1 /* Copyright (c) 2007-2017. 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. */
7 * Modeling the proportional fairness using the Lagrangian Optimization Approach. For a detailed description see:
8 * "ssh://username@scm.gforge.inria.fr/svn/memo/people/pvelho/lagrange/ppf.ps".
10 #include "src/kernel/lmm/maxmin.hpp"
12 #include "xbt/sysdep.h"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_lagrange, surf, "Logging specific to SURF (lagrange)");
21 XBT_LOG_NEW_SUBCATEGORY(surf_lagrange_dichotomy, surf_lagrange, "Logging specific to SURF (lagrange dichotomy)");
23 #define SHOW_EXPR(expr) XBT_CDEBUG(surf_lagrange, #expr " = %g", expr);
24 #define VEGAS_SCALING 1000.0
25 #define RENO_SCALING 1.0
26 #define RENO2_SCALING 1.0
32 double (*func_f_def)(const s_lmm_variable_t&, double);
33 double (*func_fp_def)(const s_lmm_variable_t&, double);
34 double (*func_fpi_def)(const s_lmm_variable_t&, double);
37 * Local prototypes to implement the Lagrangian optimization with optimal step, also called dichotomy.
39 // solves the proportional fairness using a Lagrangian optimization with dichotomy step
40 void lagrange_solve(lmm_system_t sys);
41 // computes the value of the dichotomy using a initial values, init, with a specific variable or constraint
42 static double dichotomy(double init, double diff(double, const s_lmm_constraint_t&), const s_lmm_constraint_t& cnst,
44 // computes the value of the differential of constraint cnst applied to lambda
45 static double partial_diff_lambda(double lambda, const s_lmm_constraint_t& cnst);
47 static int __check_feasible(xbt_swag_t cnst_list, xbt_swag_t var_list, int warn)
52 xbt_swag_t elem_list = nullptr;
53 lmm_element_t elem = nullptr;
54 lmm_constraint_t cnst = nullptr;
55 lmm_variable_t var = nullptr;
57 xbt_swag_foreach(_cnst, cnst_list)
59 cnst = static_cast<lmm_constraint_t>(_cnst);
61 elem_list = &(cnst->enabled_element_set);
62 xbt_swag_foreach(_elem, elem_list)
64 elem = static_cast<lmm_element_t>(_elem);
66 xbt_assert(var->sharing_weight > 0);
70 if (double_positive(tmp - cnst->bound, sg_maxmin_precision)) {
72 XBT_WARN("The link (%p) is over-used. Expected less than %f and got %f", cnst, cnst->bound, tmp);
75 XBT_DEBUG("Checking feasability for constraint (%p): sat = %f, lambda = %f ", cnst, tmp - cnst->bound,
79 xbt_swag_foreach(_var, var_list)
81 var = static_cast<lmm_variable_t>(_var);
82 if (not var->sharing_weight)
86 XBT_DEBUG("Checking feasability for variable (%p): sat = %f mu = %f", var, var->value - var->bound, var->mu);
88 if (double_positive(var->value - var->bound, sg_maxmin_precision)) {
90 XBT_WARN("The variable (%p) is too large. Expected less than %f and got %f", var, var->bound, var->value);
97 static double new_value(const s_lmm_variable_t& var)
101 for (s_lmm_element_t const& elem : var.cnsts) {
102 tmp += elem.constraint->lambda;
106 XBT_DEBUG("\t Working on var (%p). cost = %e; Weight = %e", &var, tmp, var.sharing_weight);
107 // uses the partial differential inverse function
108 return var.func_fpi(var, tmp);
111 static double new_mu(const s_lmm_variable_t& var)
114 double sigma_i = 0.0;
116 for (s_lmm_element_t const& elem : var.cnsts) {
117 sigma_i += elem.constraint->lambda;
119 mu_i = var.func_fp(var, var.bound) - sigma_i;
125 static double dual_objective(xbt_swag_t var_list, xbt_swag_t cnst_list)
129 lmm_constraint_t cnst = nullptr;
130 lmm_variable_t var = nullptr;
134 xbt_swag_foreach(_var, var_list)
136 var = static_cast<lmm_variable_t>(_var);
137 double sigma_i = 0.0;
139 if (not var->sharing_weight)
142 for (s_lmm_element_t const& elem : var->cnsts)
143 sigma_i += elem.constraint->lambda;
148 XBT_DEBUG("var %p : sigma_i = %1.20f", var, sigma_i);
150 obj += var->func_f(*var, var->func_fpi(*var, sigma_i)) - sigma_i * var->func_fpi(*var, sigma_i);
153 obj += var->mu * var->bound;
156 xbt_swag_foreach(_cnst, cnst_list)
158 cnst = static_cast<lmm_constraint_t>(_cnst);
159 obj += cnst->lambda * cnst->bound;
165 void lagrange_solve(lmm_system_t sys)
167 /* Lagrange Variables. */
168 int max_iterations = 100;
169 double epsilon_min_error = 0.00001; /* this is the precision on the objective function so it's none of the
170 configurable values and this value is the legacy one */
171 double dichotomy_min_error = 1e-14;
172 double overall_modification = 1;
174 XBT_DEBUG("Iterative method configuration snapshot =====>");
175 XBT_DEBUG("#### Maximum number of iterations : %d", max_iterations);
176 XBT_DEBUG("#### Minimum error tolerated : %e", epsilon_min_error);
177 XBT_DEBUG("#### Minimum error tolerated (dichotomy) : %e", dichotomy_min_error);
179 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
183 if (not sys->modified)
186 /* Initialize lambda. */
187 xbt_swag_t cnst_list = &(sys->active_constraint_set);
189 xbt_swag_foreach(_cnst, cnst_list)
191 lmm_constraint_t cnst = (lmm_constraint_t)_cnst;
193 cnst->new_lambda = 2.0;
194 XBT_DEBUG("#### cnst(%p)->lambda : %e", cnst, cnst->lambda);
198 * Initialize the var list variable with only the active variables.
199 * Associate an index in the swag variables. Initialize mu.
201 xbt_swag_t var_list = &(sys->variable_set);
203 xbt_swag_foreach(_var, var_list)
205 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
206 if (not var->sharing_weight)
209 if (var->bound < 0.0) {
210 XBT_DEBUG("#### NOTE var(%p) is a boundless variable", var);
216 var->value = new_value(*var);
217 XBT_DEBUG("#### var(%p) ->weight : %e", var, var->sharing_weight);
218 XBT_DEBUG("#### var(%p) ->mu : %e", var, var->mu);
219 XBT_DEBUG("#### var(%p) ->weight: %e", var, var->sharing_weight);
220 XBT_DEBUG("#### var(%p) ->bound: %e", var, var->bound);
221 auto weighted = std::find_if(begin(var->cnsts), end(var->cnsts),
222 [](s_lmm_element_t const& x) { return x.consumption_weight != 0.0; });
223 if (weighted == end(var->cnsts))
228 /* Compute dual objective. */
229 double obj = dual_objective(var_list, cnst_list);
231 /* While doesn't reach a minimum error or a number maximum of iterations. */
233 while (overall_modification > epsilon_min_error && iteration < max_iterations) {
235 XBT_DEBUG("************** ITERATION %d **************", iteration);
236 XBT_DEBUG("-------------- Gradient Descent ----------");
238 /* Improve the value of mu_i */
239 xbt_swag_foreach(_var, var_list)
241 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
242 if (var->sharing_weight && var->bound >= 0) {
243 XBT_DEBUG("Working on var (%p)", var);
244 var->new_mu = new_mu(*var);
245 XBT_DEBUG("Updating mu : var->mu (%p) : %1.20f -> %1.20f", var, var->mu, var->new_mu);
246 var->mu = var->new_mu;
248 double new_obj = dual_objective(var_list, cnst_list);
249 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
250 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
255 /* Improve the value of lambda_i */
256 xbt_swag_foreach(_cnst, cnst_list)
258 lmm_constraint_t cnst = static_cast<lmm_constraint_t>(_cnst);
259 XBT_DEBUG("Working on cnst (%p)", cnst);
260 cnst->new_lambda = dichotomy(cnst->lambda, partial_diff_lambda, *cnst, dichotomy_min_error);
261 XBT_DEBUG("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", cnst, cnst->lambda, cnst->new_lambda);
262 cnst->lambda = cnst->new_lambda;
264 double new_obj = dual_objective(var_list, cnst_list);
265 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
266 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
270 /* Now computes the values of each variable (\rho) based on the values of \lambda and \mu. */
271 XBT_DEBUG("-------------- Check convergence ----------");
272 overall_modification = 0;
273 xbt_swag_foreach(_var, var_list)
275 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
276 if (var->sharing_weight <= 0)
279 double tmp = new_value(*var);
281 overall_modification = std::max(overall_modification, fabs(var->value - tmp));
284 XBT_DEBUG("New value of var (%p) = %e, overall_modification = %e", var, var->value, overall_modification);
288 XBT_DEBUG("-------------- Check feasability ----------");
289 if (not __check_feasible(cnst_list, var_list, 0))
290 overall_modification = 1.0;
291 XBT_DEBUG("Iteration %d: overall_modification : %f", iteration, overall_modification);
294 __check_feasible(cnst_list, var_list, 1);
296 if (overall_modification <= epsilon_min_error) {
297 XBT_DEBUG("The method converges in %d iterations.", iteration);
299 if (iteration >= max_iterations) {
300 XBT_DEBUG("Method reach %d iterations, which is the maximum number of iterations allowed.", iteration);
303 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
309 * Returns a double value corresponding to the result of a dichotomy process with respect to a given
310 * variable/constraint (\mu in the case of a variable or \lambda in case of a constraint) and a initial value init.
312 * @param init initial value for \mu or \lambda
313 * @param diff a function that computes the differential of with respect a \mu or \lambda
314 * @param var_cnst a pointer to a variable or constraint
315 * @param min_erro a minimum error tolerated
317 * @return a double corresponding to the result of the dichotomy process
319 static double dichotomy(double init, double diff(double, const s_lmm_constraint_t&), const s_lmm_constraint_t& cnst,
324 double overall_error;
331 if (fabs(init) < 1e-20) {
338 diff_0 = diff(1e-16, cnst);
340 XBT_CDEBUG(surf_lagrange_dichotomy, "returning 0.0 (diff = %e)", diff_0);
345 double min_diff = diff(min, cnst);
346 double max_diff = diff(max, cnst);
348 while (overall_error > min_error) {
349 XBT_CDEBUG(surf_lagrange_dichotomy, "[min, max] = [%1.20f, %1.20f] || diffmin, diffmax = %1.20f, %1.20f", min, max,
352 if (min_diff > 0 && max_diff > 0) {
354 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing min");
356 min_diff = diff(min, cnst);
358 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
362 } else if (min_diff < 0 && max_diff < 0) {
364 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing max");
366 max_diff = diff(max, cnst);
368 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
372 } else if (min_diff < 0 && max_diff > 0) {
373 middle = (max + min) / 2.0;
374 XBT_CDEBUG(surf_lagrange_dichotomy, "Trying (max+min)/2 : %1.20f", middle);
376 if ((fabs(min - middle) < 1e-20) || (fabs(max - middle) < 1e-20)) {
377 XBT_CWARN(surf_lagrange_dichotomy,
378 "Cannot improve the convergence! min=max=middle=%1.20f, diff = %1.20f."
379 " Reaching the 'double' limits. Maybe scaling your function would help ([%1.20f,%1.20f]).",
380 min, max - min, min_diff, max_diff);
383 middle_diff = diff(middle, cnst);
385 if (middle_diff < 0) {
386 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
388 overall_error = max_diff - middle_diff;
389 min_diff = middle_diff;
390 } else if (middle_diff > 0) {
391 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
393 overall_error = max_diff - middle_diff;
394 max_diff = middle_diff;
398 } else if (fabs(min_diff) < 1e-20) {
401 } else if (fabs(max_diff) < 1e-20) {
404 } else if (min_diff > 0 && max_diff < 0) {
405 XBT_CWARN(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0");
408 XBT_CWARN(surf_lagrange_dichotomy,
409 "diffmin (%1.20f) or diffmax (%1.20f) are something I don't know, taking no action.", min_diff,
415 XBT_CDEBUG(surf_lagrange_dichotomy, "returning %e", (min + max) / 2.0);
417 return ((min + max) / 2.0);
420 static double partial_diff_lambda(double lambda, const s_lmm_constraint_t& cnst)
426 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing diff of cnst (%p)", &cnst);
428 const_xbt_swag_t elem_list = &cnst.enabled_element_set;
430 xbt_swag_foreach(_elem, elem_list)
432 lmm_element_t elem = static_cast<lmm_element_t>(_elem);
433 lmm_variable_t var = elem->variable;
434 xbt_assert(var->sharing_weight > 0);
435 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing sigma_i for var (%p)", var);
436 // Initialize the summation variable
437 double sigma_i = 0.0;
440 for (s_lmm_element_t const& elem : var->cnsts) {
441 sigma_i += elem.constraint->lambda;
444 // add mu_i if this flow has a RTT constraint associated
448 // replace value of cnst.lambda by the value of parameter lambda
449 sigma_i = (sigma_i - cnst.lambda) + lambda;
451 diff += -var->func_fpi(*var, sigma_i);
456 XBT_CDEBUG(surf_lagrange_dichotomy, "d D/d lambda for cnst (%p) at %1.20f = %1.20f", &cnst, lambda, diff);
461 /** \brief Attribute the value bound to var->bound.
463 * \param func_fpi inverse of the partial differential of f (f prime inverse, (f')^{-1})
465 * Set default functions to the ones passed as parameters. This is a polymorphism in C pure, enjoy the roots of
469 void lmm_set_default_protocol_function(double (*func_f)(const s_lmm_variable_t& var, double x),
470 double (*func_fp)(const s_lmm_variable_t& var, double x),
471 double (*func_fpi)(const s_lmm_variable_t& var, double x))
474 func_fp_def = func_fp;
475 func_fpi_def = func_fpi;
478 /**************** Vegas and Reno functions *************************/
479 /* NOTE for Reno: all functions consider the network coefficient (alpha) equal to 1. */
482 * For Vegas: $f(x) = \alpha D_f\ln(x)$
483 * Therefore: $fp(x) = \frac{\alpha D_f}{x}$
484 * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$
486 double func_vegas_f(const s_lmm_variable_t& var, double x)
488 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
489 return VEGAS_SCALING * var.sharing_weight * log(x);
492 double func_vegas_fp(const s_lmm_variable_t& var, double x)
494 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
495 return VEGAS_SCALING * var.sharing_weight / x;
498 double func_vegas_fpi(const s_lmm_variable_t& var, double x)
500 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
501 return var.sharing_weight / (x / VEGAS_SCALING);
505 * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$
506 * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$
507 * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$
509 double func_reno_f(const s_lmm_variable_t& var, double x)
511 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
513 return RENO_SCALING * sqrt(3.0 / 2.0) / var.sharing_weight * atan(sqrt(3.0 / 2.0) * var.sharing_weight * x);
516 double func_reno_fp(const s_lmm_variable_t& var, double x)
518 return RENO_SCALING * 3.0 / (3.0 * var.sharing_weight * var.sharing_weight * x * x + 2.0);
521 double func_reno_fpi(const s_lmm_variable_t& var, double x)
525 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
526 xbt_assert(x > 0.0, "Don't call me with stupid values!");
528 res_fpi = 1.0 / (var.sharing_weight * var.sharing_weight * (x / RENO_SCALING)) -
529 2.0 / (3.0 * var.sharing_weight * var.sharing_weight);
532 return sqrt(res_fpi);
535 /* Implementing new Reno-2
536 * For Reno-2: $f(x) = U_f(x_f) = \frac{{2}{D_f}}*ln(2+x*D_f)$
537 * Therefore: $fp(x) = 2/(Weight*x + 2)
538 * Therefore: $fpi(x) = (2*Weight)/x - 4
540 double func_reno2_f(const s_lmm_variable_t& var, double x)
542 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
543 return RENO2_SCALING * (1.0 / var.sharing_weight) *
544 log((x * var.sharing_weight) / (2.0 * x * var.sharing_weight + 3.0));
547 double func_reno2_fp(const s_lmm_variable_t& var, double x)
549 return RENO2_SCALING * 3.0 / (var.sharing_weight * x * (2.0 * var.sharing_weight * x + 3.0));
552 double func_reno2_fpi(const s_lmm_variable_t& var, double x)
554 xbt_assert(x > 0.0, "Don't call me with stupid values!");
555 double tmp = x * var.sharing_weight * var.sharing_weight;
556 double res_fpi = tmp * (9.0 * x + 24.0);
561 res_fpi = RENO2_SCALING * (-3.0 * tmp + sqrt(res_fpi)) / (4.0 * tmp);