Flow123d  DF_profiler_memory_monitor-e298ba0
sys_profiler.hh
Go to the documentation of this file.
1 /*!
2  *
3  * Copyright (C) 2015 Technical University of Liberec. All rights reserved.
4  *
5  * This program is free software; you can redistribute it and/or modify it under
6  * the terms of the GNU General Public License version 3 as published by the
7  * Free Software Foundation. (http://www.gnu.org/licenses/gpl-3.0.en.html)
8  *
9  * This program is distributed in the hope that it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
11  * FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
12  *
13  *
14  * @file sys_profiler.hh
15  * @brief
16  * @todo
17  * - START_GLOBAL_TIMER(tag) - this calls the start_timer, which creates local timer on the correct place in the hierarchy,
18  * further this timer is added to the list of global timers, this contains groups of timers with same tag, and
19  * collect/sum data from these timers in the report.
20  *
21  * - Allow output even during calculation (not complete, but at least some thing)
22  * Report should contain time of start as well as time of creation of the report or time from start of the program.
23  *
24  * - When generating report we has to deal with possibly different trees at every MPI process.
25  *
26  * - test memory profiling
27  * in our own new and xmalloc functions - register allocatied and deallocated memory to active Profiler frame.
28  *
29  * - test in parallel
30  * - extended output:
31  * cas na jedno volani (jina redukce nez pro kumulativni cas, pokud je pocet volani ruzny)
32  * procenta vuci predkovi
33  * code point (az nekde na konci radky)
34  *
35  *
36  * !!! Unfortunately using constexpr is worse (without optimization).
37  * This is probably due to use of static variable for
38  * CodePoint, the access could be slow, and computation of hash is done only once. Actually timing results
39  * are:
40  *
41  * OPTIONS OVERHEAD (compared to call 2x clock())
42  * -g, no c++11 : 18%
43  * -g, c++11 : 60%
44  * -O3,no c++11 : 6%
45  * -O3, c++11 : 6%
46  */
47 
48 #ifndef PROFILER_H
49 #define PROFILER_H
50 
51 #include "global_defs.h"
52 
53 #include <mpi.h>
54 #include <ostream>
55 #include <unordered_map>
56 
57 namespace boost { template <class T> struct hash; }
58 #include <boost/functional/hash/hash.hpp> // for hash
59 #include <boost/ref.hpp>
60 #include <boost/tuple/detail/tuple_basic.hpp> // for get
61 
62 #include <nlohmann/json.hpp>
63 
64 #include "time_point.hh"
65 #include "petscsys.h"
66 #include "simple_allocator.hh"
67 
68 //instead of #include "mpi.h"
69 //mpi declarations follows:
71 public:
72  static int sum(int* val, MPI_Comm comm);
73  static double sum(double* val, MPI_Comm comm);
74  static long sum(long* val, MPI_Comm comm);
75 
76  static int min(int* val, MPI_Comm comm);
77  static double min(double* val, MPI_Comm comm);
78  static long min(long* val, MPI_Comm comm);
79 
80  static int max(int* val, MPI_Comm comm);
81  static double max(double* val, MPI_Comm comm);
82  static long max(long* val, MPI_Comm comm);
83 };
84 
85 // Assuming all compilers support constexpr
86 #define CONSTEXPR_ constexpr
87 
88 using namespace std;
89 
90 
91 // These helper macros are necessary due to use of _LINE_ variable in START_TIMER macro.
92 #define _PASTE(a,b) a ## b
93 #define PASTE(a,b) _PASTE(a, b)
94 
95 
96 
97 /**
98  * \def START_TIMER(tag)
99  *
100  * @brief Starts a timer with specified tag.
101  *
102  * In fact it creates an static constant expression that identifies the point in the code and
103  * contains tag of the involved timer and its hash. Then it creates local variable that
104  * calls @p Profiler::start_timer() in constructor and @p Profiler::stop_timer() in destructor.
105  * This way the timer is automatically closed at the end of current block.
106  *
107  * ATTENTION: This macro expands to two statements so following code is illegal:
108  * @code
109  * if (some_condition) START_TIMER(tag);
110  * @endcode
111  */
112 #ifdef FLOW123D_DEBUG_PROFILER
113 #define START_TIMER(tag) static CONSTEXPR_ CodePoint PASTE(cp_,__LINE__) = CODE_POINT(tag); TimerFrame PASTE(timer_,__LINE__) = TimerFrame( PASTE(cp_,__LINE__) )
114 #else
115 #define START_TIMER(tag)
116 #endif
117 
118 /**
119  * \def START_TIMER_EXT (tag, subtag)
120  *
121  * @brief Starts a timer with specified tag and subtag.
122  *
123  * In fact it creates an static constant expression that identifies the point in the code and
124  * contains tag and subtag of the involved timer and its hash. Then it creates local variable that
125  * calls @p Profiler::start_timer() in constructor and @p Profiler::stop_timer() in destructor.
126  * This way the timer is automatically closed at the end of current block.
127  *
128  * ATTENTION: This macro expands to two statements so following code is illegal:
129  * @code
130  * if (some_condition) START_TIMER_EXT(tag, subtag);
131  * @endcode
132  */
133 #ifdef FLOW123D_DEBUG_PROFILER
134 #define START_TIMER_EXT(tag, subtag) static CONSTEXPR_ CodePoint PASTE(cp_,__LINE__) = CODE_POINT_EXT(tag, subtag); TimerFrame PASTE(timer_,__LINE__) = TimerFrame( PASTE(cp_,__LINE__) )
135 #else
136 #define START_TIMER_EXT(tag, subtag)
137 #endif
138 
139 /**
140  * \def END_TIMER(tag)
141  *
142  * @brief Ends a timer with specified tag.
143  *
144  * Use only if you want to end timer before the end of block. Again this expands into two lines, see ATTENTION in previous macro.
145  */
146 #ifdef FLOW123D_DEBUG_PROFILER
147 #define END_TIMER(tag) static CONSTEXPR_ CodePoint PASTE(cp_,__LINE__) = CODE_POINT(tag); Profiler::instance()->stop_timer( PASTE(cp_,__LINE__) )
148 #else
149 #define END_TIMER(tag)
150 #endif
151 
152 /**
153  * \def START_MEMORY_MONITORING
154  *
155  * @brief Allow to use monitoring only in part of code.
156  */
157 #ifdef FLOW123D_DEBUG_PROFILER
158 #define START_MEMORY_MONITORING Profiler::instance()->start_memory_monitoring()
159 #else
160 #define START_MEMORY_MONITORING
161 #endif
162 
163 /**
164  * \def END_START_TIMER(tag)
165  *
166  * Ends current timer and starts the new one with given tag. Again this expands into two lines, see ATTENTION in previous macro.
167  */
168 #ifdef FLOW123D_DEBUG_PROFILER
169 #define END_START_TIMER(tag) Profiler::instance()->stop_timer(); START_TIMER(tag);
170 #else
171 #define END_START_TIMER(tag)
172 #endif
173 
174 
175 /**
176  * \def ADD_CALLS(n_calls)
177  *
178  * @brief Increase number of calls in actual timer.
179  *
180  * Some time you want to measure a performance of a cycle with body that is below resolution of the Timer implementation.
181  * If you know number of cycles, you can use this macro in following way:
182  *
183  * @code
184  * START_TIMER("cycle");
185  * unsigned int i;
186  * for(i =0; i<1000000; i++) i*i*i;
187  * ADD_CALLS(i);
188  * END_TIMER("cycle");
189  * @endcode
190  *
191  * In the profiler report you get the total time spent in the cycle, and time per one call which will be average
192  * time spent in the body of the cycle.
193  */
194 #ifdef FLOW123D_DEBUG_PROFILER
195 #define ADD_CALLS(n_calls) Profiler::instance()->add_calls(n_calls)
196 #else
197 #define ADD_CALLS(n_calls)
198 #endif
199 
200 
201 #ifdef FLOW123D_DEBUG_PROFILER
202 #define CUMUL_TIMER(tag) Profiler::instance()->find_timer(tag).cumulative_time()
203 #else
204 #define CUMUL_TIMER(tag) 0
205 #endif
206 
207 //////////////////////////////////////////////////////////////////////////////////////////////
208 #ifdef FLOW123D_DEBUG_PROFILER
209 
210 /**
211  * Variable which represents value when no subtag was specified in CodePoint class
212  */
213 #define PROFILER_EMPTY_SUBTAG ""
214 
215 /**
216  * Variable used for default value in hash process
217  */
218 #define PROFILER_HASH_DEFAULT 0
219 
220 /**
221  * @brief Function for compile-time hash computation. (Needs C++x11 standard.)
222  * Input, @p str, is constant null terminated string, result is unsigned int (usually 4 bytes).
223  * Function has to be recursive, since standard requires that the body consists only from the return statement.
224  *
225  * SALT is hash for the empty string. Currently zero for simpler testing.
226  */
227 inline CONSTEXPR_ unsigned int str_hash(const char * str, unsigned int default_value) {
228  #define SALT 0 //0xef50e38f
229  return (*str == 0 ? SALT : default_value + str_hash(str+1, PROFILER_HASH_DEFAULT) * 101 + (unsigned int)(*str) );
230 }
231 
232 /**
233  * Macro to generate constexpr CodePoint object.
234  */
235 #define CODE_POINT(tag) CodePoint(tag, __FILE__, __func__, __LINE__)
236 
237 /**
238  * Macro to generate constexpr CodePoint object.
239  */
240 #define CODE_POINT_EXT(tag, subtag) CodePoint(tag, subtag, __FILE__, __func__, __LINE__)
241 
242 
243 
244 
245 /**
246  * @brief Class that represents point in the code.
247  *
248  * This class allow construction at compile time. And includes the information about the code point as well
249  * as the 'tag' of the timer and cimpile-time computed hashes of this 'tag'. The @p hash_ is long one with
250  * very small probability of collisions - this we use for comparison of tags. The @p hash_idx_ is the long hash modulo
251  * length of the array of Timer's children, this is used for fast loop up into this array that servers as a simple hash table.
252  */
253 class CodePoint {
254 public:
255  CONSTEXPR_ CodePoint(const char *tag, const char * file, const char * func, const unsigned int line)
256  : tag_(tag), subtag_(PROFILER_EMPTY_SUBTAG), file_(file), func_(func), line_(line),
257  hash_(str_hash(tag, PROFILER_HASH_DEFAULT)),
258  hash_idx_( str_hash(tag, PROFILER_HASH_DEFAULT)%max_n_timer_childs )
259  {};
260  CONSTEXPR_ CodePoint(const char *tag, const char *subtag, const char * file, const char * func, const unsigned int line)
261  : tag_(tag), subtag_(subtag), file_(file), func_(func), line_(line),
262  hash_(str_hash(subtag, str_hash(tag, PROFILER_HASH_DEFAULT))),
263  hash_idx_( str_hash(subtag, str_hash(tag, PROFILER_HASH_DEFAULT))%max_n_timer_childs )
264  {};
265 
266  /// Size of child arrays in timer nodes.
267  static const unsigned int max_n_timer_childs=13;
268 
269  /// Tag of the code point.
270  const char * const tag_;
271 
272  /// Subtag of the code point.
273  const char * const subtag_;
274 
275  /// file name of the code point
276  const char * const file_;
277 
278  /// file name of the code point
279  const char * const func_;
280 
281  /// file name of the code point
282  const unsigned int line_;
283 
284  /// Full 32-bit hash of the tag ( practically no chance of collision)
285  unsigned int hash_;
286 
287  /// Hash modulo size of array of timer childs ( we have to check full hash to prevent collision)
288  unsigned int hash_idx_;
289 };
290 
291 
292 
293 /**
294  * @brief Class for profiling tree nodes.
295  *
296  * One Timer represents one particular time frame in the execution tree.
297  * It collects information about total time, number of calls, allocated and deallocated memory.
298  *
299  * It should be accessed only through Profiler, which is its friend class.
300  *
301  * TODO: for better performance: move copy hash_ and hash_idx_ into Timer since CodePoint are in static
302  * variables, that may be slower to acces.
303  *
304  */
305 class Timer {
306 
307 
308 public:
309  /// Size of array @p child_timers, the hash table containing descendants in the call tree.
310  static const unsigned int max_n_childs=CodePoint::max_n_timer_childs;
311 
312  /**
313  * Creates the timer node object. Should not be called directly, but through the START_TIMER macro.
314  */
315  Timer(const CodePoint &cp, int parent);
316 
317 
318  /**
319  * Start the timer. If it is already started, just increase number of starts (recursions) and calls.
320  */
321  void start();
322 
323  /**
324  * If number of starts (recursions) drop back to zero, we stop the timer and add the period to the cumulative time.
325  * This method do not take care of its childs (it has no access to the other timers).
326  * When the parameter 2p forced is 'true', we stop the timer immediately regardless the number of recursions.
327  * Returns true if the timer is not closed (recursions didn't drop to zero yet).
328  */
329  bool stop(bool forced = false);
330 
331 
332  /// Getter for the 'tag'.
333  inline string tag() const {
334  string buf(code_point_->tag_);
335  buf.append(code_point_->subtag_);
336  return buf;
337  }
338 
339  /// Returns true if the timer is open, number of starts (recursions) is nonzero.
340  inline bool running() const
341  { return start_count >0; }
342 
343  /// Returns string with description of the code point where the timer was first started.
344  std::string code_point_str() const;
345 
346  /**
347  * Returns cumulative time of the timer in seconds.
348  */
349  double cumulative_time() const;
350 
351  /*
352  * Adds given index @p child_index of the timer @p child to the correct place in the hash table.
353  */
354  void add_child(int child_index, const Timer &child);
355 
356  /**
357  * Start PETSC memory monitoring in actual time frame
358  */
359  void start_memory_monitoring();
360 
361  inline bool memory_monitor_on() const {
362  return memory_monitor_on_;
363  }
364 
365 
366 protected:
367 
368  /**
369  * Pauses current timer, save measured petsc memory information util resume.
370  * We get Petsc maximum memory usage via PetscMemoryGetMaximumUsage call
371  * and save this value into temp value. (we override local maximum if temp
372  * value is greater).
373  *
374  * Method is not used now. It can be used if necessary.
375  */
376  void pause();
377  /**
378  * Resume current timer. e tell Petsc to monitor the maximum memory
379  * usage once again. We call PetscMemorySetGetMaximumUsage so later in
380  * resume() method will PetscMemoryGetMaximumUsage method work.
381  *
382  * Method is not used now. It can be used if necessary.
383  */
384  void resume();
385 
386  /**
387  * Start time when frame opens.
388  */
389  TimePoint start_time;
390  /**
391  * Cumulative time spent in the frame.
392  */
393  double cumul_time;
394  /**
395  * Total number of opening of the frame.
396  */
397  unsigned int call_count;
398  /**
399  * Number of recursive openings.
400  */
401  unsigned int start_count;
402 
403 
404  /**
405  * Code point of the first START_TIMER for the particular tag. The 'tag' identifies timer
406  * and is used in reported profiler table.
407  */
408  const CodePoint *code_point_;
409  /// Full tag hash. Copy from code_point_.
410  unsigned int full_hash_;
411  /// Hash modulo size of array of timer childs. Copy from code_point_.
412  unsigned int hash_idx_;
413 
414  /**
415  * Index of the parent timer node in the tree. Negative value means 'not set'.
416  */
417  int parent_timer;
418  /**
419  * Indices of the child timers in the Profiler::timers_ vector. Negative values means 'not set'.
420  */
421  int child_timers[max_n_childs];
422 
423  /**
424  * Total number of bytes allocated in this frame. After
425  * Profiler::propagate_timers call will also contain values from children.
426  */
427  size_t total_allocated_;
428  /**
429  * Total number of bytes deallocated in this frame. After
430  * Profiler::propagate_timers call, will also contain values from children.
431  */
432  size_t total_deallocated_;
433  /**
434  * Maximum number of bytes allocated at one time in this frame. After
435  * Profiler::propagate_timers call, maximum value will be taken from this
436  * Timer and also from all children Timers.
437  */
438  size_t max_allocated_;
439  /**
440  * Current number of bytes allocated in this frame at the same time.
441  * This value is used to monitor maximum bytes allocated. When notify_free
442  * and notify_malloc is called this values is changed and new maximum
443  * is tested.
444  */
445  size_t current_allocated_;
446 
447  /**
448  * Number of times new/new[] operator was used in this scope
449  */
450  int alloc_called;
451  /**
452  * Number of times delete/delete[] operator was used in this scope
453  */
454  int dealloc_called;
455  /**
456  * True if memory monitoring was switched on by macro START_MEMORY_MONITORING.
457  */
458  bool memory_monitor_on_;
459 
460  #ifdef FLOW123D_HAVE_PETSC
461  /**
462  * Number of bytes used by Petsc at the start of time-frame
463  */
464  PetscLogDouble petsc_start_memory;
465  /**
466  * Number of bytes used by Petsc at the end of time-frame
467  */
468  PetscLogDouble petsc_end_memory;
469  /**
470  * Difference between start and end of a petsc memory usage
471  */
472  PetscLogDouble petsc_memory_difference;
473  /**
474  * Maximum amount of memory used that was PetscMalloc()ed at any time
475  * during this run.
476  *
477  * The memory usage reported here includes all Fortran arrays (that may be
478  * used in application-defined sections of code).
479  */
480  PetscLogDouble petsc_peak_memory;
481  /**
482  * Local maximum amount of memory used that was PetscMalloc()ed
483  * used during time-frame pause/resume. Auxilary variable for storing
484  * local memory used when pause is called.
485  */
486  PetscLogDouble petsc_local_peak_memory;
487  #endif // FLOW123D_HAVE_PETSC
488 
489  friend class Profiler;
490  friend std::ostream & operator <<(std::ostream&, const Timer&);
491 
492  /**
493  * if under unit testing, specify friend so protected members can be tested
494  */
495  #ifdef __UNIT_TEST__
496  friend ProfilerTest;
497  #endif /* __UNIT_TEST__ */
498 
499 };
500 
501 /*
502 struct SimpleTranslator {
503  typedef std::string internal_type;
504  typedef int external_type;
505 
506  // Converts a string to int
507  boost::optional<external_type> get_value(const internal_type& str) {
508  return boost::optional<external_type>(std::stoi(str));
509  }
510 
511  // Converts a bool to string
512  boost::optional<internal_type> put_value(const external_type& i){
513  return boost::optional<internal_type>(std::to_string(i));
514  }
515 };
516 
517 namespace boost {
518 namespace property_tree {
519 
520 template<typename Ch, typename Traits, typename Alloc>
521 struct translator_between<std::basic_string< Ch, Traits, Alloc >, int> {
522  typedef SimpleTranslator type;
523 };
524 
525 
526 } // namespace property_tree
527 } // namespace boost
528 */
529 /**
530  *
531  * @brief Main class for profiling by measuring time intervals.
532  *
533  * These time intervals form a tree structure where each interval is represented
534  * by a Timer object. The root node of the tree is automatically created and
535  * started after creating the Profiler object and cannot be stopped manually.
536  *
537  * The class implements a singleton pattern and all the functions are accessible trough
538  * Profiler::instance(), but in most cases the programmer will access the profiler
539  * functions via the #START_TIMER and #END_TIMER macros. The #START_TIMER macro
540  * is responsible for the fact that we don't have to call #END_TIMER macro to stop the timer and
541  * the timer will be stopped at the end of the block in which #START_TIMER was used.
542  * These macros internally use the TimerFrame objects and the programmer should
543  * not use the TimerFrame objects directly.
544  *
545  * By using #SET_TIMER_SUBFRAMES macro, the programmer can specify the number of subframes (eg. iterations)
546  * for the currently active timer.
547  *
548  *
549  * Currently the Profiler system is not thread safe. No idea how to do this.
550  *
551  */
552 class Profiler {
553 public:
554  /**
555  * Returns unique Profiler object.
556  * if clear flag is set, will delete profiiler isntance
557  */
558  static Profiler* instance(bool clear = false);
559  /**
560  * Sets task specific information. The string @p description with textual description of the task and the
561  * number of elements of the mesh (parameter @p size). This is used for weak scaling graphs so it should
562  * measure size of the task of the same type (same description).
563  *
564  */
565  void set_task_info(string description, int size);
566  /**
567  * Sets informations about program version. This consists of @p program_version (includes program name), @p branch in the repository or rather full URL of the branch,
568  * and SVN @p revision (or hash for GIT).
569  *
570  */
571  void set_program_info(string program_name, string program_version, string branch, string revision, string build);
572 
573 
574  /**
575  * Starts a timer with code point, tag and hashes specified by CodePoint object @p cp.
576  * If the timer is not already created, it creates a new one. It returns index of
577  * the actual timer.
578  */
579  int start_timer(const CodePoint &cp);
580  /**
581  * Stops actual timer. It check if the hash of given code point match hash of the
582  * tag of actual timer node. If not we print out warning and try to find the correct tag
583  * towards the tree root closing all nodes we pass through.
584  *
585  * If FLOW123D_DEBUG is set, we check that all children are closed.
586  */
587  void stop_timer(const CodePoint &cp);
588 
589  /**
590  * Stop timer with index given by @p timer_index. If this is not equal to @p actual_node, we
591  * traverse the tree towards root while force closing nodes by the way.
592  *
593  * Negative @p timer_index means close @p actual_node
594  */
595  void stop_timer(int timer_index = -1);
596 
597  /**
598  * Start memory monitoring if it is switched off.
599  *
600  * Memory monitoring is automatically turn off at the end of actual tag.
601  */
602  void start_memory_monitoring();
603 
604  /**
605  * Adds @p n_calls - 1 to the total number of calls of the current timer. Minus one, since one call is counted when
606  * timer was started. You should use macro ADD_CALLS above.
607  */
608  void add_calls(unsigned int n_calls);
609  /**
610  * Notification about allocation of given size.
611  * Increase total allocated memory in current profiler frame.
612  */
613  void notify_malloc(const size_t size, const long p);
614  /**
615  * Notification about freeing memory of given size.
616  * Increase total deallocated memory in current profiler frame.
617  */
618  void notify_free(const long p);
619 
620  /**
621  * Return average profiler timer resolution in seconds
622  * based on 100 measurements
623  */
624  static double get_resolution ();
625 
626  /**
627  * Find a first timer matching the tag.
628  * O(n) complexity.
629  */
630  Timer find_timer(string tag);
631 
632 
633 
634 
635 #ifdef FLOW123D_HAVE_MPI
636  /**
637  * @brief Output current timing information into the given stream.
638  *
639  * COLECTIVE - all processes in the communicator have to call this
640  * method. All timers are finished, all processes are synchronized, collect
641  * profiling informations are collected and written to the given stream.
642  *
643  * Pass through the profiling tree (collective over processors)
644  * Print cumulative times average, balance (max/min), count (denote differences)
645  *
646  */
647  void output(MPI_Comm comm, std::ostream &os);
648 
649  /**
650  * Same as previous, but output to the file with default name: "profiler_info_YYMMDD_HH::MM:SS.log".
651  * Empty body if macro FLOW123D_DEBUG_PROFILER is not defined.
652  */
653  string output(MPI_Comm comm, string profiler_path = "");
654 
655 #endif /* FLOW123D_HAVE_MPI */
656  /**
657  * @brief Output current timing information into the given stream.
658  *
659  * It temporally stops all timers, synchronize all processes, collect
660  * profiling informations and write it to the given stream.
661  *
662  * Pass through the profiling tree (collective over processors)
663  * Print cumulative times average, balance (max/min), count (denote differences)
664  *
665  */
666  void output(std::ostream &os);
667 
668  /**
669  * Same as previous, but output to the file with default name: "profiler_info_YYMMDD_HH::MM:SS.log".
670  * Empty body if macro FLOW123D_DEBUG_PROFILER is not defined.
671  */
672  string output(string profiler_path = "");
673 
674  /**
675  * Method will transform last profiler json file to desired format
676  */
677 // void transform_profiler_data (const string &output_file_suffix, const string &formatter);
678  /**
679  * Stop all timers and destroys the Profiler object.
680  * If you want some output call @p output method just before.
681  */
682  static void uninitialize();
683 
684  /**
685  * Class-specific allocation function new. Called by the usual
686  * single-object new-expressions if allocating an object of type Profiler.
687  */
688  static void* operator new (size_t sz);
689  /**
690  * Class-specific allocation function delete. Deallocates storage
691  * previously allocated by a matching operator new. These deallocation
692  * functions are called by delete-expressions.
693  */
694  static void operator delete (void* p);
695  /// Sized deallocator, doesthe same as operator delete (void* p)
696  static void operator delete (void* p, std::size_t);
697 
698  /**
699  * Public setter to turn on/off memory monitoring
700  * @param global_monitor whether to turn global monitoring on or off
701  * @param petsc_monitor petsc monitoring
702  */
703  void static set_memory_monitoring(const bool global_monitor);
704 
705  /**
706  * Public getter to memory monitoring
707  * @return memory monitoring status
708  */
709  inline bool static get_global_memory_monitoring() {
710  return global_monitor_memory;
711  }
712 
713  /**
714  * Run calibration frame "UNIT PAYLOAD".
715  * That should be about 100x timer resolution.
716  */
717  void calibrate();
718 
719  /**
720  * Time of a unit payload, result of a single measurement. Can be used for raw calibration.
721  */
722  double calibration_time() {
723  if (calibration_time_ < 0) calibrate();
724  return calibration_time_;
725  }
726  /**
727  * if under unit testing, specify friend so protected members can be tested
728  */
729  #ifdef __UNIT_TEST__
730  friend ProfilerTest;
731  #endif /* __UNIT_TEST__ */
732 
733 
734 protected:
735 
736  /**
737  * Whether to monitor operator 'new/delete'
738  */
739  static bool global_monitor_memory;
740 
741  /**
742  * Whether to monitor petsc memory usage
743  */
744  static bool petsc_monitor_memory;
745 
746  /**
747  * When creating Profiler also reserve some bytes in malloc_map so overhead
748  * of creating single items is lowered. This value is passed as parameter in
749  * map.reserve() method so it indicates how many objects (pointers) are
750  * allocated at first.
751  */
752  static const long malloc_map_reserve;
753 
754 
755  /**
756  * Method will propagate values from children timers to its parents
757  */
758  void propagate_timers ();
759 
760  /**
761  * Method for exchanging metrics from child timer to its parent timer
762  */
763  void accept_from_child (Timer &parent, Timer &child);
764 
765  /**
766  * Try to find timer with tag (in fact only its 32-bit hash) from given code point @p cp.
767  * Returns -1 if it is not found otherwise it returns its index.
768  */
769  int find_child(const CodePoint &cp);
770 
771 
772  /**
773  * Method will prepare construct specific details about the run (time start and time end)
774  * and write them along with basic informations about the run (name, description, ...)
775  * into nlohmann:json object
776  */
777  void output_header (nlohmann::json &root, int mpi_size);
778 
779  /**
780  * Open a new file for profiler output with default name based on the
781  * actual time and date. Returns a pointer to the stream of the output file.
782  */
783  //std::shared_ptr<std::ostream> get_output_stream(string path);
784 
785  /// Vector of all timers. Whole tree is stored in this array.
787 
788  /// Index of the actual timer node. Negative value means 'unset'.
789  unsigned int actual_node;
790 
791  /// MPI communicator used for final reduce of the timer node tree.
792  //MPI_Comm communicator_;
793  /// MPI_rank
794  //int mpi_rank_;
795 
796  /**
797  * flag indicating that collection of timer details will be
798  * using MPI
799  bool mpi_used;
800  */
801  // header informations
802 
803  /// Some measure of the size of the task in the set of the tasks that differs
804  /// only by size - used for scaling tests.
805  int task_size_;
806  /// Task description and identifier in possible database of all Profiler results.
807  string task_description_;
808  /// Time and date of the start of the task solution. In fact start of the Profiler.
809  time_t start_time;
810 
811  /// Name of the program.
812  string flow_name_;
813  /// Version of the program.
814  string flow_version_;
815  /// Http address of the branch in a repository.
816  string flow_branch_;
817  /// Revision or GIT hash.
818  string flow_revision_;
819  /// Build date and time.
820  string flow_build_;
821  /// Variable which stores last json log filepath
822  //string json_filepath;
823 
824  Timer none_timer_;
825 
826  /// Time of a unit payload, result of single measurement. Can be used for raw calibration.
827  double calibration_time_;
828 
829 protected:
830  /**
831  * Use DFS to pass through the tree and collect information about all timers reduced from the processes in the communicator.
832  * For every timer the information strings are stored in the struct TimerInfo in order to pad fields correctly
833  * to have alligned columns on the output. The alligning is performed in the output() method.
834  */
835  template<typename ReduceFunctor>
836  void add_timer_info(ReduceFunctor reduce, nlohmann::json* node, int timer_idx, double parent_time);
837 
838  //Profiler(MPI_Comm comm); // private constructor
839  Profiler(); // private constructor
840  Profiler(Profiler const&); // copy constructor is private
841  Profiler & operator=(Profiler const&); // assignment operator is private
842 
843 };
844 
845 
846 
847 
848 
849 
850 /**
851  *
852  * @brief Class for automatic timer closing. This class is used by #START_TIMER macro
853  * and is responsible for the fact that we don't have to call #END_TIMER macro to stop the timer,
854  * the timer will be stopped at the end of the block in which #START_TIMER was used.
855  *
856  * The main idea of the approach described is that the TimerFrame variable will be destroyed
857  * at the end of the block where #START_TIMER macro was used. In order to work properly
858  * in situations where #END_TIMER was used to stop the timer manually before (but there is still the
859  * variable which will be later destroyed), we have to store references to these variables and
860  * destroy them on-demand.
861  *
862  * TODO:
863  * Should only contain pointer to the Timer. And destructor, that close the timer.
864  */
865 class TimerFrame {
866 private:
867  int const timer_index_;
868 public:
869  inline TimerFrame(const CodePoint &cp)
870  : timer_index_( Profiler::instance()->start_timer(cp) )
871  {}
872 
873  ~TimerFrame() {
874  Profiler::instance()->stop_timer(timer_index_);
875  }
876 };
877 
878 
879 /**
880  * Simple class providing static map variable storing address and alloc size
881  */
882 // gcc version 4.9 and lower has following bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59751
883 // fix in version 4.9: https://gcc.gnu.org/gcc-4.9/changes.html#cxx
884 // typedef unordered_map<long, int, hash<long>, equal_to<long>, internal::SimpleAllocator<pair<const long, int>>> unordered_map_with_alloc;
885 typedef std::unordered_map<long, int, boost::hash<long>, equal_to<long>, internal::SimpleAllocator<std::pair<const long, int>>> unordered_map_with_alloc;
886 class MemoryAlloc {
887 public:
888  /**
889  * Create static map containing <allocation address, allocation size> pairs
890  * map is used for storing allocations and deallocations of all object not
891  * related to profiler after profiler initialization phase
892  */
893  static unordered_map_with_alloc & malloc_map();
894 };
895 
896 
897 
898 
899 #else // FLOW123D_DEBUG_PROFILER
900 
901 
902 // dummy declaration of Profiler class
903 class Profiler {
904 public:
905  static Profiler* instance(bool clear = false);
906 
907  void set_task_info(string, int)
908  {}
909  void set_program_info(string, string, string, string, string)
910  {}
911  void notify_malloc(const size_t )
912  {}
913  void notify_free(const size_t )
914  {}
915  void output(MPI_Comm, ostream &)
916  {}
917  string output(MPI_Comm, string)
918  {return "";}
919  void output(std::ostream &)
920  {}
921  string output(string)
922  {return "";}
923 // void output(MPI_Comm)
924 // {}
925 // string output()
926 // {}
927 // void transform_profiler_data(const string &, const string &)
928 // {}
929  double get_resolution () const
930  { return 0.0; }
931  const char *actual_tag() const
932  { return NULL; }
933  inline unsigned int actual_count() const
934  { return 0; }
935  inline double actual_cumulative_time() const
936  { return 0.0; }
937  static void uninitialize();
938  void calibrate();
939  double calibration_time() {
940  return -2;
941  }
942  static void set_memory_monitoring(bool)
943  {}
944 private:
945  Profiler() {}
946 };
947 
948 
949 
950 
951 #endif
952 
953 
954 #endif
955 
static int sum(int *val, MPI_Comm comm)
Definition: sys_profiler.cc:42
static int min(int *val, MPI_Comm comm)
Definition: sys_profiler.cc:60
static int max(int *val, MPI_Comm comm)
Definition: sys_profiler.cc:78
double calibration_time()
void notify_malloc(const size_t)
void output(MPI_Comm, ostream &)
void notify_free(const size_t)
static void set_memory_monitoring(bool)
static Profiler * instance(bool clear=false)
double actual_cumulative_time() const
void output(std::ostream &)
string output(string)
unsigned int actual_count() const
void set_program_info(string, string, string, string, string)
const char * actual_tag() const
string output(MPI_Comm, string)
void set_task_info(string, int)
void calibrate()
double get_resolution() const
Definition: memory.cc:33
a class to store JSON values
Definition: json.hpp:174
Global macros to enhance readability and debugging, general constants.
int MPI_Comm
Definition: mpi.h:141
#define CONSTEXPR_
Definition: sys_profiler.hh:86
STREAM & operator<<(STREAM &s, UpdateFlags u)