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;
25 namespace simgrid::smpi::utils {
27 double total_benched_time=0;
28 unsigned long total_malloc_size=0;
29 unsigned long total_shared_size=0;
30 unsigned int total_shared_calls=0;
31 struct alloc_metadata_t {
33 unsigned int numcall = 0;
38 struct current_buffer_metadata_t {
39 alloc_metadata_t alloc;
43 alloc_metadata_t max_malloc;
44 F2C* current_handle = nullptr;
45 current_buffer_metadata_t current_buffer1;
46 current_buffer_metadata_t current_buffer2;
48 std::unordered_map<const void*, alloc_metadata_t> allocs;
50 std::unordered_map<int, std::vector<std::string>> collective_calls;
52 void FactorSet::parse(const std::string& values)
54 factors_ = parse_factor(values);
58 FactorSet::FactorSet(const std::string& name, double default_value,
59 std::function<double(s_smpi_factor_t const&)> const& lambda)
60 : name_(name), default_value_(default_value), lambda_(lambda)
64 double FactorSet::operator()()
66 return lambda_(factors_.front());
69 double FactorSet::operator()(double size)
72 return default_value_;
74 for (long unsigned i = 0; i < factors_.size(); i++) {
75 auto const& fact = factors_[i];
77 if (size <= fact.factor) { // Too large already, use the previous value
79 if (i == 0) { // Before the first boundary: use the default value
80 XBT_DEBUG("%s: %f <= %zu return default %f", name_.c_str(), size, fact.factor, default_value_);
81 return default_value_;
83 double val = lambda_(factors_[i - 1]);
84 XBT_DEBUG("%s: %f <= %zu return %f", name_.c_str(), size, fact.factor, val);
88 double val = lambda_(factors_.back());
90 XBT_DEBUG("%s: %f > %zu return %f", name_.c_str(), size, factors_.back().factor, val);
94 std::vector<s_smpi_factor_t> parse_factor(const std::string& smpi_coef_string)
96 std::vector<s_smpi_factor_t> smpi_factor;
98 /** Setup the tokenizer that parses the string **/
99 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
100 boost::char_separator<char> sep(";");
101 boost::char_separator<char> factor_separator(":");
102 Tokenizer tokens(smpi_coef_string, sep);
105 * Iterate over patterns like A:B:C:D;E:F;G:H
106 * These will be broken down into:
111 for (auto token_iter = tokens.begin(); token_iter != tokens.end(); ++token_iter) {
112 XBT_DEBUG("token: %s", token_iter->c_str());
113 Tokenizer factor_values(*token_iter, factor_separator);
114 s_smpi_factor_t fact;
115 xbt_assert(factor_values.begin() != factor_values.end(), "Malformed radical for smpi factor: '%s'",
116 smpi_coef_string.c_str());
117 unsigned int iteration = 0;
118 for (auto factor_iter = factor_values.begin(); factor_iter != factor_values.end(); ++factor_iter) {
121 if (factor_iter == factor_values.begin()) { /* first element */
123 fact.factor = std::stoi(*factor_iter);
124 } catch (const std::invalid_argument&) {
125 throw std::invalid_argument(std::string("Invalid factor in chunk ") + std::to_string(smpi_factor.size() + 1) +
126 ": " + *factor_iter);
130 fact.values.push_back(xbt_parse_get_time(surf_parsed_filename, surf_parse_lineno, *factor_iter, ""));
131 } catch (const std::invalid_argument&) {
132 throw std::invalid_argument(std::string("Invalid factor value ") + std::to_string(iteration) + " in chunk " +
133 std::to_string(smpi_factor.size() + 1) + ": " + *factor_iter);
138 smpi_factor.push_back(fact);
139 XBT_DEBUG("smpi_factor:\t%zu: %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]);
141 std::sort(smpi_factor.begin(), smpi_factor.end(), [](const s_smpi_factor_t &pa, const s_smpi_factor_t &pb) {
142 return (pa.factor < pb.factor);
144 for (auto const& fact : smpi_factor) {
145 XBT_DEBUG("smpi_factor:\t%zu: %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]);
147 smpi_factor.shrink_to_fit();
152 void add_benched_time(double time){
153 total_benched_time += time;
156 void account_malloc_size(size_t size, std::string_view file, int line, const void* ptr)
158 if (smpi_cfg_display_alloc()) {
159 alloc_metadata_t metadata;
160 metadata.size = size;
161 metadata.line = line;
162 metadata.numcall = 1;
163 metadata.file = file;
164 allocs.try_emplace(ptr, metadata);
166 total_malloc_size += size;
167 if(size > max_malloc.size){
168 max_malloc.size = size;
169 max_malloc.line = line;
170 max_malloc.numcall = 1;
171 max_malloc.file = file;
172 } else if (size == max_malloc.size && max_malloc.line == line && max_malloc.file == file) {
173 max_malloc.numcall++;
178 void account_shared_size(size_t size){
179 if (smpi_cfg_display_alloc()) {
180 total_shared_size += size;
181 total_shared_calls++;
185 void print_time_analysis(double global_time){
186 if (simgrid::config::get_value<bool>("smpi/display-timing")) {
187 XBT_INFO("Simulated time: %g seconds. \n\n"
188 "The simulation took %g seconds (after parsing and platform setup)\n"
189 "%g seconds were actual computation of the application",
190 simgrid_get_clock(), global_time, total_benched_time);
191 if (total_benched_time/global_time>=0.75)
192 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
193 "You may want to use sampling functions or trace replay to reduce this.");
197 static void print_leaked_handles()
199 // Put the leaked non-default handles in a vector to sort them by id
200 std::vector<std::pair<unsigned int, smpi::F2C*>> handles;
201 if (simgrid::smpi::F2C::lookup() != nullptr)
202 std::copy_if(simgrid::smpi::F2C::lookup()->begin(), simgrid::smpi::F2C::lookup()->end(),
203 std::back_inserter(handles),
204 [](auto const& entry) { return entry.first >= simgrid::smpi::F2C::get_num_default_handles(); });
208 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
209 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed MPI handles:";
211 message += "\nHINT: Display types and addresses (n max) with --cfg=smpi/list-leaks:n.\n"
212 "Running smpirun with -wrapper \"valgrind --leak-check=full\" can provide more information";
213 XBT_INFO(message.c_str(), handles.size());
217 // we cannot trust F2C::lookup()->size() > F2C::get_num_default_handles() because some default handles are already
218 // freed at this point
219 bool display_advice = false;
220 std::map<std::string, int, std::less<>> count;
221 for (const auto& [_, elem] : handles) {
222 std::string key = elem->name();
223 if ((not xbt_log_no_loc) && (not elem->call_location().empty()))
224 key += " at " + elem->call_location();
226 display_advice = true;
227 auto& result = count.try_emplace(key, 0).first->second;
231 XBT_WARN("To get more information (location of allocations), compile your code with -trace-call-location flag of "
234 for (const auto& [key, value] : count) {
236 XBT_INFO("leaked handle of type %s", key.c_str());
238 XBT_INFO("%d leaked handles of type %s", value, key.c_str());
243 if (max < count.size())
244 XBT_INFO("(%lu more handle leaks hidden as you wanted to see only %lu of them)", count.size() - max, max);
247 static void print_leaked_buffers()
252 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
253 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed buffers:";
255 message += "display types and addresses (n max) with --cfg=smpi/list-leaks:n.\nRunning smpirun with -wrapper "
256 "\"valgrind --leak-check=full\" can provide more information";
257 XBT_INFO(message.c_str(), allocs.size());
262 // gather by allocation origin (only one group reported in case of no-loc or if trace-call-location is not used)
269 std::map<std::string, struct buff_leak, std::less<>> leaks_aggreg;
270 for (const auto& [_, elem] : allocs) {
271 std::string key = "leaked allocations";
272 if (not xbt_log_no_loc)
273 key = elem.file + ":" + std::to_string(elem.line) + ": " + key;
274 auto& result = leaks_aggreg.try_emplace(key, buff_leak{0, 0, elem.size, elem.size}).first->second;
276 result.total_size += elem.size;
277 if (elem.size > result.max_size)
278 result.max_size = elem.size;
279 else if (elem.size < result.min_size)
280 result.min_size = elem.size;
282 // now we can order by total size.
283 std::vector<std::pair<std::string, buff_leak>> leaks(leaks_aggreg.begin(), leaks_aggreg.end());
284 std::sort(leaks.begin(), leaks.end(),
285 [](auto const& a, auto const& b) { return a.second.total_size > b.second.total_size; });
288 for (const auto& [key, value] : leaks) {
289 if (value.min_size == value.max_size)
290 XBT_INFO("%s of total size %zu, called %d times, each with size %zu", key.c_str(), value.total_size, value.count,
293 XBT_INFO("%s of total size %zu, called %d times, with minimum size %zu and maximum size %zu", key.c_str(),
294 value.total_size, value.count, value.min_size, value.max_size);
299 if (max < leaks_aggreg.size())
300 XBT_INFO("(more buffer leaks hidden as you wanted to see only %lu of them)", max);
303 void print_memory_analysis()
305 if (smpi_cfg_display_alloc()) {
306 print_leaked_handles();
307 print_leaked_buffers();
309 if(total_malloc_size != 0)
310 XBT_INFO("Memory Usage: Simulated application allocated %lu bytes during its lifetime through malloc/calloc calls.\n"
311 "Largest allocation at once from a single process was %zu bytes, at %s:%d. It was called %u times during the whole simulation.\n"
312 "If this is too much, consider sharing allocations for computation buffers.\n"
313 "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",
314 total_malloc_size, max_malloc.size, simgrid::xbt::Path(max_malloc.file).get_base_name().c_str(), max_malloc.line, max_malloc.numcall
318 "Allocations analysis asked, but 0 bytes were allocated through malloc/calloc calls intercepted by SMPI.\n"
319 "The code may not use malloc() to allocate memory, or it was built with SMPI_NO_OVERRIDE_MALLOC");
320 if(total_shared_size != 0)
321 XBT_INFO("%lu bytes were automatically shared between processes, in %u calls\n", total_shared_size, total_shared_calls);
325 void set_current_handle(F2C* handle){
326 current_handle=handle;
329 void print_current_handle(){
331 if(current_handle->call_location().empty())
332 XBT_INFO("To get handle location information, pass -trace-call-location flag to smpicc/f90 as well");
334 XBT_INFO("Handle %s was allocated by a call at %s", current_handle->name().c_str(),
335 (char*)(current_handle->call_location().c_str()));
339 void set_current_buffer(int i, const char* name, const void* buf){
342 if(not current_buffer1.name.empty()){
343 current_buffer1.name="";
345 if(not current_buffer2.name.empty()){
346 current_buffer2.name="";
349 auto meta = allocs.find(buf);
350 if (meta == allocs.end()) {
351 XBT_DEBUG("Buffer %p was not allocated with malloc/calloc", buf);
355 current_buffer1.alloc = meta->second;
356 current_buffer1.name = name;
358 current_buffer2.alloc=meta->second;
359 current_buffer2.name=name;
363 void print_buffer_info()
365 if (not current_buffer1.name.empty())
366 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer1.name.c_str(),
367 current_buffer1.alloc.file.c_str(), current_buffer1.alloc.line, current_buffer1.alloc.size);
368 if (not current_buffer2.name.empty())
369 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer2.name.c_str(),
370 current_buffer2.alloc.file.c_str(), current_buffer2.alloc.line, current_buffer2.alloc.size);
373 size_t get_buffer_size(const void* buf){
374 auto meta = allocs.find(buf);
375 if (meta == allocs.end()) {
376 //we don't know this buffer (on stack or feature disabled), assume it's fine.
377 return std::numeric_limits<std::size_t>::max();
379 return meta->second.size;
382 void account_free(const void* ptr){
383 if (smpi_cfg_display_alloc()) {
388 int check_collectives_ordering(MPI_Comm comm, const std::string& call)
390 unsigned int count = comm->get_collectives_count();
391 comm->increment_collectives_count();
392 if (auto vec = collective_calls.find(comm->id()); vec == collective_calls.end()) {
393 collective_calls.try_emplace(comm->id(), std::vector<std::string>{call});
395 // are we the first ? add the call
396 if (vec->second.size() == count) {
397 vec->second.emplace_back(call);
398 } else if (vec->second.size() > count) {
399 if (vec->second[count] != call) {
400 XBT_WARN("Collective operation mismatch. For process %ld, expected %s, got %s",
401 simgrid::s4u::this_actor::get_pid(), vec->second[count].c_str(), call.c_str());
402 return MPI_ERR_OTHER;
410 } // namespace simgrid::smpi::utils