1 /* Copyright (c) 2009-2021. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "simdag_private.hpp"
8 #include "simgrid/simdag.h"
9 #include "xbt/file.hpp"
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(sd_daxparse, sd, "Parsing DAX files");
21 /* Ensure that transfer tasks have unique names even though a file is used several times */
22 void uniq_transfer_task_name(SD_task_t task)
24 const_SD_task_t child = *(task->successors->begin());
25 const_SD_task_t parent = *(task->predecessors->begin());
27 std::string new_name =
28 std::string(SD_task_get_name(parent)) + "_" + SD_task_get_name(task) + "_" + SD_task_get_name(child);
30 SD_task_set_name(task, new_name.c_str());
33 static bool children_are_marked(const_SD_task_t task)
35 return std::none_of(task->successors->begin(), task->successors->end(),
36 [](const SD_task_t& elm) { return not elm->marked; }) &&
37 std::none_of(task->outputs->begin(), task->outputs->end(),
38 [](const SD_task_t& elm) { return not elm->marked; });
41 static bool parents_are_marked(const_SD_task_t task)
43 return std::none_of(task->predecessors->begin(), task->predecessors->end(),
44 [](const SD_task_t& elm) { return not elm->marked; }) &&
45 std::none_of(task->inputs->begin(), task->inputs->end(), [](const SD_task_t& elm) { return not elm->marked; });
48 bool acyclic_graph_detail(const_xbt_dynar_t dag)
51 SD_task_t task = nullptr;
52 std::vector<SD_task_t> current;
53 xbt_dynar_foreach (dag, count, task)
54 if (task->kind != SD_TASK_COMM_E2E && task->successors->empty() && task->outputs->empty())
55 current.push_back(task);
57 while (not current.empty()) {
58 std::vector<SD_task_t> next;
59 for (auto const& t : current) {
62 for (SD_task_t const& input : *t->inputs) {
64 // Inputs are communication, hence they can have only one predecessor
65 SD_task_t input_pred = *(input->predecessors->begin());
66 if (children_are_marked(input_pred))
67 next.push_back(input_pred);
69 for (SD_task_t const& pred : *t->predecessors) {
70 if (children_are_marked(pred))
78 bool all_marked = true;
79 //test if all tasks are marked
80 xbt_dynar_foreach(dag,count,task){
81 if (task->kind != SD_TASK_COMM_E2E && not task->marked) {
82 XBT_WARN("the task %s is not marked",task->name);
89 XBT_VERB("there is at least one cycle in your task graph");
90 xbt_dynar_foreach(dag,count,task){
91 if(task->kind != SD_TASK_COMM_E2E && task->predecessors->empty() && task->inputs->empty()){
93 current.push_back(task);
96 //test if something has to be done for the next iteration
97 while (not current.empty()) {
98 std::vector<SD_task_t> next;
99 //test if the current iteration is done
100 for (auto const& t : current) {
102 for (SD_task_t const& output : *t->outputs) {
103 output->marked = true;
104 // outputs are communication, hence they can have only one successor
105 SD_task_t output_succ = *(output->successors->begin());
106 if (parents_are_marked(output_succ))
107 next.push_back(output_succ);
109 for (SD_task_t const& succ : *t->successors) {
110 if (parents_are_marked(succ))
111 next.push_back(succ);
119 xbt_dynar_foreach(dag,count,task){
120 if (task->kind != SD_TASK_COMM_E2E && not task->marked) {
121 XBT_WARN("the task %s is in a cycle",task->name);
129 static YY_BUFFER_STATE input_buffer;
131 static xbt_dynar_t result;
132 static std::map<std::string, SD_task_t, std::less<>> jobs;
133 static std::map<std::string, SD_task_t, std::less<>> files;
134 static SD_task_t current_job;
136 /** @brief loads a DAX file describing a DAG
138 * See https://confluence.pegasus.isi.edu/display/pegasus/WorkflowGenerator for more details.
140 xbt_dynar_t SD_daxload(const char *filename)
143 FILE* in_file = fopen(filename, "r");
144 xbt_assert(in_file, "Unable to open \"%s\"\n", filename);
145 input_buffer = dax__create_buffer(in_file, 10);
146 dax__switch_to_buffer(input_buffer);
149 result = xbt_dynar_new(sizeof(SD_task_t), nullptr);
150 SD_task_t root_task = SD_task_create_comp_seq("root", nullptr, 0);
151 /* by design the root task is always SCHEDULABLE */
152 SD_task_set_state(root_task, SD_SCHEDULABLE);
154 xbt_dynar_push(result, &root_task);
155 SD_task_t end_task = SD_task_create_comp_seq("end", nullptr, 0);
159 xbt_die("Parse error in %s: %s", filename, dax__parse_err_msg());
160 dax__delete_buffer(input_buffer);
164 /* And now, post-process the files.
165 * We want a file task per pair of computation tasks exchanging the file. Duplicate on need
166 * Files not produced in the system are said to be produced by root task (top of DAG).
167 * Files not consumed in the system are said to be consumed by end task (bottom of DAG).
170 for (auto const& elm : files) {
173 if (file->predecessors->empty()) {
174 for (SD_task_t const& it : *file->successors) {
175 newfile = SD_task_create_comm_e2e(file->name, nullptr, file->amount);
176 SD_task_dependency_add(root_task, newfile);
177 SD_task_dependency_add(newfile, it);
178 xbt_dynar_push(result, &newfile);
181 if (file->successors->empty()) {
182 for (SD_task_t const& it : *file->predecessors) {
183 newfile = SD_task_create_comm_e2e(file->name, nullptr, file->amount);
184 SD_task_dependency_add(it, newfile);
185 SD_task_dependency_add(newfile, end_task);
186 xbt_dynar_push(result, &newfile);
189 for (SD_task_t const& it : *file->predecessors) {
190 for (SD_task_t const& it2 : *file->successors) {
192 XBT_WARN("File %s is produced and consumed by task %s."
193 "This loop dependency will prevent the execution of the task.",
194 file->name, it->name);
196 newfile = SD_task_create_comm_e2e(file->name, nullptr, file->amount);
197 SD_task_dependency_add(it, newfile);
198 SD_task_dependency_add(newfile, it2);
199 xbt_dynar_push(result, &newfile);
202 /* Free previous copy of the files */
203 SD_task_destroy(file);
206 /* Push end task last */
207 xbt_dynar_push(result, &end_task);
210 xbt_dynar_foreach(result, cpt, file) {
211 if (SD_task_get_kind(file) == SD_TASK_COMM_E2E) {
212 uniq_transfer_task_name(file);
214 /* If some tasks do not take files as input, connect them to the root
215 * if they don't produce files, connect them to the end node.
217 if ((file != root_task) && (file != end_task)) {
218 if (file->inputs->empty())
219 SD_task_dependency_add(root_task, file);
220 if (file->outputs->empty())
221 SD_task_dependency_add(file, end_task);
226 if (not acyclic_graph_detail(result)) {
227 XBT_ERROR("The DAX described in %s is not a DAG. It contains a cycle.",
228 simgrid::xbt::Path(filename).get_base_name().c_str());
229 xbt_dynar_foreach(result, cpt, file)
230 SD_task_destroy(file);
231 xbt_dynar_free_container(&result);
238 void STag_dax__adag()
241 double version = std::stod(std::string(A_dax__adag_version));
242 xbt_assert(version == 2.1, "Expected version 2.1 in <adag> tag, got %f. Fix the parser or your file", version);
243 } catch (const std::invalid_argument&) {
244 throw std::invalid_argument(std::string("Parse error: ") + A_dax__adag_version + " is not a double");
251 double runtime = std::stod(std::string(A_dax__job_runtime));
253 std::string name = std::string(A_dax__job_id) + "@" + A_dax__job_name;
254 runtime *= 4200000000.; /* Assume that timings were done on a 4.2GFlops machine. I mean, why not? */
255 XBT_DEBUG("See <job id=%s runtime=%s %.0f>", A_dax__job_id, A_dax__job_runtime, runtime);
256 current_job = SD_task_create_comp_seq(name.c_str(), nullptr, runtime);
257 jobs.insert({A_dax__job_id, current_job});
258 xbt_dynar_push(result, ¤t_job);
259 } catch (const std::invalid_argument&) {
260 throw std::invalid_argument(std::string("Parse error: ") + A_dax__job_runtime + " is not a double");
264 void STag_dax__uses()
268 size = std::stod(std::string(A_dax__uses_size));
269 } catch (const std::invalid_argument&) {
270 throw std::invalid_argument(std::string("Parse error: ") + A_dax__uses_size + " is not a double");
272 bool is_input = (A_dax__uses_link == A_dax__uses_link_input);
274 XBT_DEBUG("See <uses file=%s %s>",A_dax__uses_file,(is_input?"in":"out"));
275 auto it = files.find(A_dax__uses_file);
277 if (it == files.end()) {
278 file = SD_task_create_comm_e2e(A_dax__uses_file, nullptr, size);
279 sd_global->initial_tasks.erase(file);
280 files[A_dax__uses_file] = file;
283 if (file->amount < size || file->amount > size) {
284 XBT_WARN("Ignore file %s size redefinition from %.0f to %.0f", A_dax__uses_file, SD_task_get_amount(file), size);
288 SD_task_dependency_add(file, current_job);
290 SD_task_dependency_add(current_job, file);
291 if ((file->predecessors->size() + file->inputs->size()) > 1) {
292 XBT_WARN("File %s created at more than one location...", file->name);
297 static SD_task_t current_child;
298 void STag_dax__child()
300 auto job = jobs.find(A_dax__child_ref);
301 if (job != jobs.end()) {
302 current_child = job->second;
304 throw std::out_of_range(std::string("Parse error on line ") + std::to_string(dax_lineno) +
305 ": Asked to add dependencies to the non-existent " + A_dax__child_ref + "task");
309 void ETag_dax__child()
311 current_child = nullptr;
314 void STag_dax__parent()
316 auto job = jobs.find(A_dax__parent_ref);
317 if (job != jobs.end()) {
318 SD_task_t parent = job->second;
319 SD_task_dependency_add(parent, current_child);
320 XBT_DEBUG("Control-flow dependency from %s to %s", current_child->name, parent->name);
322 throw std::out_of_range(std::string("Parse error on line ") + std::to_string(dax_lineno) +
323 ": Asked to add a dependency from " + current_child->name + " to " + A_dax__parent_ref +
324 ", but " + A_dax__parent_ref + " does not exist");
328 void ETag_dax__adag()
330 XBT_DEBUG("See </adag>");
335 current_job = nullptr;
336 XBT_DEBUG("See </job>");
339 void ETag_dax__parent()
341 XBT_DEBUG("See </parent>");
344 void ETag_dax__uses()
346 XBT_DEBUG("See </uses>");