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 template <class CnstList, class VarList>
48 static int __check_feasible(const CnstList& cnst_list, const VarList& var_list, int warn)
51 const_xbt_swag_t elem_list = nullptr;
52 lmm_element_t elem = nullptr;
54 for (s_lmm_constraint_t const& cnst : cnst_list) {
56 elem_list = &cnst.enabled_element_set;
57 xbt_swag_foreach(_elem, elem_list)
59 elem = static_cast<lmm_element_t>(_elem);
60 lmm_variable_t var = elem->variable;
61 xbt_assert(var->sharing_weight > 0);
65 if (double_positive(tmp - cnst.bound, sg_maxmin_precision)) {
67 XBT_WARN("The link (%p) is over-used. Expected less than %f and got %f", &cnst, cnst.bound, tmp);
70 XBT_DEBUG("Checking feasability for constraint (%p): sat = %f, lambda = %f ", &cnst, tmp - cnst.bound, cnst.lambda);
73 for (s_lmm_variable_t const& var : var_list) {
74 if (not var.sharing_weight)
78 XBT_DEBUG("Checking feasability for variable (%p): sat = %f mu = %f", &var, var.value - var.bound, var.mu);
80 if (double_positive(var.value - var.bound, sg_maxmin_precision)) {
82 XBT_WARN("The variable (%p) is too large. Expected less than %f and got %f", &var, var.bound, var.value);
89 static double new_value(const s_lmm_variable_t& var)
93 for (s_lmm_element_t const& elem : var.cnsts) {
94 tmp += elem.constraint->lambda;
98 XBT_DEBUG("\t Working on var (%p). cost = %e; Weight = %e", &var, tmp, var.sharing_weight);
99 // uses the partial differential inverse function
100 return var.func_fpi(var, tmp);
103 static double new_mu(const s_lmm_variable_t& var)
106 double sigma_i = 0.0;
108 for (s_lmm_element_t const& elem : var.cnsts) {
109 sigma_i += elem.constraint->lambda;
111 mu_i = var.func_fp(var, var.bound) - sigma_i;
117 template <class VarList, class CnstList>
118 static double dual_objective(const VarList& var_list, const CnstList& cnst_list)
122 for (s_lmm_variable_t const& var : var_list) {
123 double sigma_i = 0.0;
125 if (not var.sharing_weight)
128 for (s_lmm_element_t const& elem : var.cnsts)
129 sigma_i += elem.constraint->lambda;
134 XBT_DEBUG("var %p : sigma_i = %1.20f", &var, sigma_i);
136 obj += var.func_f(var, var.func_fpi(var, sigma_i)) - sigma_i * var.func_fpi(var, sigma_i);
139 obj += var.mu * var.bound;
142 for (s_lmm_constraint_t const& cnst : cnst_list)
143 obj += cnst.lambda * cnst.bound;
148 void lagrange_solve(lmm_system_t sys)
150 /* Lagrange Variables. */
151 int max_iterations = 100;
152 double epsilon_min_error = 0.00001; /* this is the precision on the objective function so it's none of the
153 configurable values and this value is the legacy one */
154 double dichotomy_min_error = 1e-14;
155 double overall_modification = 1;
157 XBT_DEBUG("Iterative method configuration snapshot =====>");
158 XBT_DEBUG("#### Maximum number of iterations : %d", max_iterations);
159 XBT_DEBUG("#### Minimum error tolerated : %e", epsilon_min_error);
160 XBT_DEBUG("#### Minimum error tolerated (dichotomy) : %e", dichotomy_min_error);
162 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
166 if (not sys->modified)
169 /* Initialize lambda. */
170 auto& cnst_list = sys->active_constraint_set;
171 for (s_lmm_constraint_t& cnst : cnst_list) {
173 cnst.new_lambda = 2.0;
174 XBT_DEBUG("#### cnst(%p)->lambda : %e", &cnst, cnst.lambda);
178 * Initialize the var list variable with only the active variables.
179 * Associate an index in the swag variables. Initialize mu.
181 auto& var_list = sys->variable_set;
182 for (s_lmm_variable_t& var : var_list) {
183 if (not var.sharing_weight)
186 if (var.bound < 0.0) {
187 XBT_DEBUG("#### NOTE var(%p) is a boundless variable", &var);
193 var.value = new_value(var);
194 XBT_DEBUG("#### var(%p) ->weight : %e", &var, var.sharing_weight);
195 XBT_DEBUG("#### var(%p) ->mu : %e", &var, var.mu);
196 XBT_DEBUG("#### var(%p) ->weight: %e", &var, var.sharing_weight);
197 XBT_DEBUG("#### var(%p) ->bound: %e", &var, var.bound);
198 auto weighted = std::find_if(begin(var.cnsts), end(var.cnsts),
199 [](s_lmm_element_t const& x) { return x.consumption_weight != 0.0; });
200 if (weighted == end(var.cnsts))
205 /* Compute dual objective. */
206 double obj = dual_objective(var_list, cnst_list);
208 /* While doesn't reach a minimum error or a number maximum of iterations. */
210 while (overall_modification > epsilon_min_error && iteration < max_iterations) {
212 XBT_DEBUG("************** ITERATION %d **************", iteration);
213 XBT_DEBUG("-------------- Gradient Descent ----------");
215 /* Improve the value of mu_i */
216 for (s_lmm_variable_t& var : var_list) {
217 if (var.sharing_weight && var.bound >= 0) {
218 XBT_DEBUG("Working on var (%p)", &var);
219 var.new_mu = new_mu(var);
220 XBT_DEBUG("Updating mu : var->mu (%p) : %1.20f -> %1.20f", &var, var.mu, var.new_mu);
223 double new_obj = dual_objective(var_list, cnst_list);
224 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
225 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
230 /* Improve the value of lambda_i */
231 for (s_lmm_constraint_t& cnst : cnst_list) {
232 XBT_DEBUG("Working on cnst (%p)", &cnst);
233 cnst.new_lambda = dichotomy(cnst.lambda, partial_diff_lambda, cnst, dichotomy_min_error);
234 XBT_DEBUG("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", &cnst, cnst.lambda, cnst.new_lambda);
235 cnst.lambda = cnst.new_lambda;
237 double new_obj = dual_objective(var_list, cnst_list);
238 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
239 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
243 /* Now computes the values of each variable (\rho) based on the values of \lambda and \mu. */
244 XBT_DEBUG("-------------- Check convergence ----------");
245 overall_modification = 0;
246 for (s_lmm_variable_t& var : var_list) {
247 if (var.sharing_weight <= 0)
250 double tmp = new_value(var);
252 overall_modification = std::max(overall_modification, fabs(var.value - tmp));
255 XBT_DEBUG("New value of var (%p) = %e, overall_modification = %e", &var, var.value, overall_modification);
259 XBT_DEBUG("-------------- Check feasability ----------");
260 if (not __check_feasible(cnst_list, var_list, 0))
261 overall_modification = 1.0;
262 XBT_DEBUG("Iteration %d: overall_modification : %f", iteration, overall_modification);
265 __check_feasible(cnst_list, var_list, 1);
267 if (overall_modification <= epsilon_min_error) {
268 XBT_DEBUG("The method converges in %d iterations.", iteration);
270 if (iteration >= max_iterations) {
271 XBT_DEBUG("Method reach %d iterations, which is the maximum number of iterations allowed.", iteration);
274 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
280 * Returns a double value corresponding to the result of a dichotomy process with respect to a given
281 * variable/constraint (\mu in the case of a variable or \lambda in case of a constraint) and a initial value init.
283 * @param init initial value for \mu or \lambda
284 * @param diff a function that computes the differential of with respect a \mu or \lambda
285 * @param var_cnst a pointer to a variable or constraint
286 * @param min_erro a minimum error tolerated
288 * @return a double corresponding to the result of the dichotomy process
290 static double dichotomy(double init, double diff(double, const s_lmm_constraint_t&), const s_lmm_constraint_t& cnst,
295 double overall_error;
302 if (fabs(init) < 1e-20) {
309 diff_0 = diff(1e-16, cnst);
311 XBT_CDEBUG(surf_lagrange_dichotomy, "returning 0.0 (diff = %e)", diff_0);
316 double min_diff = diff(min, cnst);
317 double max_diff = diff(max, cnst);
319 while (overall_error > min_error) {
320 XBT_CDEBUG(surf_lagrange_dichotomy, "[min, max] = [%1.20f, %1.20f] || diffmin, diffmax = %1.20f, %1.20f", min, max,
323 if (min_diff > 0 && max_diff > 0) {
325 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing min");
327 min_diff = diff(min, cnst);
329 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
333 } else if (min_diff < 0 && max_diff < 0) {
335 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing max");
337 max_diff = diff(max, cnst);
339 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
343 } else if (min_diff < 0 && max_diff > 0) {
344 middle = (max + min) / 2.0;
345 XBT_CDEBUG(surf_lagrange_dichotomy, "Trying (max+min)/2 : %1.20f", middle);
347 if ((fabs(min - middle) < 1e-20) || (fabs(max - middle) < 1e-20)) {
348 XBT_CWARN(surf_lagrange_dichotomy,
349 "Cannot improve the convergence! min=max=middle=%1.20f, diff = %1.20f."
350 " Reaching the 'double' limits. Maybe scaling your function would help ([%1.20f,%1.20f]).",
351 min, max - min, min_diff, max_diff);
354 middle_diff = diff(middle, cnst);
356 if (middle_diff < 0) {
357 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
359 overall_error = max_diff - middle_diff;
360 min_diff = middle_diff;
361 } else if (middle_diff > 0) {
362 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
364 overall_error = max_diff - middle_diff;
365 max_diff = middle_diff;
369 } else if (fabs(min_diff) < 1e-20) {
372 } else if (fabs(max_diff) < 1e-20) {
375 } else if (min_diff > 0 && max_diff < 0) {
376 XBT_CWARN(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0");
379 XBT_CWARN(surf_lagrange_dichotomy,
380 "diffmin (%1.20f) or diffmax (%1.20f) are something I don't know, taking no action.", min_diff,
386 XBT_CDEBUG(surf_lagrange_dichotomy, "returning %e", (min + max) / 2.0);
388 return ((min + max) / 2.0);
391 static double partial_diff_lambda(double lambda, const s_lmm_constraint_t& cnst)
397 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing diff of cnst (%p)", &cnst);
399 const_xbt_swag_t elem_list = &cnst.enabled_element_set;
401 xbt_swag_foreach(_elem, elem_list)
403 lmm_element_t elem = static_cast<lmm_element_t>(_elem);
404 s_lmm_variable_t& var = *elem->variable;
405 xbt_assert(var.sharing_weight > 0);
406 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing sigma_i for var (%p)", &var);
407 // Initialize the summation variable
408 double sigma_i = 0.0;
411 for (s_lmm_element_t const& elem : var.cnsts) {
412 sigma_i += elem.constraint->lambda;
415 // add mu_i if this flow has a RTT constraint associated
419 // replace value of cnst.lambda by the value of parameter lambda
420 sigma_i = (sigma_i - cnst.lambda) + lambda;
422 diff += -var.func_fpi(var, sigma_i);
427 XBT_CDEBUG(surf_lagrange_dichotomy, "d D/d lambda for cnst (%p) at %1.20f = %1.20f", &cnst, lambda, diff);
432 /** \brief Attribute the value bound to var->bound.
434 * \param func_fpi inverse of the partial differential of f (f prime inverse, (f')^{-1})
436 * Set default functions to the ones passed as parameters. This is a polymorphism in C pure, enjoy the roots of
440 void lmm_set_default_protocol_function(double (*func_f)(const s_lmm_variable_t& var, double x),
441 double (*func_fp)(const s_lmm_variable_t& var, double x),
442 double (*func_fpi)(const s_lmm_variable_t& var, double x))
445 func_fp_def = func_fp;
446 func_fpi_def = func_fpi;
449 /**************** Vegas and Reno functions *************************/
450 /* NOTE for Reno: all functions consider the network coefficient (alpha) equal to 1. */
453 * For Vegas: $f(x) = \alpha D_f\ln(x)$
454 * Therefore: $fp(x) = \frac{\alpha D_f}{x}$
455 * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$
457 double func_vegas_f(const s_lmm_variable_t& var, double x)
459 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
460 return VEGAS_SCALING * var.sharing_weight * log(x);
463 double func_vegas_fp(const s_lmm_variable_t& var, double x)
465 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
466 return VEGAS_SCALING * var.sharing_weight / x;
469 double func_vegas_fpi(const s_lmm_variable_t& var, double x)
471 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
472 return var.sharing_weight / (x / VEGAS_SCALING);
476 * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$
477 * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$
478 * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$
480 double func_reno_f(const s_lmm_variable_t& var, double x)
482 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
484 return RENO_SCALING * sqrt(3.0 / 2.0) / var.sharing_weight * atan(sqrt(3.0 / 2.0) * var.sharing_weight * x);
487 double func_reno_fp(const s_lmm_variable_t& var, double x)
489 return RENO_SCALING * 3.0 / (3.0 * var.sharing_weight * var.sharing_weight * x * x + 2.0);
492 double func_reno_fpi(const s_lmm_variable_t& var, double x)
496 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
497 xbt_assert(x > 0.0, "Don't call me with stupid values!");
499 res_fpi = 1.0 / (var.sharing_weight * var.sharing_weight * (x / RENO_SCALING)) -
500 2.0 / (3.0 * var.sharing_weight * var.sharing_weight);
503 return sqrt(res_fpi);
506 /* Implementing new Reno-2
507 * For Reno-2: $f(x) = U_f(x_f) = \frac{{2}{D_f}}*ln(2+x*D_f)$
508 * Therefore: $fp(x) = 2/(Weight*x + 2)
509 * Therefore: $fpi(x) = (2*Weight)/x - 4
511 double func_reno2_f(const s_lmm_variable_t& var, double x)
513 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
514 return RENO2_SCALING * (1.0 / var.sharing_weight) *
515 log((x * var.sharing_weight) / (2.0 * x * var.sharing_weight + 3.0));
518 double func_reno2_fp(const s_lmm_variable_t& var, double x)
520 return RENO2_SCALING * 3.0 / (var.sharing_weight * x * (2.0 * var.sharing_weight * x + 3.0));
523 double func_reno2_fpi(const s_lmm_variable_t& var, double x)
525 xbt_assert(x > 0.0, "Don't call me with stupid values!");
526 double tmp = x * var.sharing_weight * var.sharing_weight;
527 double res_fpi = tmp * (9.0 * x + 24.0);
532 res_fpi = RENO2_SCALING * (-3.0 * tmp + sqrt(res_fpi)) / (4.0 * tmp);