1 /* Copyright (c) 2016-2023. 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 #include "smpi_utils.hpp"
9 #include "smpi_config.hpp"
10 #include "src/kernel/xml/platf.hpp"
12 #include "xbt/file.hpp"
14 #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)");
24 double total_benched_time=0;
25 unsigned long total_malloc_size=0;
26 unsigned long total_shared_size=0;
27 unsigned int total_shared_calls=0;
28 struct alloc_metadata_t {
30 unsigned int numcall = 0;
35 struct current_buffer_metadata_t {
36 alloc_metadata_t alloc;
40 alloc_metadata_t max_malloc;
41 simgrid::smpi::F2C* current_handle = nullptr;
42 current_buffer_metadata_t current_buffer1;
43 current_buffer_metadata_t current_buffer2;
45 std::unordered_map<const void*, alloc_metadata_t> allocs;
47 std::unordered_map<int, std::vector<std::string>> collective_calls;
51 namespace simgrid::smpi::utils {
53 void add_benched_time(double time){
54 total_benched_time += time;
57 void account_malloc_size(size_t size, std::string_view file, int line, const void* ptr)
59 if (smpi_cfg_display_alloc()) {
60 alloc_metadata_t metadata;
65 allocs.try_emplace(ptr, metadata);
67 total_malloc_size += size;
68 if(size > max_malloc.size){
69 max_malloc.size = size;
70 max_malloc.line = line;
71 max_malloc.numcall = 1;
72 max_malloc.file = file;
73 } else if (size == max_malloc.size && max_malloc.line == line && max_malloc.file == file) {
79 void account_shared_size(size_t size){
80 if (smpi_cfg_display_alloc()) {
81 total_shared_size += size;
86 void print_time_analysis(double global_time){
87 if (simgrid::config::get_value<bool>("smpi/display-timing")) {
88 XBT_INFO("Simulated time: %g seconds. \n\n"
89 "The simulation took %g seconds (after parsing and platform setup)\n"
90 "%g seconds were actual computation of the application",
91 simgrid_get_clock(), global_time, total_benched_time);
92 if (total_benched_time/global_time>=0.75)
93 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
94 "You may want to use sampling functions or trace replay to reduce this.");
98 static void print_leaked_handles()
100 // Put the leaked non-default handles in a vector to sort them by id
101 std::vector<std::pair<unsigned int, smpi::F2C*>> handles;
102 if (simgrid::smpi::F2C::lookup() != nullptr)
103 std::copy_if(simgrid::smpi::F2C::lookup()->begin(), simgrid::smpi::F2C::lookup()->end(),
104 std::back_inserter(handles),
105 [](auto const& entry) { return entry.first >= simgrid::smpi::F2C::get_num_default_handles(); });
109 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
110 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed MPI handles:";
112 message += "\nHINT: Display types and addresses (n max) with --cfg=smpi/list-leaks:n.\n"
113 "Running smpirun with -wrapper \"valgrind --leak-check=full\" can provide more information";
114 XBT_INFO(message.c_str(), handles.size());
118 // we cannot trust F2C::lookup()->size() > F2C::get_num_default_handles() because some default handles are already
119 // freed at this point
120 bool display_advice = false;
121 std::map<std::string, int, std::less<>> count;
122 for (const auto& [_, elem] : handles) {
123 std::string key = elem->name();
124 if ((not xbt_log_no_loc) && (not elem->call_location().empty()))
125 key += " at " + elem->call_location();
127 display_advice = true;
128 auto& result = count.try_emplace(key, 0).first->second;
132 XBT_WARN("To get more information (location of allocations), compile your code with -trace-call-location flag of "
135 for (const auto& [key, value] : count) {
137 XBT_INFO("leaked handle of type %s", key.c_str());
139 XBT_INFO("%d leaked handles of type %s", value, key.c_str());
144 if (max < count.size())
145 XBT_INFO("(%lu more handle leaks hidden as you wanted to see only %lu of them)", count.size() - max, max);
148 static void print_leaked_buffers()
153 auto max = static_cast<unsigned long>(simgrid::config::get_value<int>("smpi/list-leaks"));
154 std::string message = "Probable memory leaks in your code: SMPI detected %zu unfreed buffers:";
156 message += "display types and addresses (n max) with --cfg=smpi/list-leaks:n.\nRunning smpirun with -wrapper "
157 "\"valgrind --leak-check=full\" can provide more information";
158 XBT_INFO(message.c_str(), allocs.size());
163 // gather by allocation origin (only one group reported in case of no-loc or if trace-call-location is not used)
170 std::map<std::string, struct buff_leak, std::less<>> leaks_aggreg;
171 for (const auto& [_, elem] : allocs) {
172 std::string key = "leaked allocations";
173 if (not xbt_log_no_loc)
174 key = elem.file + ":" + std::to_string(elem.line) + ": " + key;
175 auto& result = leaks_aggreg.try_emplace(key, buff_leak{0, 0, elem.size, elem.size}).first->second;
177 result.total_size += elem.size;
178 if (elem.size > result.max_size)
179 result.max_size = elem.size;
180 else if (elem.size < result.min_size)
181 result.min_size = elem.size;
183 // now we can order by total size.
184 std::vector<std::pair<std::string, buff_leak>> leaks(leaks_aggreg.begin(), leaks_aggreg.end());
185 std::sort(leaks.begin(), leaks.end(),
186 [](auto const& a, auto const& b) { return a.second.total_size > b.second.total_size; });
189 for (const auto& [key, value] : leaks) {
190 if (value.min_size == value.max_size)
191 XBT_INFO("%s of total size %zu, called %d times, each with size %zu", key.c_str(), value.total_size, value.count,
194 XBT_INFO("%s of total size %zu, called %d times, with minimum size %zu and maximum size %zu", key.c_str(),
195 value.total_size, value.count, value.min_size, value.max_size);
200 if (max < leaks_aggreg.size())
201 XBT_INFO("(more buffer leaks hidden as you wanted to see only %lu of them)", max);
204 void print_memory_analysis()
206 if (smpi_cfg_display_alloc()) {
207 print_leaked_handles();
208 print_leaked_buffers();
210 if(total_malloc_size != 0)
211 XBT_INFO("Memory Usage: Simulated application allocated %lu bytes during its lifetime through malloc/calloc calls.\n"
212 "Largest allocation at once from a single process was %zu bytes, at %s:%d. It was called %u times during the whole simulation.\n"
213 "If this is too much, consider sharing allocations for computation buffers.\n"
214 "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",
215 total_malloc_size, max_malloc.size, simgrid::xbt::Path(max_malloc.file).get_base_name().c_str(), max_malloc.line, max_malloc.numcall
219 "Allocations analysis asked, but 0 bytes were allocated through malloc/calloc calls intercepted by SMPI.\n"
220 "The code may not use malloc() to allocate memory, or it was built with SMPI_NO_OVERRIDE_MALLOC");
221 if(total_shared_size != 0)
222 XBT_INFO("%lu bytes were automatically shared between processes, in %u calls\n", total_shared_size, total_shared_calls);
226 void set_current_handle(F2C* handle){
227 current_handle=handle;
230 void print_current_handle(){
232 if(current_handle->call_location().empty())
233 XBT_INFO("To get handle location information, pass -trace-call-location flag to smpicc/f90 as well");
235 XBT_INFO("Handle %s was allocated by a call at %s", current_handle->name().c_str(),
236 (char*)(current_handle->call_location().c_str()));
240 void set_current_buffer(int i, const char* name, const void* buf){
243 if(not current_buffer1.name.empty()){
244 current_buffer1.name="";
246 if(not current_buffer2.name.empty()){
247 current_buffer2.name="";
250 auto meta = allocs.find(buf);
251 if (meta == allocs.end()) {
252 XBT_DEBUG("Buffer %p was not allocated with malloc/calloc", buf);
256 current_buffer1.alloc = meta->second;
257 current_buffer1.name = name;
259 current_buffer2.alloc=meta->second;
260 current_buffer2.name=name;
264 void print_buffer_info()
266 if (not current_buffer1.name.empty())
267 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer1.name.c_str(),
268 current_buffer1.alloc.file.c_str(), current_buffer1.alloc.line, current_buffer1.alloc.size);
269 if (not current_buffer2.name.empty())
270 XBT_INFO("Buffer %s was allocated from %s line %d, with size %zu", current_buffer2.name.c_str(),
271 current_buffer2.alloc.file.c_str(), current_buffer2.alloc.line, current_buffer2.alloc.size);
274 size_t get_buffer_size(const void* buf){
275 auto meta = allocs.find(buf);
276 if (meta == allocs.end()) {
277 //we don't know this buffer (on stack or feature disabled), assume it's fine.
278 return std::numeric_limits<std::size_t>::max();
280 return meta->second.size;
283 void account_free(const void* ptr){
284 if (smpi_cfg_display_alloc()) {
289 int check_collectives_ordering(MPI_Comm comm, const std::string& call)
291 unsigned int count = comm->get_collectives_count();
292 comm->increment_collectives_count();
293 if (auto vec = collective_calls.find(comm->id()); vec == collective_calls.end()) {
294 collective_calls.try_emplace(comm->id(), std::vector<std::string>{call});
296 // are we the first ? add the call
297 if (vec->second.size() == count) {
298 vec->second.emplace_back(call);
299 } else if (vec->second.size() > count) {
300 if (vec->second[count] != call) {
301 XBT_WARN("Collective operation mismatch. For process %ld, expected %s, got %s",
302 simgrid::s4u::this_actor::get_pid(), vec->second[count].c_str(), call.c_str());
303 return MPI_ERR_OTHER;
311 } // namespace simgrid::smpi::utils