1 /* Copyright (c) 2007-2018. 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"
18 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_lagrange, surf, "Logging specific to SURF (lagrange)");
19 XBT_LOG_NEW_SUBCATEGORY(surf_lagrange_dichotomy, surf_lagrange, "Logging specific to SURF (lagrange dichotomy)");
21 static constexpr double VEGAS_SCALING = 1000.0;
22 static constexpr double RENO_SCALING = 1.0;
23 static constexpr double RENO2_SCALING = 1.0;
29 double (*func_f_def)(const Variable&, double);
30 double (*func_fp_def)(const Variable&, double);
31 double (*func_fpi_def)(const Variable&, double);
33 System* make_new_lagrange_system(bool selective_update)
35 return new Lagrange(selective_update);
38 bool Lagrange::check_feasible(bool warn)
40 for (Constraint const& cnst : active_constraint_set) {
42 for (Element const& elem : cnst.enabled_element_set) {
43 Variable* var = elem.variable;
44 xbt_assert(var->sharing_weight > 0);
48 if (double_positive(tmp - cnst.bound, sg_maxmin_precision)) {
50 XBT_WARN("The link (%p) is over-used. Expected less than %f and got %f", &cnst, cnst.bound, tmp);
53 XBT_DEBUG("Checking feasability for constraint (%p): sat = %f, lambda = %f ", &cnst, tmp - cnst.bound, cnst.lambda);
56 for (Variable const& var : variable_set) {
57 if (not var.sharing_weight)
61 XBT_DEBUG("Checking feasability for variable (%p): sat = %f mu = %f", &var, var.value - var.bound, var.mu);
63 if (double_positive(var.value - var.bound, sg_maxmin_precision)) {
65 XBT_WARN("The variable (%p) is too large. Expected less than %f and got %f", &var, var.bound, var.value);
72 double Lagrange::new_value(const Variable& var)
76 for (Element const& elem : var.cnsts) {
77 tmp += elem.constraint->lambda;
81 XBT_DEBUG("\t Working on var (%p). cost = %e; Weight = %e", &var, tmp, var.sharing_weight);
82 // uses the partial differential inverse function
83 return var.func_fpi(var, tmp);
86 double Lagrange::new_mu(const Variable& var)
91 for (Element const& elem : var.cnsts) {
92 sigma_i += elem.constraint->lambda;
94 mu_i = var.func_fp(var, var.bound) - sigma_i;
100 double Lagrange::dual_objective()
104 for (Variable const& var : variable_set) {
105 double sigma_i = 0.0;
107 if (not var.sharing_weight)
110 for (Element const& elem : var.cnsts)
111 sigma_i += elem.constraint->lambda;
116 XBT_DEBUG("var %p : sigma_i = %1.20f", &var, sigma_i);
118 obj += var.func_f(var, var.func_fpi(var, sigma_i)) - sigma_i * var.func_fpi(var, sigma_i);
121 obj += var.mu * var.bound;
124 for (Constraint const& cnst : active_constraint_set)
125 obj += cnst.lambda * cnst.bound;
130 // solves the proportional fairness using a Lagrangian optimization with dichotomy step
131 void Lagrange::lagrange_solve()
133 /* Lagrange Variables. */
134 int max_iterations = 100;
135 double epsilon_min_error = 0.00001; /* this is the precision on the objective function so it's none of the
136 configurable values and this value is the legacy one */
137 double dichotomy_min_error = 1e-14;
138 double overall_modification = 1;
140 XBT_DEBUG("Iterative method configuration snapshot =====>");
141 XBT_DEBUG("#### Maximum number of iterations : %d", max_iterations);
142 XBT_DEBUG("#### Minimum error tolerated : %e", epsilon_min_error);
143 XBT_DEBUG("#### Minimum error tolerated (dichotomy) : %e", dichotomy_min_error);
145 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
152 /* Initialize lambda. */
153 auto& cnst_list = active_constraint_set;
154 for (Constraint& cnst : cnst_list) {
156 cnst.new_lambda = 2.0;
157 XBT_DEBUG("#### cnst(%p)->lambda : %e", &cnst, cnst.lambda);
161 * Initialize the var_list variable with only the active variables. Initialize mu.
163 auto& var_list = variable_set;
164 for (Variable& var : var_list) {
165 if (not var.sharing_weight)
168 if (var.bound < 0.0) {
169 XBT_DEBUG("#### NOTE var(%p) is a boundless variable", &var);
175 var.value = new_value(var);
176 XBT_DEBUG("#### var(%p) ->weight : %e", &var, var.sharing_weight);
177 XBT_DEBUG("#### var(%p) ->mu : %e", &var, var.mu);
178 XBT_DEBUG("#### var(%p) ->weight: %e", &var, var.sharing_weight);
179 XBT_DEBUG("#### var(%p) ->bound: %e", &var, var.bound);
181 std::find_if(begin(var.cnsts), end(var.cnsts), [](Element const& x) { return x.consumption_weight != 0.0; });
182 if (weighted == end(var.cnsts))
187 /* Compute dual objective. */
188 double obj = dual_objective();
190 /* While doesn't reach a minimum error or a number maximum of iterations. */
192 while (overall_modification > epsilon_min_error && iteration < max_iterations) {
194 XBT_DEBUG("************** ITERATION %d **************", iteration);
195 XBT_DEBUG("-------------- Gradient Descent ----------");
197 /* Improve the value of mu_i */
198 for (Variable& var : var_list) {
199 if (var.sharing_weight && var.bound >= 0) {
200 XBT_DEBUG("Working on var (%p)", &var);
201 var.new_mu = new_mu(var);
202 XBT_DEBUG("Updating mu : var->mu (%p) : %1.20f -> %1.20f", &var, var.mu, var.new_mu);
205 double new_obj = dual_objective();
206 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
207 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
212 /* Improve the value of lambda_i */
213 for (Constraint& cnst : cnst_list) {
214 XBT_DEBUG("Working on cnst (%p)", &cnst);
215 cnst.new_lambda = dichotomy(cnst.lambda, partial_diff_lambda, cnst, dichotomy_min_error);
216 XBT_DEBUG("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", &cnst, cnst.lambda, cnst.new_lambda);
217 cnst.lambda = cnst.new_lambda;
219 double new_obj = dual_objective();
220 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
221 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
225 /* Now computes the values of each variable (\rho) based on the values of \lambda and \mu. */
226 XBT_DEBUG("-------------- Check convergence ----------");
227 overall_modification = 0;
228 for (Variable& var : var_list) {
229 if (var.sharing_weight <= 0)
232 double tmp = new_value(var);
234 overall_modification = std::max(overall_modification, fabs(var.value - tmp));
237 XBT_DEBUG("New value of var (%p) = %e, overall_modification = %e", &var, var.value, overall_modification);
241 XBT_DEBUG("-------------- Check feasability ----------");
242 if (not check_feasible(false))
243 overall_modification = 1.0;
244 XBT_DEBUG("Iteration %d: overall_modification : %f", iteration, overall_modification);
247 check_feasible(true);
249 if (overall_modification <= epsilon_min_error) {
250 XBT_DEBUG("The method converges in %d iterations.", iteration);
252 if (iteration >= max_iterations) {
253 XBT_DEBUG("Method reach %d iterations, which is the maximum number of iterations allowed.", iteration);
256 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
262 * Returns a double value corresponding to the result of a dichotomy process with respect to a given
263 * variable/constraint (\mu in the case of a variable or \lambda in case of a constraint) and a initial value init.
265 * @param init initial value for \mu or \lambda
266 * @param diff a function that computes the differential of with respect a \mu or \lambda
267 * @param var_cnst a pointer to a variable or constraint
268 * @param min_erro a minimum error tolerated
270 * @return a double corresponding to the result of the dichotomy process
272 double Lagrange::dichotomy(double init, double diff(double, const Constraint&), const Constraint& cnst,
277 double overall_error;
284 if (fabs(init) < 1e-20) {
291 diff_0 = diff(1e-16, cnst);
293 XBT_CDEBUG(surf_lagrange_dichotomy, "returning 0.0 (diff = %e)", diff_0);
298 double min_diff = diff(min, cnst);
299 double max_diff = diff(max, cnst);
301 while (overall_error > min_error) {
302 XBT_CDEBUG(surf_lagrange_dichotomy, "[min, max] = [%1.20f, %1.20f] || diffmin, diffmax = %1.20f, %1.20f", min, max,
305 if (min_diff > 0 && max_diff > 0) {
307 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing min");
309 min_diff = diff(min, cnst);
311 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
315 } else if (min_diff < 0 && max_diff < 0) {
317 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing max");
319 max_diff = diff(max, cnst);
321 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
325 } else if (min_diff < 0 && max_diff > 0) {
326 middle = (max + min) / 2.0;
327 XBT_CDEBUG(surf_lagrange_dichotomy, "Trying (max+min)/2 : %1.20f", middle);
329 if ((fabs(min - middle) < 1e-20) || (fabs(max - middle) < 1e-20)) {
330 XBT_CWARN(surf_lagrange_dichotomy,
331 "Cannot improve the convergence! min=max=middle=%1.20f, diff = %1.20f."
332 " Reaching the 'double' limits. Maybe scaling your function would help ([%1.20f,%1.20f]).",
333 min, max - min, min_diff, max_diff);
336 middle_diff = diff(middle, cnst);
338 if (middle_diff < 0) {
339 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
341 overall_error = max_diff - middle_diff;
342 min_diff = middle_diff;
343 } else if (middle_diff > 0) {
344 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
346 overall_error = max_diff - middle_diff;
347 max_diff = middle_diff;
351 } else if (fabs(min_diff) < 1e-20) {
354 } else if (fabs(max_diff) < 1e-20) {
357 } else if (min_diff > 0 && max_diff < 0) {
358 XBT_CWARN(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0");
361 XBT_CWARN(surf_lagrange_dichotomy,
362 "diffmin (%1.20f) or diffmax (%1.20f) are something I don't know, taking no action.", min_diff,
368 XBT_CDEBUG(surf_lagrange_dichotomy, "returning %e", (min + max) / 2.0);
370 return ((min + max) / 2.0);
373 double Lagrange::partial_diff_lambda(double lambda, const Constraint& cnst)
379 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing diff of cnst (%p)", &cnst);
381 for (Element const& elem : cnst.enabled_element_set) {
382 Variable& var = *elem.variable;
383 xbt_assert(var.sharing_weight > 0);
384 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing sigma_i for var (%p)", &var);
385 // Initialize the summation variable
386 double sigma_i = 0.0;
389 for (Element const& elem2 : var.cnsts)
390 sigma_i += elem2.constraint->lambda;
392 // add mu_i if this flow has a RTT constraint associated
396 // replace value of cnst.lambda by the value of parameter lambda
397 sigma_i = (sigma_i - cnst.lambda) + lambda;
399 diff += -var.func_fpi(var, sigma_i);
404 XBT_CDEBUG(surf_lagrange_dichotomy, "d D/d lambda for cnst (%p) at %1.20f = %1.20f", &cnst, lambda, diff);
409 /** \brief Attribute the value bound to var->bound.
411 * \param func_fpi inverse of the partial differential of f (f prime inverse, (f')^{-1})
413 * Set default functions to the ones passed as parameters. This is a polymorphism in C pure, enjoy the roots of
417 void set_default_protocol_function(double (*func_f)(const Variable& var, double x),
418 double (*func_fp)(const Variable& var, double x),
419 double (*func_fpi)(const Variable& var, double x))
422 func_fp_def = func_fp;
423 func_fpi_def = func_fpi;
426 /**************** Vegas and Reno functions *************************/
427 /* NOTE for Reno: all functions consider the network coefficient (alpha) equal to 1. */
430 * For Vegas: $f(x) = \alpha D_f\ln(x)$
431 * Therefore: $fp(x) = \frac{\alpha D_f}{x}$
432 * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$
434 double func_vegas_f(const Variable& var, double x)
436 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
437 return VEGAS_SCALING * var.sharing_weight * log(x);
440 double func_vegas_fp(const Variable& var, double x)
442 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
443 return VEGAS_SCALING * var.sharing_weight / x;
446 double func_vegas_fpi(const Variable& var, double x)
448 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
449 return var.sharing_weight / (x / VEGAS_SCALING);
453 * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$
454 * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$
455 * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$
457 double func_reno_f(const Variable& var, double x)
459 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
461 return RENO_SCALING * sqrt(3.0 / 2.0) / var.sharing_weight * atan(sqrt(3.0 / 2.0) * var.sharing_weight * x);
464 double func_reno_fp(const Variable& var, double x)
466 return RENO_SCALING * 3.0 / (3.0 * var.sharing_weight * var.sharing_weight * x * x + 2.0);
469 double func_reno_fpi(const Variable& var, double x)
473 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
474 xbt_assert(x > 0.0, "Don't call me with stupid values!");
476 res_fpi = 1.0 / (var.sharing_weight * var.sharing_weight * (x / RENO_SCALING)) -
477 2.0 / (3.0 * var.sharing_weight * var.sharing_weight);
480 return sqrt(res_fpi);
483 /* Implementing new Reno-2
484 * For Reno-2: $f(x) = U_f(x_f) = \frac{{2}{D_f}}*ln(2+x*D_f)$
485 * Therefore: $fp(x) = 2/(Weight*x + 2)
486 * Therefore: $fpi(x) = (2*Weight)/x - 4
488 double func_reno2_f(const Variable& var, double x)
490 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
491 return RENO2_SCALING * (1.0 / var.sharing_weight) *
492 log((x * var.sharing_weight) / (2.0 * x * var.sharing_weight + 3.0));
495 double func_reno2_fp(const Variable& var, double x)
497 return RENO2_SCALING * 3.0 / (var.sharing_weight * x * (2.0 * var.sharing_weight * x + 3.0));
500 double func_reno2_fpi(const Variable& var, double x)
502 xbt_assert(x > 0.0, "Don't call me with stupid values!");
503 double tmp = x * var.sharing_weight * var.sharing_weight;
504 double res_fpi = tmp * (9.0 * x + 24.0);
509 res_fpi = RENO2_SCALING * (-3.0 * tmp + sqrt(res_fpi)) / (4.0 * tmp);