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 "surf/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
31 double (*func_f_def)(lmm_variable_t, double);
32 double (*func_fp_def)(lmm_variable_t, double);
33 double (*func_fpi_def)(lmm_variable_t, double);
36 * Local prototypes to implement the Lagrangian optimization with optimal step, also called dichotomy.
38 // solves the proportional fairness using a Lagrangian optimization with dichotomy step
39 void lagrange_solve(lmm_system_t sys);
40 // computes the value of the dichotomy using a initial values, init, with a specific variable or constraint
41 static double dichotomy(double init, double diff(double, void*), void* var_cnst, double min_error);
42 // computes the value of the differential of constraint param_cnst applied to lambda
43 static double partial_diff_lambda(double lambda, void* param_cnst);
45 static int __check_feasible(xbt_swag_t cnst_list, xbt_swag_t var_list, int warn)
50 xbt_swag_t elem_list = nullptr;
51 lmm_element_t elem = nullptr;
52 lmm_constraint_t cnst = nullptr;
53 lmm_variable_t var = nullptr;
55 xbt_swag_foreach(_cnst, cnst_list)
57 cnst = static_cast<lmm_constraint_t>(_cnst);
59 elem_list = &(cnst->enabled_element_set);
60 xbt_swag_foreach(_elem, elem_list)
62 elem = static_cast<lmm_element_t>(_elem);
64 xbt_assert(var->sharing_weight > 0);
68 if (double_positive(tmp - cnst->bound, sg_maxmin_precision)) {
70 XBT_WARN("The link (%p) is over-used. Expected less than %f and got %f", cnst, cnst->bound, tmp);
73 XBT_DEBUG("Checking feasability for constraint (%p): sat = %f, lambda = %f ", cnst, tmp - cnst->bound,
77 xbt_swag_foreach(_var, var_list)
79 var = static_cast<lmm_variable_t>(_var);
80 if (not var->sharing_weight)
84 XBT_DEBUG("Checking feasability for variable (%p): sat = %f mu = %f", var, var->value - var->bound, var->mu);
86 if (double_positive(var->value - var->bound, sg_maxmin_precision)) {
88 XBT_WARN("The variable (%p) is too large. Expected less than %f and got %f", var, var->bound, var->value);
95 static double new_value(lmm_variable_t var)
99 for (s_lmm_element_t const& elem : var->cnsts) {
100 tmp += elem.constraint->lambda;
104 XBT_DEBUG("\t Working on var (%p). cost = %e; Weight = %e", var, tmp, var->sharing_weight);
105 // uses the partial differential inverse function
106 return var->func_fpi(var, tmp);
109 static double new_mu(lmm_variable_t var)
112 double sigma_i = 0.0;
114 for (s_lmm_element_t const& elem : var->cnsts) {
115 sigma_i += elem.constraint->lambda;
117 mu_i = var->func_fp(var, var->bound) - sigma_i;
123 static double dual_objective(xbt_swag_t var_list, xbt_swag_t cnst_list)
127 lmm_constraint_t cnst = nullptr;
128 lmm_variable_t var = nullptr;
132 xbt_swag_foreach(_var, var_list)
134 var = static_cast<lmm_variable_t>(_var);
135 double sigma_i = 0.0;
137 if (not var->sharing_weight)
140 for (s_lmm_element_t const& elem : var->cnsts)
141 sigma_i += elem.constraint->lambda;
146 XBT_DEBUG("var %p : sigma_i = %1.20f", var, sigma_i);
148 obj += var->func_f(var, var->func_fpi(var, sigma_i)) - sigma_i * var->func_fpi(var, sigma_i);
151 obj += var->mu * var->bound;
154 xbt_swag_foreach(_cnst, cnst_list)
156 cnst = static_cast<lmm_constraint_t>(_cnst);
157 obj += cnst->lambda * cnst->bound;
163 void lagrange_solve(lmm_system_t sys)
165 /* Lagrange Variables. */
166 int max_iterations = 100;
167 double epsilon_min_error = 0.00001; /* this is the precision on the objective function so it's none of the
168 configurable values and this value is the legacy one */
169 double dichotomy_min_error = 1e-14;
170 double overall_modification = 1;
172 XBT_DEBUG("Iterative method configuration snapshot =====>");
173 XBT_DEBUG("#### Maximum number of iterations : %d", max_iterations);
174 XBT_DEBUG("#### Minimum error tolerated : %e", epsilon_min_error);
175 XBT_DEBUG("#### Minimum error tolerated (dichotomy) : %e", dichotomy_min_error);
177 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
181 if (not sys->modified)
184 /* Initialize lambda. */
185 xbt_swag_t cnst_list = &(sys->active_constraint_set);
187 xbt_swag_foreach(_cnst, cnst_list)
189 lmm_constraint_t cnst = (lmm_constraint_t)_cnst;
191 cnst->new_lambda = 2.0;
192 XBT_DEBUG("#### cnst(%p)->lambda : %e", cnst, cnst->lambda);
196 * Initialize the var list variable with only the active variables.
197 * Associate an index in the swag variables. Initialize mu.
199 xbt_swag_t var_list = &(sys->variable_set);
201 xbt_swag_foreach(_var, var_list)
203 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
204 if (not var->sharing_weight)
207 if (var->bound < 0.0) {
208 XBT_DEBUG("#### NOTE var(%p) is a boundless variable", var);
214 var->value = new_value(var);
215 XBT_DEBUG("#### var(%p) ->weight : %e", var, var->sharing_weight);
216 XBT_DEBUG("#### var(%p) ->mu : %e", var, var->mu);
217 XBT_DEBUG("#### var(%p) ->weight: %e", var, var->sharing_weight);
218 XBT_DEBUG("#### var(%p) ->bound: %e", var, var->bound);
219 auto weighted = std::find_if(begin(var->cnsts), end(var->cnsts),
220 [](s_lmm_element_t const& x) { return x.consumption_weight != 0.0; });
221 if (weighted == end(var->cnsts))
226 /* Compute dual objective. */
227 double obj = dual_objective(var_list, cnst_list);
229 /* While doesn't reach a minimum error or a number maximum of iterations. */
231 while (overall_modification > epsilon_min_error && iteration < max_iterations) {
233 XBT_DEBUG("************** ITERATION %d **************", iteration);
234 XBT_DEBUG("-------------- Gradient Descent ----------");
236 /* Improve the value of mu_i */
237 xbt_swag_foreach(_var, var_list)
239 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
240 if (var->sharing_weight && var->bound >= 0) {
241 XBT_DEBUG("Working on var (%p)", var);
242 var->new_mu = new_mu(var);
243 XBT_DEBUG("Updating mu : var->mu (%p) : %1.20f -> %1.20f", var, var->mu, var->new_mu);
244 var->mu = var->new_mu;
246 double new_obj = dual_objective(var_list, cnst_list);
247 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
248 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
253 /* Improve the value of lambda_i */
254 xbt_swag_foreach(_cnst, cnst_list)
256 lmm_constraint_t cnst = static_cast<lmm_constraint_t>(_cnst);
257 XBT_DEBUG("Working on cnst (%p)", cnst);
258 cnst->new_lambda = dichotomy(cnst->lambda, partial_diff_lambda, cnst, dichotomy_min_error);
259 XBT_DEBUG("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", cnst, cnst->lambda, cnst->new_lambda);
260 cnst->lambda = cnst->new_lambda;
262 double new_obj = dual_objective(var_list, cnst_list);
263 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
264 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
268 /* Now computes the values of each variable (\rho) based on the values of \lambda and \mu. */
269 XBT_DEBUG("-------------- Check convergence ----------");
270 overall_modification = 0;
271 xbt_swag_foreach(_var, var_list)
273 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
274 if (var->sharing_weight <= 0)
277 double tmp = new_value(var);
279 overall_modification = std::max(overall_modification, fabs(var->value - tmp));
282 XBT_DEBUG("New value of var (%p) = %e, overall_modification = %e", var, var->value, overall_modification);
286 XBT_DEBUG("-------------- Check feasability ----------");
287 if (not __check_feasible(cnst_list, var_list, 0))
288 overall_modification = 1.0;
289 XBT_DEBUG("Iteration %d: overall_modification : %f", iteration, overall_modification);
292 __check_feasible(cnst_list, var_list, 1);
294 if (overall_modification <= epsilon_min_error) {
295 XBT_DEBUG("The method converges in %d iterations.", iteration);
297 if (iteration >= max_iterations) {
298 XBT_DEBUG("Method reach %d iterations, which is the maximum number of iterations allowed.", iteration);
301 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
307 * Returns a double value corresponding to the result of a dichotomy process with respect to a given
308 * variable/constraint (\mu in the case of a variable or \lambda in case of a constraint) and a initial value init.
310 * @param init initial value for \mu or \lambda
311 * @param diff a function that computes the differential of with respect a \mu or \lambda
312 * @param var_cnst a pointer to a variable or constraint
313 * @param min_erro a minimum error tolerated
315 * @return a double corresponding to the result of the dichotomy process
317 static double dichotomy(double init, double diff(double, void*), void* var_cnst, double min_error)
321 double overall_error;
328 if (fabs(init) < 1e-20) {
335 diff_0 = diff(1e-16, var_cnst);
337 XBT_CDEBUG(surf_lagrange_dichotomy, "returning 0.0 (diff = %e)", diff_0);
342 double min_diff = diff(min, var_cnst);
343 double max_diff = diff(max, var_cnst);
345 while (overall_error > min_error) {
346 XBT_CDEBUG(surf_lagrange_dichotomy, "[min, max] = [%1.20f, %1.20f] || diffmin, diffmax = %1.20f, %1.20f", min, max,
349 if (min_diff > 0 && max_diff > 0) {
351 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing min");
353 min_diff = diff(min, var_cnst);
355 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
359 } else if (min_diff < 0 && max_diff < 0) {
361 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing max");
363 max_diff = diff(max, var_cnst);
365 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
369 } else if (min_diff < 0 && max_diff > 0) {
370 middle = (max + min) / 2.0;
371 XBT_CDEBUG(surf_lagrange_dichotomy, "Trying (max+min)/2 : %1.20f", middle);
373 if ((fabs(min - middle) < 1e-20) || (fabs(max - middle) < 1e-20)) {
374 XBT_CWARN(surf_lagrange_dichotomy,
375 "Cannot improve the convergence! min=max=middle=%1.20f, diff = %1.20f."
376 " Reaching the 'double' limits. Maybe scaling your function would help ([%1.20f,%1.20f]).",
377 min, max - min, min_diff, max_diff);
380 middle_diff = diff(middle, var_cnst);
382 if (middle_diff < 0) {
383 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
385 overall_error = max_diff - middle_diff;
386 min_diff = middle_diff;
387 } else if (middle_diff > 0) {
388 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
390 overall_error = max_diff - middle_diff;
391 max_diff = middle_diff;
395 } else if (fabs(min_diff) < 1e-20) {
398 } else if (fabs(max_diff) < 1e-20) {
401 } else if (min_diff > 0 && max_diff < 0) {
402 XBT_CWARN(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0");
405 XBT_CWARN(surf_lagrange_dichotomy,
406 "diffmin (%1.20f) or diffmax (%1.20f) are something I don't know, taking no action.", min_diff,
412 XBT_CDEBUG(surf_lagrange_dichotomy, "returning %e", (min + max) / 2.0);
414 return ((min + max) / 2.0);
417 static double partial_diff_lambda(double lambda, void* param_cnst)
419 lmm_constraint_t cnst = static_cast<lmm_constraint_t>(param_cnst);
424 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing diff of cnst (%p)", cnst);
426 xbt_swag_t elem_list = &(cnst->enabled_element_set);
428 xbt_swag_foreach(_elem, elem_list)
430 lmm_element_t elem = static_cast<lmm_element_t>(_elem);
431 lmm_variable_t var = elem->variable;
432 xbt_assert(var->sharing_weight > 0);
433 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing sigma_i for var (%p)", var);
434 // Initialize the summation variable
435 double sigma_i = 0.0;
438 for (s_lmm_element_t const& elem : var->cnsts) {
439 sigma_i += elem.constraint->lambda;
442 // add mu_i if this flow has a RTT constraint associated
446 // replace value of cnst->lambda by the value of parameter lambda
447 sigma_i = (sigma_i - cnst->lambda) + lambda;
449 diff += -var->func_fpi(var, sigma_i);
454 XBT_CDEBUG(surf_lagrange_dichotomy, "d D/d lambda for cnst (%p) at %1.20f = %1.20f", cnst, lambda, diff);
459 /** \brief Attribute the value bound to var->bound.
461 * \param func_fpi inverse of the partial differential of f (f prime inverse, (f')^{-1})
463 * Set default functions to the ones passed as parameters. This is a polymorphism in C pure, enjoy the roots of
467 void lmm_set_default_protocol_function(double (*func_f)(lmm_variable_t var, double x),
468 double (*func_fp)(lmm_variable_t var, double x),
469 double (*func_fpi)(lmm_variable_t var, double x))
472 func_fp_def = func_fp;
473 func_fpi_def = func_fpi;
476 /**************** Vegas and Reno functions *************************/
477 /* NOTE for Reno: all functions consider the network coefficient (alpha) equal to 1. */
480 * For Vegas: $f(x) = \alpha D_f\ln(x)$
481 * Therefore: $fp(x) = \frac{\alpha D_f}{x}$
482 * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$
484 double func_vegas_f(lmm_variable_t var, double x)
486 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
487 return VEGAS_SCALING * var->sharing_weight * log(x);
490 double func_vegas_fp(lmm_variable_t var, double x)
492 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
493 return VEGAS_SCALING * var->sharing_weight / x;
496 double func_vegas_fpi(lmm_variable_t var, double x)
498 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
499 return var->sharing_weight / (x / VEGAS_SCALING);
503 * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$
504 * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$
505 * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$
507 double func_reno_f(lmm_variable_t var, double x)
509 xbt_assert(var->sharing_weight > 0.0, "Don't call me with stupid values!");
511 return RENO_SCALING * sqrt(3.0 / 2.0) / var->sharing_weight * atan(sqrt(3.0 / 2.0) * var->sharing_weight * x);
514 double func_reno_fp(lmm_variable_t var, double x)
516 return RENO_SCALING * 3.0 / (3.0 * var->sharing_weight * var->sharing_weight * x * x + 2.0);
519 double func_reno_fpi(lmm_variable_t var, double x)
523 xbt_assert(var->sharing_weight > 0.0, "Don't call me with stupid values!");
524 xbt_assert(x > 0.0, "Don't call me with stupid values!");
526 res_fpi = 1.0 / (var->sharing_weight * var->sharing_weight * (x / RENO_SCALING)) -
527 2.0 / (3.0 * var->sharing_weight * var->sharing_weight);
530 return sqrt(res_fpi);
533 /* Implementing new Reno-2
534 * For Reno-2: $f(x) = U_f(x_f) = \frac{{2}{D_f}}*ln(2+x*D_f)$
535 * Therefore: $fp(x) = 2/(Weight*x + 2)
536 * Therefore: $fpi(x) = (2*Weight)/x - 4
538 double func_reno2_f(lmm_variable_t var, double x)
540 xbt_assert(var->sharing_weight > 0.0, "Don't call me with stupid values!");
541 return RENO2_SCALING * (1.0 / var->sharing_weight) *
542 log((x * var->sharing_weight) / (2.0 * x * var->sharing_weight + 3.0));
545 double func_reno2_fp(lmm_variable_t var, double x)
547 return RENO2_SCALING * 3.0 / (var->sharing_weight * x * (2.0 * var->sharing_weight * x + 3.0));
550 double func_reno2_fpi(lmm_variable_t var, double x)
552 xbt_assert(x > 0.0, "Don't call me with stupid values!");
553 double tmp = x * var->sharing_weight * var->sharing_weight;
554 double res_fpi = tmp * (9.0 * x + 24.0);
559 res_fpi = RENO2_SCALING * (-3.0 * tmp + sqrt(res_fpi)) / (4.0 * tmp);