1 /* Copyright (c) 2016-2022. 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 "smpi_utils.hpp"
10 #include "smpi_config.hpp"
11 #include "src/surf/xml/platf.hpp"
12 #include "xbt/file.hpp"
15 #include "xbt/parse_units.hpp"
16 #include "xbt/sysdep.h"
18 #include <boost/tokenizer.hpp>
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_utils, smpi, "Logging specific to SMPI (utils)");
22 extern std::string surf_parsed_filename;
23 extern int surf_parse_lineno;
29 double total_benched_time=0;
30 unsigned long total_malloc_size=0;
31 unsigned long total_shared_size=0;
32 unsigned int total_shared_calls=0;
33 struct alloc_metadata_t {
35 unsigned int numcall = 0;
40 struct current_buffer_metadata_t {
41 alloc_metadata_t alloc;
45 alloc_metadata_t max_malloc;
46 F2C* current_handle = nullptr;
47 current_buffer_metadata_t current_buffer1;
48 current_buffer_metadata_t current_buffer2;
50 std::unordered_map<const void*, alloc_metadata_t> allocs;
52 std::unordered_map<int, std::vector<std::string>> collective_calls;
54 std::vector<s_smpi_factor_t> parse_factor(const std::string& smpi_coef_string)
56 std::vector<s_smpi_factor_t> smpi_factor;
58 /** Setup the tokenizer that parses the string **/
59 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
60 boost::char_separator<char> sep(";");
61 boost::char_separator<char> factor_separator(":");
62 Tokenizer tokens(smpi_coef_string, sep);
65 * Iterate over patterns like A:B:C:D;E:F;G:H
66 * These will be broken down into:
71 for (auto token_iter = tokens.begin(); token_iter != tokens.end(); ++token_iter) {
72 XBT_DEBUG("token: %s", token_iter->c_str());
73 Tokenizer factor_values(*token_iter, factor_separator);
75 xbt_assert(factor_values.begin() != factor_values.end(), "Malformed radical for smpi factor: '%s'",
76 smpi_coef_string.c_str());
77 unsigned int iteration = 0;
78 for (auto factor_iter = factor_values.begin(); factor_iter != factor_values.end(); ++factor_iter) {
81 if (factor_iter == factor_values.begin()) { /* first element */
83 fact.factor = std::stoi(*factor_iter);
84 } catch (const std::invalid_argument&) {
85 throw std::invalid_argument(std::string("Invalid factor in chunk ") + std::to_string(smpi_factor.size() + 1) +
90 fact.values.push_back(xbt_parse_get_time(surf_parsed_filename, surf_parse_lineno, *factor_iter, ""));
91 } catch (const std::invalid_argument&) {
92 throw std::invalid_argument(std::string("Invalid factor value ") + std::to_string(iteration) + " in chunk " +
93 std::to_string(smpi_factor.size() + 1) + ": " + *factor_iter);
98 smpi_factor.push_back(fact);
99 XBT_DEBUG("smpi_factor:\t%zu: %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]);
101 std::sort(smpi_factor.begin(), smpi_factor.end(), [](const s_smpi_factor_t &pa, const s_smpi_factor_t &pb) {
102 return (pa.factor < pb.factor);
104 for (auto const& fact : smpi_factor) {
105 XBT_DEBUG("smpi_factor:\t%zu: %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]);
107 smpi_factor.shrink_to_fit();
112 void add_benched_time(double time){
113 total_benched_time += time;
116 void account_malloc_size(size_t size, const std::string& file, int line, void* ptr)
118 if (smpi_cfg_display_alloc()) {
119 alloc_metadata_t metadata;
120 metadata.size = size;
121 metadata.line = line;
122 metadata.numcall = 1;
123 metadata.file = file;
124 allocs.emplace(ptr, metadata);
126 total_malloc_size += size;
127 if(size > max_malloc.size){
128 max_malloc.size = size;
129 max_malloc.line = line;
130 max_malloc.numcall = 1;
131 max_malloc.file = file;
132 } else if (size == max_malloc.size && max_malloc.line == line && max_malloc.file == file) {
133 max_malloc.numcall++;
138 void account_shared_size(size_t size){
139 if (smpi_cfg_display_alloc()) {
140 total_shared_size += size;
141 total_shared_calls++;
145 void print_time_analysis(double global_time){
146 if (simgrid::config::get_value<bool>("smpi/display-timing")) {
147 XBT_INFO("Simulated time: %g seconds. \n\n"
148 "The simulation took %g seconds (after parsing and platform setup)\n"
149 "%g seconds were actual computation of the application",
150 simgrid_get_clock(), global_time, total_benched_time);
151 if (total_benched_time/global_time>=0.75)
152 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
153 "You may want to use sampling functions or trace replay to reduce this.");
157 static void print_leaked_handles()
159 // Put the leaked non-default handles in a vector to sort them by id
160 std::vector<std::pair<unsigned int, smpi::F2C*>> handles;
161 if (simgrid::smpi::F2C::lookup() != nullptr)
162 std::copy_if(simgrid::smpi::F2C::lookup()->begin(), simgrid::smpi::F2C::lookup()->end(),
163 std::back_inserter(handles),
164 [](auto const& entry) { return entry.first >= simgrid::smpi::F2C::get_num_default_handles(); });
168 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
169 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed MPI handles:";
171 message += "\nHINT: Display types and addresses (n max) with --cfg=smpi/list-leaks:n.\n"
172 "Running smpirun with -wrapper \"valgrind --leak-check=full\" can provide more information";
173 XBT_INFO(message.c_str(), handles.size());
177 // we cannot trust F2C::lookup()->size() > F2C::get_num_default_handles() because some default handles are already
178 // freed at this point
179 bool display_advice = false;
180 std::map<std::string, int, std::less<>> count;
181 for (const auto& elem : handles) {
182 std::string key = elem.second->name();
183 if ((not xbt_log_no_loc) && (not elem.second->call_location().empty()))
184 key += " at " + elem.second->call_location();
186 display_advice = true;
187 auto result = count.emplace(key, 1);
188 if (result.second == false)
189 result.first->second++;
192 XBT_WARN("To get more information (location of allocations), compile your code with -trace-call-location flag of "
195 for (const auto& p : count) {
197 XBT_INFO("leaked handle of type %s", p.first.c_str());
199 XBT_INFO("%d leaked handles of type %s", p.second, p.first.c_str());
204 if (max < count.size())
205 XBT_INFO("(%lu more handle leaks hidden as you wanted to see only %lu of them)", count.size() - max, max);
208 static void print_leaked_buffers()
213 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
214 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed buffers:";
216 message += "display types and addresses (n max) with --cfg=smpi/list-leaks:n.\nRunning smpirun with -wrapper "
217 "\"valgrind --leak-check=full\" can provide more information";
218 XBT_INFO(message.c_str(), allocs.size());
223 // gather by allocation origin (only one group reported in case of no-loc or if trace-call-location is not used)
230 std::map<std::string, struct buff_leak, std::less<>> leaks_aggreg;
231 for (const auto& elem : allocs) {
232 std::string key = "leaked allocations";
233 if (not xbt_log_no_loc)
234 key = elem.second.file + ":" + std::to_string(elem.second.line) + ": " + key;
235 auto result = leaks_aggreg.emplace(key, buff_leak{1, elem.second.size, elem.second.size, elem.second.size});
236 if (result.second == false) {
237 result.first->second.count++;
238 result.first->second.total_size += elem.second.size;
239 if (elem.second.size > result.first->second.max_size)
240 result.first->second.max_size = elem.second.size;
241 else if (elem.second.size < result.first->second.min_size)
242 result.first->second.min_size = elem.second.size;
245 // now we can order by total size.
246 std::vector<std::pair<std::string, buff_leak>> leaks(leaks_aggreg.begin(), leaks_aggreg.end());
247 std::sort(leaks.begin(), leaks.end(),
248 [](auto const& a, auto const& b) { return a.second.total_size > b.second.total_size; });
251 for (const auto& p : leaks) {
252 if (p.second.min_size == p.second.max_size)
253 XBT_INFO("%s of total size %zu, called %d times, each with size %zu", p.first.c_str(), p.second.total_size,
254 p.second.count, p.second.min_size);
256 XBT_INFO("%s of total size %zu, called %d times, with minimum size %zu and maximum size %zu", p.first.c_str(),
257 p.second.total_size, p.second.count, p.second.min_size, p.second.max_size);
262 if (max < leaks_aggreg.size())
263 XBT_INFO("(more buffer leaks hidden as you wanted to see only %lu of them)", max);
266 void print_memory_analysis()
268 if (smpi_cfg_display_alloc()) {
269 print_leaked_handles();
270 print_leaked_buffers();
272 if(total_malloc_size != 0)
273 XBT_INFO("Memory Usage: Simulated application allocated %lu bytes during its lifetime through malloc/calloc calls.\n"
274 "Largest allocation at once from a single process was %zu bytes, at %s:%d. It was called %u times during the whole simulation.\n"
275 "If this is too much, consider sharing allocations for computation buffers.\n"
276 "This can be done automatically by setting --cfg=smpi/auto-shared-malloc-thresh to the minimum size wanted size (this can alter execution if data content is necessary)\n",
277 total_malloc_size, max_malloc.size, simgrid::xbt::Path(max_malloc.file).get_base_name().c_str(), max_malloc.line, max_malloc.numcall
281 "Allocations analysis asked, but 0 bytes were allocated through malloc/calloc calls intercepted by SMPI.\n"
282 "The code may not use malloc() to allocate memory, or it was built with SMPI_NO_OVERRIDE_MALLOC");
283 if(total_shared_size != 0)
284 XBT_INFO("%lu bytes were automatically shared between processes, in %u calls\n", total_shared_size, total_shared_calls);
288 void set_current_handle(F2C* handle){
289 current_handle=handle;
292 void print_current_handle(){
294 if(current_handle->call_location().empty())
295 XBT_INFO("To get handle location information, pass -trace-call-location flag to smpicc/f90 as well");
297 XBT_INFO("Handle %s was allocated by a call at %s", current_handle->name().c_str(),
298 (char*)(current_handle->call_location().c_str()));
302 void set_current_buffer(int i, const char* name, const void* buf){
305 if(not current_buffer1.name.empty()){
306 current_buffer1.name="";
308 if(not current_buffer2.name.empty()){
309 current_buffer2.name="";
312 auto meta = allocs.find(buf);
313 if (meta == allocs.end()) {
314 XBT_DEBUG("Buffer %p was not allocated with malloc/calloc", buf);
318 current_buffer1.alloc = meta->second;
319 current_buffer1.name = name;
321 current_buffer2.alloc=meta->second;
322 current_buffer2.name=name;
326 void print_buffer_info()
328 if (not current_buffer1.name.empty())
329 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer1.name.c_str(),
330 current_buffer1.alloc.file.c_str(), current_buffer1.alloc.line, current_buffer1.alloc.size);
331 if (not current_buffer2.name.empty())
332 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer2.name.c_str(),
333 current_buffer2.alloc.file.c_str(), current_buffer2.alloc.line, current_buffer2.alloc.size);
336 size_t get_buffer_size(const void* buf){
337 auto meta = allocs.find(buf);
338 if (meta == allocs.end()) {
339 //we don't know this buffer (on stack or feature disabled), assume it's fine.
340 return std::numeric_limits<std::size_t>::max();
342 return meta->second.size;
345 void account_free(const void* ptr){
346 if (smpi_cfg_display_alloc()) {
351 int check_collectives_ordering(MPI_Comm comm, const std::string& call)
353 unsigned int count = comm->get_collectives_count();
354 comm->increment_collectives_count();
355 auto vec = collective_calls.find(comm->id());
356 if (vec == collective_calls.end()) {
357 collective_calls.emplace(comm->id(), std::vector<std::string>{call});
359 // are we the first ? add the call
360 if (vec->second.size() == count) {
361 vec->second.emplace_back(call);
362 } else if (vec->second.size() > count) {
363 if (vec->second[count] != call) {
364 XBT_WARN("Collective operation mismatch. For process %ld, expected %s, got %s",
365 simgrid::s4u::this_actor::get_pid(), vec->second[count].c_str(), call.c_str());
366 return MPI_ERR_OTHER;
376 } // namespace simgrid