cpu_model->model_private->share_resources =
494 cpu_share_resources_full;
495 cpu_model->model_private->update_actions_state =
496 cpu_update_actions_state_full;
498 xbt_die("Invalid cpu update mechanism!");
500 cpu_model->model_private->update_resource_state =
501 cpu_update_resource_state;
502 cpu_model->model_private->finalize = cpu_finalize;
504 cpu_model->suspend = surf_action_suspend;
505 cpu_model->resume = surf_action_resume;
506 cpu_model->is_suspended = surf_action_is_suspended;
507 cpu_model->set_max_duration = surf_action_set_max_duration;
508 cpu_model->set_priority = surf_action_set_priority;
509 cpu_model->set_bound = surf_action_set_bound;
510 cpu_model->set_affinity = cpu_action_set_affinity;
512 cpu_model->set_category = surf_action_set_category;
514 cpu_model->get_remains = surf_action_get_remains;
516 cpu_model->extension.cpu.execute = cpu_execute;
517 cpu_model->extension.cpu.sleep = cpu_action_sleep;
519 cpu_model->extension.cpu.get_state = cpu_get_state;
520 cpu_model->extension.cpu.set_state = cpu_set_state;
521 cpu_model->extension.cpu.get_core = cpu_get_core;
522 cpu_model->extension.cpu.get_speed = cpu_get_speed;
523 cpu_model->extension.cpu.get_available_speed =
524 cpu_get_available_speed;
525 cpu_model->extension.cpu.add_traces = cpu_add_traces_cpu;
527 if (!cpu_model->model_private->maxmin_system) {
528 cpu_model->model_private->maxmin_system = lmm_system_new(cpu_model->model_private->selective_update);
530 if (cpu_model->model_private->update_mechanism == UM_LAZY) {
531 cpu_model->model_private->action_heap = xbt_heap_new(8, NULL);
532 xbt_heap_set_update_callback(cpu_model->model_private->action_heap,
533 surf_action_lmm_update_index_heap);
534 cpu_model->model_private->modified_set =
535 xbt_swag_new(xbt_swag_offset(comp, generic_lmm_action.action_list_hookup));
536 cpu_model->model_private->maxmin_system->keep_track = cpu_model->model_private->modified_set;
542 /*********************************************************************/
543 /* Basic sharing model for CPU: that is where all this started... ;) */
544 /*********************************************************************/
545 /* @InProceedings{casanova01simgrid, */
546 /* author = "H. Casanova", */
547 /* booktitle = "Proceedings of the IEEE Symposium on Cluster Computing */
548 /* and the Grid (CCGrid'01)", */
549 /* publisher = "IEEE Computer Society", */
550 /* title = "Simgrid: {A} Toolkit for the Simulation of Application */
554 /* note = "Available at */
555 /* \url{http://grail.sdsc.edu/papers/simgrid_ccgrid01.ps.gz}." */
559 void surf_cpu_model_init_Cas01(void)
561 char *optim = xbt_cfg_get_string(_sg_cfg_set, "cpu/optim");
563 xbt_assert(!surf_cpu_model_pm);
564 xbt_assert(!surf_cpu_model_vm);
566 if (strcmp(optim, "TI") == 0) {
567 /* FIXME: do we have to supprot TI? for VM */
568 surf_cpu_model_pm = surf_cpu_model_init_ti();
569 XBT_INFO("TI model is used (it will crashed since this is the hypervisor branch)");
571 surf_cpu_model_pm = surf_cpu_model_init_cas01();
572 surf_cpu_model_vm = surf_cpu_model_init_cas01();
574 /* cpu_model is registered only to model_list, and not to
575 * model_list_invoke. The shared_resource callback function will be called
576 * from that of the workstation model. */
577 xbt_dynar_push(model_list, &surf_cpu_model_pm);
578 xbt_dynar_push(model_list, &surf_cpu_model_vm);
580 cpu_define_callbacks_cas01();
584 /* TODO: do we address nested virtualization later? */
586 surf_model_t cpu_model_cas01(int level){
587 // TODO this table should be allocated
588 if(!surf_cpu_model[level])
590 return surf_cpu_model[level];