1 /* Copyright (c) 2004-2023. 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. */
6 #ifndef SIMGRID_KERNEL_LMM_BMF_HPP
7 #define SIMGRID_KERNEL_LMM_BMF_HPP
9 #include "src/kernel/lmm/System.hpp"
10 #include "xbt/config.hpp"
15 // Ignore deprecation warnings with Eigen < 4.0 (see https://gitlab.com/libeigen/eigen/-/issues/1850)
16 #pragma clang diagnostic push
17 #pragma clang diagnostic ignored "-Wdeprecated-declarations"
19 #include <Eigen/Dense>
21 #pragma clang diagnostic pop
24 #include <unordered_set>
26 namespace simgrid::kernel::lmm {
28 /** @brief Generate all combinations of valid allocation */
29 class XBT_PUBLIC AllocationGenerator {
31 explicit AllocationGenerator(Eigen::MatrixXd A);
34 * @brief Get next valid allocation
36 * @param next_alloc Allocation (OUTPUT)
37 * @return true if there's an allocation not tested yet, false otherwise
39 bool next(std::vector<int>& next_alloc);
43 std::vector<int> alloc_;
50 * Despite the simplicity of BMF fairness definition, it's quite hard to
51 * find a BMF allocation in the general case.
53 * This solver implements one possible algorithm to find a BMF, as proposed
54 * at: https://hal.archives-ouvertes.fr/hal-01552739.
56 * The idea of this algorithm is that each player/flow "selects" a resource to
57 * saturate. Then, we calculate the rate each flow would have with this allocation.
58 * If the allocation is a valid BMF and no one needs to move, it's over. Otherwise,
59 * each player selects a new resource to saturate based on the minimim rate possible
60 * between all resources.
63 * 1) Given an initial allocation B_i
64 * 2) Build a matrix A'_ji and C'_ji which assures that the player receives the most
65 * share at selected resources
66 * 3) Solve: A'_ji * rho_i = C'_j
67 * 4) Calculate the minimum fair rate for each resource j: f_j. The f_j represents
68 * the maximum each flow can receive at the resource j.
69 * 5) Builds a new vector B'_i = arg min(f_j/A_ji).
70 * 6) Stop if B == B' (nobody needs to move), go to step 2 otherwise
72 * Despite the overall good performance of this algorithm, which converges in a few
73 * iterations, we don't have any assurance about its convergence. In the worst case,
74 * it may be needed to test all possible combination of allocations (which is exponential).
78 class XBT_PUBLIC BmfSolver {
79 inline static simgrid::config::Flag<int> cfg_bmf_max_iteration{
80 "bmf/max-iterations", "Maximum number of steps to be performed while searching for a BMF allocation", 1000};
82 inline static simgrid::config::Flag<double> cfg_bmf_precision{
83 "bmf/precision", {"precision/bmf"}, "Numerical precision used when computing resource sharing", 1E-12};
87 * @brief Instantiate the BMF solver
89 * @param A A_ji: consumption of player i on resource j
90 * @param maxA maxA_ji: consumption of larger player i on resource j
91 * @param C Resource capacity
92 * @param shared Is resource shared between player or each player receives the full capacity (FATPIPE links)
93 * @param phi Bound for each player
95 BmfSolver(Eigen::MatrixXd A, Eigen::MatrixXd maxA, Eigen::VectorXd C, std::vector<bool> shared, Eigen::VectorXd phi);
96 /** @brief Solve equation system to find a fair-sharing of resources */
97 Eigen::VectorXd solve();
100 using allocation_map_t = std::unordered_map<int, std::unordered_set<int>>;
102 * @brief Get actual resource capacity considering bounded players
104 * Calculates the resource capacity considering that some players on it may be bounded by user,
105 * i.e. an explicit limit in speed was configured
107 * @param resource Internal index of resource in C_ vector
108 * @param bounded_players List of players that are externally bounded
109 * @return Actual resource capacity
111 double get_resource_capacity(int resource, const std::vector<int>& bounded_players) const;
113 * @brief Get maxmin share of the resource
115 * @param resource Internal index of resource in C_ vector
116 * @param bounded_players List of players that are externally bounded
117 * @return maxmin share
119 double get_maxmin_share(int resource, const std::vector<int>& bounded_players) const;
121 * @brief Auxiliary method to get list of bounded player from allocation
123 * @param alloc Current allocation
124 * @return list of bounded players
126 std::vector<int> get_bounded_players(const allocation_map_t& alloc) const;
129 * @brief Given an allocation calculates the speed/rho for each player
132 * Builds 2 auxiliares matrices A' and C' and solves the system: rho_i = inv(A'_ji) * C'_j
134 * All resources in A' and C' are saturated, i.e., sum(A'_j * rho_i) = C'_j.
136 * The matrix A' is built as follows:
137 * - For each resource j in alloc: copy row A_j to A'
138 * - If 2 players (i, k) share a same resource, assure fairness by adding a row in A' such as:
139 * - A_ji*rho_i - Ajk*rho_j = 0
141 * @param alloc for each resource, players that chose to saturate it
142 * @return Vector rho with "players' speed"
144 Eigen::VectorXd equilibrium(const allocation_map_t& alloc) const;
147 * @brief Given a fair_sharing vector, gets the allocation
149 * The allocation for player i is given by: min(bound, f_j/A_ji).
150 * The minimum between all fair-sharing and the external bound (if any)
152 * The algorithm dictates a random initial allocation. For simplicity, we opt to use the same
153 * logic with the fair_sharing vector.
155 * @param fair_sharing Fair sharing vector
156 * @param initial Is this the initial allocation?
157 * @return allocation vector
159 bool get_alloc(const Eigen::VectorXd& fair_sharing, const allocation_map_t& last_alloc, allocation_map_t& alloc,
162 bool disturb_allocation(allocation_map_t& alloc, std::vector<int>& alloc_by_player);
164 * @brief Calculates the fair sharing for each resource
166 * Basically 3 options:
167 * 1) resource in allocation: A_ji*rho_i since all players who selected this resource have the same share
168 * 2) resource not selected by saturated (fully used): divide it by the number of players C_/n_players
169 * 3) resource not selected and not-saturated: no limitation
171 * @param alloc Allocation map (resource-> players)
172 * @param rho Speed for each player i
173 * @param fair_sharing Output vector, fair sharing for each resource j
175 void set_fair_sharing(const allocation_map_t& alloc, const Eigen::VectorXd& rho, Eigen::VectorXd& fair_sharing) const;
178 * @brief Check if allocation is BMF
180 * To be a bmf allocation it must:
181 * - respect the capacity of all resources
182 * - saturate at least 1 resource
183 * - every player receives maximum share in at least 1 saturated resource
184 * @param rho Allocation
185 * @return true if BMF false otherwise
187 bool is_bmf(const Eigen::VectorXd& rho) const;
188 std::vector<int> alloc_map_to_vector(const allocation_map_t& alloc) const;
191 * @brief Set of debug functions to print the different objects
193 template <typename T> std::string debug_eigen(const T& obj) const;
194 template <typename C> std::string debug_vector(const C& container) const;
195 std::string debug_alloc(const allocation_map_t& alloc) const;
197 Eigen::MatrixXd A_; //!< A_ji: resource usage matrix, each row j represents a resource and col i a flow/player
198 Eigen::MatrixXd maxA_; //!< maxA_ji, similar as A_, but containing the maximum consumption of player i (if player a
199 //!< single flow it's equal to A_)
200 Eigen::VectorXd C_; //!< C_j Capacity of each resource
201 std::vector<bool> C_shared_; //!< shared_j Resource j is shared or not
202 Eigen::VectorXd phi_; //!< phi_i bound for each player
204 std::set<std::vector<int>> allocations_; //!< set of already tested allocations, since last identified loop
205 AllocationGenerator gen_;
206 static constexpr int NO_RESOURCE = -1; //!< flag to indicate player has selected no resource
207 int max_iteration_ = cfg_bmf_max_iteration; //!< number maximum of iterations of BMF algorithm
213 * A BMF (bottleneck max fairness) solver to resolve inequation systems.
215 * Usually, SimGrid relies on a *max-min fairness* solver to share the resources.
216 * Max-min is great when sharing homogenous resources, however it cannot be used with heterogeneous resources.
218 * BMF is a natural alternative to max-min, providing a fair-sharing of heterogeneous resources (CPU, network, disk).
219 * It is specially relevant for the implementation of parallel tasks whose sharing involves different
220 * kinds of resources.
222 * BMF assures that every flow receives the maximum share possible in at least 1 bottleneck (fully used) resource.
224 * The BMF is characterized by:
225 * - A_ji: a matrix of requirement for flows/player. For each resource j, and flow i, A_ji represents the utilization
226 * of resource j for 1 unit of the flow i.
227 * - rho_i: the rate allocated for flow i (same among all resources)
228 * - C_j: the capacity of each resource (can be bytes/s, flops/s, etc)
230 * Therefore, these conditions need to satisfied to an allocation be considered a BMF:
231 * 1) All constraints are respected (flows cannot use more than the resource has available)
232 * - for all resource j and player i: A_ji * rho_i <= C_j
233 * 2) At least 1 resource is fully used (bottleneck).
234 * - for some resource j: A_ji * rho_i = C_j
235 * 3) Each flow (player) receives the maximum share in at least 1 bottleneck.
236 * - for all player i: exist a resource j: A_ji * rho_i >= A_jk * rho_k for all other player k
238 * Despite the prove of existence of a BMF allocation in the general case, it may not
239 * be unique, which leads to possible different rate for the applications.
241 * More details about BMF can be found at: https://hal.inria.fr/hal-01243985/document
246 * @brief Bottleneck max-fair system
248 class XBT_PUBLIC BmfSystem : public System {
250 using System::System;
253 /** @brief Implements the solve method to calculate a BMF allocation */
254 void do_solve() final;
255 using allocation_map_t = std::unordered_map<int, std::unordered_set<int>>;
257 * @brief Solve equation system to find a fair-sharing of resources
259 * @param cnst_list Constraint list (modified for selective update or active)
261 template <class CnstList> void bmf_solve(const CnstList& cnst_list);
263 * @brief Iterates over system and build the consumption matrix A_ji and maxA_ji
265 * Each row j represents a resource and each col i a player/flow
267 * Considers only active variables to build the matrix.
269 * @param number_cnsts Number of constraints in the system
270 * @param A Consumption matrix (OUTPUT)
271 * @param maxA Max subflow consumption matrix (OUTPUT)
272 * @param phi Bounds for variables
274 void get_flows_data(Eigen::Index number_cnsts, Eigen::MatrixXd& A, Eigen::MatrixXd& maxA, Eigen::VectorXd& phi);
276 * @brief Builds the vector C_ with resource's capacity
278 * @param cnst_list Constraint list (modified for selective update or active)
279 * @param C Resource capacity vector
280 * @param shared Resource is shared or not (fatpipe links)
282 template <class CnstList>
283 void get_constraint_data(const CnstList& cnst_list, Eigen::VectorXd& C, std::vector<bool>& shared);
285 std::unordered_map<int, Variable*> idx2Var_; //!< Map player index (and position in matrices) to system's variable
286 std::unordered_map<const Constraint*, int> cnst2idx_; //!< Conversely map constraint to index
289 } // namespace simgrid::kernel::lmm