From c9ac9b66d4880cbbc9aebad711b3288701577fa5 Mon Sep 17 00:00:00 2001 From: alegrand Date: Thu, 26 Jul 2007 18:57:43 +0000 Subject: [PATCH] Many bug fix: * func_fp can be used to compute new_mu without having to resort to a dichotomy. So let's ask users to provide it. * func_f can be used to compute the dual functionnal (d), that is useful to check the convergence of the gradient. So let's ask users to provide it. * bug fix in the dichotomy. The stopping condition on overall_error was wrong. * Compute the dual (d) to check the convergence. * BUG FIX in the RENO functions. We need to write 3.0/2.0 and not 3/2..... It took me a while to find this nasty one! :( git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/simgrid/simgrid/trunk@3895 48e7efb5-ca39-0410-a469-dd3cf9ba447f --- src/surf/lagrange.c | 256 ++++++++++++++++++++++++++++---------------- 1 file changed, 162 insertions(+), 94 deletions(-) diff --git a/src/surf/lagrange.c b/src/surf/lagrange.c index 41b022a663..caedc9dfd5 100644 --- a/src/surf/lagrange.c +++ b/src/surf/lagrange.c @@ -22,6 +22,12 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_lagrange, surf, XBT_LOG_NEW_SUBCATEGORY(surf_lagrange_dichotomy, surf_lagrange, "Logging specific to SURF (lagrange dichotomy)"); +#define SHOW_EXPR(expr) CDEBUG1(surf_lagrange,#expr " = %g",expr); + +double (* func_f_def ) (lmm_variable_t , double); +double (* func_fp_def ) (lmm_variable_t , double); +double (* func_fpi_def )(lmm_variable_t , double); + /* * Local prototypes to implement the lagrangian optimization with optimal step, also called dichotomy. */ @@ -34,9 +40,6 @@ static double dichotomy(double init, double diff(double, void *), void *var_cnst static double partial_diff_mu(double mu, void *param_var); //computes the value of the differential of constraint param_cnst applied to lambda static double partial_diff_lambda(double lambda, void *param_cnst); -//auxiliar function to compute the partial_diff -static double diff_aux(lmm_variable_t var, double x); - static int __check_feasible(xbt_swag_t cnst_list, xbt_swag_t var_list, int warn) { @@ -85,6 +88,68 @@ static int __check_feasible(xbt_swag_t cnst_list, xbt_swag_t var_list, int warn) return 1; } +static double new_value(lmm_variable_t var) +{ + double tmp = 0; + int i; + + for (i = 0; i < var->cnsts_number; i++) { + tmp += (var->cnsts[i].constraint)->lambda; + } + if (var->bound > 0) + tmp += var->mu; + DEBUG3("\t Working on var (%p). cost = %e; Df = %e", var, tmp, + var->df); + //uses the partial differential inverse function + return var->func_fpi(var, tmp); +} + +static double new_mu(lmm_variable_t var) +{ + double mu_i = 0.0; + double sigma_i = 0.0; + int j; + + for (j = 0; j < var->cnsts_number; j++) { + sigma_i += (var->cnsts[j].constraint)->lambda; + } + mu_i = var->func_fp(var,var->bound)-sigma_i; + if(mu_i<0.0) return 0.0; + return mu_i; +} + +static double dual_objective(xbt_swag_t var_list, xbt_swag_t cnst_list) +{ + lmm_constraint_t cnst = NULL; + lmm_variable_t var = NULL; + + double obj = 0.0; + + xbt_swag_foreach(var, var_list) { + double sigma_i=0.0; + int j; + + for (j = 0; j < var->cnsts_number; j++) + sigma_i += (var->cnsts[j].constraint)->lambda; + + if (var->bound > 0) + sigma_i += var->mu; + + DEBUG2("var %p : sigma_i = %1.20f",var,sigma_i); + + obj += var->func_f(var,var->func_fpi(var,sigma_i)) - + sigma_i*var->func_fpi(var,sigma_i); + + if (var->bound > 0) + obj += var->mu*var->bound; + } + + xbt_swag_foreach(cnst, cnst_list) + obj += cnst->lambda*cnst->bound; + + return obj; +} + void lagrange_solve(lmm_system_t sys) { /* @@ -92,7 +157,7 @@ void lagrange_solve(lmm_system_t sys) */ int max_iterations = 100; double epsilon_min_error = MAXMIN_PRECISION; - double dichotomy_min_error = 1e-18; + double dichotomy_min_error = 1e-14; double overall_modification = 1; /* @@ -111,7 +176,7 @@ void lagrange_solve(lmm_system_t sys) int iteration = 0; double tmp = 0; int i; - + double obj,new_obj; DEBUG0("Iterative method configuration snapshot =====>"); DEBUG1("#### Maximum number of iterations : %d", max_iterations); @@ -123,6 +188,17 @@ void lagrange_solve(lmm_system_t sys) if (!(sys->modified)) return; + + /* + * Initialize lambda. + */ + cnst_list = &(sys->active_constraint_set); + xbt_swag_foreach(cnst, cnst_list) { + cnst->lambda = 1.0; + cnst->new_lambda = 2.0; + DEBUG2("#### cnst(%p)->lambda : %e", cnst, cnst->lambda); + } + /* * Initialize the var list variable with only the active variables. * Associate an index in the swag variables. Initialize mu. @@ -133,9 +209,14 @@ void lagrange_solve(lmm_system_t sys) if ((var->bound < 0.0) || (var->weight <= 0.0)) { DEBUG1("#### NOTE var(%d) is a boundless (or inactive) variable", i); var->mu = -1.0; + if(var->weight>0.0) + var->value = new_value(var); + else + var->value = 0; } else { var->mu = 1.0; var->new_mu = 2.0; + var->value = new_value(var); } DEBUG3("#### var(%d) %p ->mu : %e", i, var, var->mu); DEBUG3("#### var(%d) %p ->weight: %e", i, var, var->weight); @@ -144,14 +225,9 @@ void lagrange_solve(lmm_system_t sys) } /* - * Initialize lambda. + * Compute dual objective. */ - cnst_list = &(sys->active_constraint_set); - xbt_swag_foreach(cnst, cnst_list) { - cnst->lambda = 1.0; - cnst->new_lambda = 2.0; - DEBUG2("#### cnst(%p)->lambda : %e", cnst, cnst->lambda); - } + obj = dual_objective(var_list,cnst_list); /* * While doesn't reach a minimun error or a number maximum of iterations. @@ -162,26 +238,30 @@ void lagrange_solve(lmm_system_t sys) iteration++; DEBUG1("************** ITERATION %d **************", iteration); DEBUG0("-------------- Gradient Descent ----------"); + /* - * Compute the value of mu_i + * Improve the value of mu_i */ - //forall mu_i in mu_1, mu_2, ..., mu_n xbt_swag_foreach(var, var_list) { if ((var->bound >= 0) && (var->weight > 0)) { DEBUG1("Working on var (%p)", var); - var->new_mu = - dichotomy(var->mu, partial_diff_mu, var, dichotomy_min_error); + var->new_mu = new_mu(var); dual_updated += (fabs(var->new_mu-var->mu)>dichotomy_min_error); DEBUG2("dual_updated (%d) : %1.20f",dual_updated,fabs(var->new_mu-var->mu)); DEBUG3("Updating mu : var->mu (%p) : %1.20f -> %1.20f", var, var->mu, var->new_mu); var->mu = var->new_mu; + + new_obj=dual_objective(var_list,cnst_list); + DEBUG3("Improvement for Objective (%g -> %g) : %g", obj, new_obj, + obj-new_obj); + xbt_assert1(obj-new_obj>=-epsilon_min_error,"Our gradient sucks! (%1.20f)",obj-new_obj); + obj = new_obj; } } /* - * Compute the value of lambda_i + * Improve the value of lambda_i */ - //forall lambda_i in lambda_1, lambda_2, ..., lambda_n xbt_swag_foreach(cnst, cnst_list) { DEBUG1("Working on cnst (%p)", cnst); cnst->new_lambda = @@ -191,6 +271,12 @@ void lagrange_solve(lmm_system_t sys) DEBUG2("dual_updated (%d) : %1.20f",dual_updated,fabs(cnst->new_lambda-cnst->lambda)); DEBUG3("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", cnst, cnst->lambda, cnst->new_lambda); cnst->lambda = cnst->new_lambda; + + new_obj=dual_objective(var_list,cnst_list); + DEBUG3("Improvement for Objective (%g -> %g) : %g", obj, new_obj, + obj-new_obj); + xbt_assert1(obj-new_obj>=-epsilon_min_error,"Our gradient sucks! (%1.20f)",obj-new_obj); + obj = new_obj; } /* @@ -203,18 +289,7 @@ void lagrange_solve(lmm_system_t sys) if (var->weight <= 0) var->value = 0.0; else { - //compute sigma_i + mu_i - tmp = 0; - for (i = 0; i < var->cnsts_number; i++) { - tmp += (var->cnsts[i].constraint)->lambda; - } - if (var->bound > 0) - tmp += var->mu; - DEBUG3("\t Working on var (%p). cost = %e; Df = %e", var, tmp, - var->df); - - //uses the partial differential inverse function - tmp = var->func_fpi(var, tmp); + tmp = new_value(var); if (overall_modification < (fabs(var->value - tmp)/tmp)) { overall_modification = (fabs(var->value - tmp)/tmp); @@ -226,6 +301,7 @@ void lagrange_solve(lmm_system_t sys) } } + DEBUG0("-------------- Check feasability ----------"); if (!__check_feasible(cnst_list, var_list, 0)) overall_modification = 1.0; DEBUG2("Iteration %d: overall_modification : %f", iteration, overall_modification); @@ -233,6 +309,10 @@ void lagrange_solve(lmm_system_t sys) WARN1("Could not improve the convergence at iteration %d. Drop it!",iteration); break; } + + /* + * Compute dual objective. + */ } @@ -308,7 +388,6 @@ static double dichotomy(double init, double diff(double, void *), void *var_cnst CDEBUG0(surf_lagrange_dichotomy, "Decreasing max"); max = min; max_diff = min_diff; - } } else if (min_diff < 0 && max_diff < 0) { if (min == max) { @@ -335,22 +414,27 @@ static double dichotomy(double init, double diff(double, void *), void *var_cnst if (middle_diff < 0) { CDEBUG0(surf_lagrange_dichotomy, "Increasing min"); min = middle; - min_diff = middle_diff; overall_error = max_diff-middle_diff; + min_diff = middle_diff; +/* SHOW_EXPR(overall_error); */ } else if (middle_diff > 0) { CDEBUG0(surf_lagrange_dichotomy, "Decreasing max"); max = middle; - max_diff = middle_diff; overall_error = max_diff-middle_diff; + max_diff = middle_diff; +/* SHOW_EXPR(overall_error); */ } else { overall_error = 0; +/* SHOW_EXPR(overall_error); */ } } else if (min_diff == 0) { max=min; overall_error = 0; +/* SHOW_EXPR(overall_error); */ } else if (max_diff == 0) { min=max; overall_error = 0; +/* SHOW_EXPR(overall_error); */ } else if (min_diff > 0 && max_diff < 0) { CWARN0(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0"); @@ -368,63 +452,34 @@ static double dichotomy(double init, double diff(double, void *), void *var_cnst return ((min + max) / 2.0); } -/* - * - */ -static double partial_diff_mu(double mu, void *param_var) -{ - double mu_partial = 0.0; - double sigma_mu = 0.0; - lmm_variable_t var = (lmm_variable_t) param_var; - int i; - XBT_IN; - //compute sigma_i - for (i = 0; i < var->cnsts_number; i++) - sigma_mu += (var->cnsts[i].constraint)->lambda; - - //compute sigma_i + mu_i - sigma_mu += mu; - - //use auxiliar function passing (sigma_i + mu_i) - mu_partial = diff_aux(var, sigma_mu); - - //add the RTT limit - mu_partial += var->bound; - - XBT_OUT; - return mu_partial; -} - -/* - * - */ static double partial_diff_lambda(double lambda, void *param_cnst) { - int i; + int j; xbt_swag_t elem_list = NULL; lmm_element_t elem = NULL; lmm_variable_t var = NULL; lmm_constraint_t cnst = (lmm_constraint_t) param_cnst; - double lambda_partial = 0.0; + double diff = 0.0; double sigma_i = 0.0; XBT_IN; elem_list = &(cnst->element_set); - CDEBUG1(surf_lagrange_dichotomy,"Computting diff of cnst (%p)", cnst); + CDEBUG1(surf_lagrange_dichotomy,"Computing diff of cnst (%p)", cnst); xbt_swag_foreach(elem, elem_list) { var = elem->variable; if (var->weight <= 0) continue; - //initilize de sumation variable + CDEBUG1(surf_lagrange_dichotomy,"Computing sigma_i for var (%p)", var); + // Initialize the summation variable sigma_i = 0.0; - //compute sigma_i of variable var - for (i = 0; i < var->cnsts_number; i++) { - sigma_i += (var->cnsts[i].constraint)->lambda; + // Compute sigma_i + for (j = 0; j < var->cnsts_number; j++) { + sigma_i += (var->cnsts[j].constraint)->lambda; } //add mu_i if this flow has a RTT constraint associated @@ -434,31 +489,16 @@ static double partial_diff_lambda(double lambda, void *param_cnst) //replace value of cnst->lambda by the value of parameter lambda sigma_i = (sigma_i - cnst->lambda) + lambda; - //use the auxiliar function passing (\sigma_i + \mu_i) - lambda_partial += diff_aux(var, sigma_i); + diff += -var->func_fpi(var, sigma_i); } - lambda_partial += cnst->bound; + diff += cnst->bound; + CDEBUG3(surf_lagrange_dichotomy,"d D/d lambda for cnst (%p) at %1.20f = %1.20f", + cnst, lambda, diff); XBT_OUT; - return lambda_partial; -} - - -static double diff_aux(lmm_variable_t var, double x) -{ - double tmp_fpi, result; - - XBT_IN2("(var (%p), x (%1.20f))", var, x); - xbt_assert0(var->func_fpi, - "Initialize the protocol functions first create variables before."); - - tmp_fpi = var->func_fpi(var, x); - result = - tmp_fpi; - - XBT_OUT; - return result; + return diff; } /** \brief Attribute the value bound to var->bound. @@ -468,8 +508,12 @@ static double diff_aux(lmm_variable_t var, double x) * Set default functions to the ones passed as parameters. This is a polimorfism in C pure, enjoy the roots of programming. * */ -void lmm_set_default_protocol_function(double (* func_fpi) (lmm_variable_t var, double x)) +void lmm_set_default_protocol_function(double (* func_f) (lmm_variable_t var, double x), + double (* func_fp) (lmm_variable_t var, double x), + double (* func_fpi) (lmm_variable_t var, double x)) { + func_f_def = func_f; + func_fp_def = func_fp; func_fpi_def = func_fpi; } @@ -482,28 +526,52 @@ void lmm_set_default_protocol_function(double (* func_fpi) (lmm_variable_t var, /* * For Vegas: $f(x) = \alpha D_f\ln(x)$ + * Therefore: $fp(x) = \frac{\alpha D_f}{x}$ * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$ */ #define VEGAS_SCALING 1000.0 -double func_vegas_fpi(lmm_variable_t var, double x){ + +double func_vegas_f(lmm_variable_t var, double x){ + xbt_assert0(x>0.0,"Don't call me with stupid values!"); + return VEGAS_SCALING*var->df*log(x); +} + +double func_vegas_fp(lmm_variable_t var, double x){ xbt_assert0(x>0.0,"Don't call me with stupid values!"); return VEGAS_SCALING*var->df/x; } +double func_vegas_fpi(lmm_variable_t var, double x){ + xbt_assert0(x>0.0,"Don't call me with stupid values!"); + return var->df/(x/VEGAS_SCALING); +} + /* * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$ + * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$ * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$ */ #define RENO_SCALING 1.0 +double func_reno_f(lmm_variable_t var, double x){ + xbt_assert0(var->df>0.0,"Don't call me with stupid values!"); + + return RENO_SCALING*sqrt(3.0/2.0)/var->df*atan(sqrt(3.0/2.0)*var->df*x); +} + +double func_reno_fp(lmm_variable_t var, double x){ + return RENO_SCALING*3.0/(3.0*var->df*var->df*x*x +2.0); +} + double func_reno_fpi(lmm_variable_t var, double x){ double res_fpi; xbt_assert0(var->df>0.0,"Don't call me with stupid values!"); xbt_assert0(x>0.0,"Don't call me with stupid values!"); - res_fpi = 1/(var->df*var->df*x) - 2/(3*var->df*var->df); + res_fpi = 1.0/(var->df*var->df*(x/RENO_SCALING)) - 2.0/(3.0*var->df*var->df); if(res_fpi<=0.0) return 0.0; /* xbt_assert0(res_fpi>0.0,"Don't call me with stupid values!"); */ - return sqrt(RENO_SCALING*res_fpi); + return sqrt(res_fpi); } + -- 2.20.1