sys_profiler.cc

Go to the documentation of this file.

/*!
 *
 * Copyright (C) 2015 Technical University of Liberec.  All rights reserved.
 * 
 * This program is free software; you can redistribute it and/or modify it under
 * the terms of the GNU General Public License version 3 as published by the
 * Free Software Foundation. (http://www.gnu.org/licenses/gpl-3.0.en.html)
 * 
 * This program is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
 * FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more details.
 *
 * 
 * @file    sys_profiler.cc
 * @ingroup system
 * @brief   Profiler
 */
 
 
// Fat header
 
#include <fstream>
#include <iomanip>
#include <sys/param.h>
#include <unordered_map>
 
#ifdef FLOW123D_HAVE_PYTHON
    #include "Python.h"
#endif // FLOW123D_HAVE_PYTHON
 
#include "sys_profiler.hh"
#include "system/system.hh"
#include "system/python_loader.hh"
#include <iostream>
#include <boost/format.hpp>
 
#include "system/file_path.hh"
#include "mpi.h"
#include "time_point.hh"
 
/*
 * These should be replaced by using boost MPI interface
 */
int MPI_Functions::sum(int* val, MPI_Comm comm) {
        int total = 0;
        MPI_Reduce(val, &total, 1, MPI_INT, MPI_SUM, 0, comm);
        return total;
    }
 
double MPI_Functions::sum(double* val, MPI_Comm comm) {
        double total = 0;
        MPI_Reduce(val, &total, 1, MPI_DOUBLE, MPI_SUM, 0, comm);
        return total;
    }
    
long MPI_Functions::sum(long* val, MPI_Comm comm) {
        long total = 0;
        MPI_Reduce(val, &total, 1, MPI_LONG, MPI_SUM, 0, comm);
        return total;
    }
 
int MPI_Functions::min(int* val, MPI_Comm comm) {
        int min = 0;
        MPI_Reduce(val, &min, 1, MPI_INT, MPI_MIN, 0, comm);
        return min;
    }
 
double MPI_Functions::min(double* val, MPI_Comm comm) {
        double min = 0;
        MPI_Reduce(val, &min, 1, MPI_DOUBLE, MPI_MIN, 0, comm);
        return min;
    }
    
long MPI_Functions::min(long* val, MPI_Comm comm) {
        long min = 0;
        MPI_Reduce(val, &min, 1, MPI_LONG, MPI_MIN, 0, comm);
        return min;
    }
 
int MPI_Functions::max(int* val, MPI_Comm comm) {
        int max = 0;
        MPI_Reduce(val, &max, 1, MPI_INT, MPI_MAX, 0, comm);
        return max;
    }
 
double MPI_Functions::max(double* val, MPI_Comm comm) {
        double max = 0;
        MPI_Reduce(val, &max, 1, MPI_DOUBLE, MPI_MAX, 0, comm);
        return max;
    }
    
long MPI_Functions::max(long* val, MPI_Comm comm) {
        long max = 0;
        MPI_Reduce(val, &max, 1, MPI_LONG, MPI_MAX, 0, comm);
        return max;
    }
 
 
#ifdef FLOW123D_DEBUG_PROFILER
/*********************************************************************************************
 * Implementation of class Timer
 */
 
const int timer_no_child=-1;
 
Timer::Timer(const CodePoint &cp, int parent)
: start_time(TimePoint()),
  cumul_time(0.0),
  call_count(0),
  start_count(0),
  code_point_(&cp),
  full_hash_(cp.hash_),
  hash_idx_(cp.hash_idx_),
  parent_timer(parent),
  total_allocated_(0),
  total_deallocated_(0),
  max_allocated_(0),
  current_allocated_(0),
  alloc_called(0),
  dealloc_called(0)
#ifdef FLOW123D_HAVE_PETSC
, petsc_start_memory(0),
  petsc_end_memory (0),
  petsc_memory_difference(0),
  petsc_peak_memory(0),
  petsc_local_peak_memory(0)
#endif // FLOW123D_HAVE_PETSC
{
    for(unsigned int i=0; i< max_n_childs ;i++)   child_timers[i]=timer_no_child;
}
 
 
/**
 * Debug information of the timer
 */
ostream & operator <<(ostream& os, const Timer& timer) {
    os << "  Timer: " << timer.tag() << endl;
    os << "          malloc: " << timer.total_allocated_ << endl;
    os << "          dalloc: " << timer.total_deallocated_ << endl;
    #ifdef FLOW123D_HAVE_PETSC
    os << "          start: " << timer.petsc_start_memory << endl;
    os << "          stop : " << timer.petsc_end_memory << endl;
    os << "          diff : " << timer.petsc_memory_difference << " (" << timer.petsc_end_memory - timer.petsc_start_memory << ")" << endl;
    os << "          peak : " << timer.petsc_peak_memory << " (" << timer.petsc_local_peak_memory << ")" << endl;
    #endif // FLOW123D_HAVE_PETSC
    os << endl;
    return os;
}
 
 
double Timer::cumulative_time() const {
    return cumul_time;
}
 
void Profiler::accept_from_child(Timer &parent, Timer &child) {
    int child_timer = 0;
    for (unsigned int i = 0; i < Timer::max_n_childs; i++) {
        child_timer = child.child_timers[i];
        if (child_timer != timer_no_child) {
            // propagate metrics from child to parent
            accept_from_child(child, timers_[child_timer]);
        }
    }
    // compute totals by adding values from child
    parent.total_allocated_ += child.total_allocated_;
    parent.total_deallocated_ += child.total_deallocated_;
    parent.alloc_called += child.alloc_called;
    parent.dealloc_called += child.dealloc_called;
    
#ifdef FLOW123D_HAVE_PETSC
    if (petsc_monitor_memory) {
        // add differences from child
        parent.petsc_memory_difference += child.petsc_memory_difference;
        parent.current_allocated_ += child.current_allocated_;
        
        // when computing maximum, we take greater value from parent and child
        // PetscMemoryGetCurrentUsage always return absolute (not relative) value
        parent.petsc_peak_memory = max(parent.petsc_peak_memory, child.petsc_peak_memory);
    }
#endif // FLOW123D_HAVE_PETSC
    
    parent.max_allocated_ = max(parent.max_allocated_, child.max_allocated_);
}
 
 
void Timer::pause() {
#ifdef FLOW123D_HAVE_PETSC
    if (Profiler::get_petsc_memory_monitoring()) {
        // get the maximum resident set size (memory used) for the program.
        PetscMemoryGetMaximumUsage(&petsc_local_peak_memory);
        if (petsc_peak_memory < petsc_local_peak_memory)
            petsc_peak_memory = petsc_local_peak_memory;
    }
#endif // FLOW123D_HAVE_PETSC
}
 
void Timer::resume() {
#ifdef FLOW123D_HAVE_PETSC
    if (Profiler::get_petsc_memory_monitoring()) {
        // tell PETSc to monitor the maximum memory usage so
        //   that PetscMemoryGetMaximumUsage() will work.
        PetscMemorySetGetMaximumUsage();
    }
#endif // FLOW123D_HAVE_PETSC
}
 
void Timer::start() {
#ifdef FLOW123D_HAVE_PETSC
    if (Profiler::get_petsc_memory_monitoring()) {
        // Tell PETSc to monitor the maximum memory usage so
        //   that PetscMemoryGetMaximumUsage() will work.
        PetscMemorySetGetMaximumUsage();
        PetscMemoryGetCurrentUsage (&petsc_start_memory);
    }
#endif // FLOW123D_HAVE_PETSC
    
    if (start_count == 0) {
        start_time = TimePoint();
    }
    call_count++;
    start_count++;
}
 
 
 
bool Timer::stop(bool forced) {
#ifdef FLOW123D_HAVE_PETSC
    if (Profiler::get_petsc_memory_monitoring()) {
        // get current memory usage
        PetscMemoryGetCurrentUsage (&petsc_end_memory);
        petsc_memory_difference += petsc_end_memory - petsc_start_memory;
        
        // get the maximum resident set size (memory used) for the program.
        PetscMemoryGetMaximumUsage(&petsc_local_peak_memory);
        if (petsc_peak_memory < petsc_local_peak_memory)
            petsc_peak_memory = petsc_local_peak_memory;
    }
#endif // FLOW123D_HAVE_PETSC
    
    if (forced) start_count=1;
 
    if (start_count == 1) {
        cumul_time += (TimePoint() - start_time);
        start_count--;
        return true;
    } else {
        start_count--;
    }
    return false;
}
 
 
 
void Timer::add_child(int child_index, const Timer &child)
{
    unsigned int idx = child.hash_idx_;
    if (child_timers[idx] != timer_no_child) {
        // hash collision, find first empty place
        unsigned int i=idx;
        do {
            i=( i < max_n_childs ? i+1 : 0);
        } while (i!=idx && child_timers[i] != timer_no_child);
        ASSERT_PERMANENT(i!=idx)(tag()).error("Too many children of the timer");
        idx=i;
    }
    //DebugOut().fmt("Adding child {} at index: {}\n", child_index, idx);
    child_timers[idx] = child_index;
}
 
 
 
string Timer::code_point_str() const {
    return fmt::format("{%s}:{%d}, {%s}()",
            code_point_->file_, code_point_->line_, code_point_->func_ );
}
 
 
/***********************************************************************************************
 * Implementation of Profiler
 */
 
 
Profiler * Profiler::instance(bool clear) {
    static Profiler * _instance = NULL;
    
    if (clear) {
        if (_instance != NULL) {
            delete _instance;
            _instance = NULL;
        }
        return _instance;
    }
 
    if (_instance == NULL) {
        MemoryAlloc::malloc_map().reserve(Profiler::malloc_map_reserve);
        _instance = new Profiler();
    }
    
    return _instance;
 } 
 
 
// static CONSTEXPR_ CodePoint main_cp = CODE_POINT("Whole Program");
const long Profiler::malloc_map_reserve = 100 * 1000;
 
Profiler::Profiler()
: actual_node(0),
  task_size_(1),
  start_time( time(NULL) ),
  json_filepath(""),
  none_timer_(CODE_POINT("NONE TIMER"), 0),
  calibration_time_(-1)
 
{
    static CONSTEXPR_ CodePoint main_cp = CODE_POINT("Whole Program");
    set_memory_monitoring(true, true);
#ifdef FLOW123D_DEBUG_PROFILER
    MemoryAlloc::malloc_map().reserve(Profiler::malloc_map_reserve);
    timers_.push_back( Timer(main_cp, 0) );
    timers_[0].start();
#endif
}
 
 
void Profiler::calibrate() {
 
    uint SIZE = 64 * 1024;
    uint HALF = SIZE / 2;
    START_TIMER("UNIT_PAYLOAD");
    Timer &timer = timers_[actual_node];
 
    uint count = 0;
    while(TimePoint() - timer.start_time < 0.1) {
        double * block = new double[SIZE];
        for(uint i=0; i<HALF; i++) {
            block[HALF+i] = block[i]*block[i] + i;
        }
        delete[] block;
        count++;
    }
 
 
 
    END_TIMER("UNIT_PAYLOAD");
 
    const double reference_count = 7730;
    // reference count is average of 5 measurements on payload free laptop:
    // DELL Latitude E5570, Intel(R) Core(TM) i7-6820HQ CPU @ 2.70GHz
    // new/delete makes about 3% of time
    calibration_time_ = 10 * timer.cumul_time * reference_count / count;
    LogOut() << "Profiler calibration count: " << count << std::endl;
    LogOut() << "Profiler calibration time: " << calibration_time_ << std::endl;
 
}
 
 
 
void Profiler::propagate_timers() {
    for (unsigned int i = 0; i < Timer::max_n_childs; i++) {
        unsigned int child_timer = timers_[0].child_timers[i];
        if ((signed int)child_timer != timer_no_child) {
            // propagate metrics from child to Whole-Program time-frame
            accept_from_child(timers_[0], timers_[child_timer]);
        }
    }
}
 
 
 
void Profiler::set_task_info(string description, int size) {
    task_description_ = description;
    task_size_ = size;
}
 
 
 
void Profiler::set_program_info(string program_name, string program_version, string branch, string revision, string build) {
    flow_name_ = program_name;
    flow_version_ = program_version;
    flow_branch_ = branch;
    flow_revision_ = revision;
    flow_build_ = build;
}
 
 
 
int  Profiler::start_timer(const CodePoint &cp) {
    unsigned int parent_node = actual_node;
    //DebugOut().fmt("Start timer: {}\n", cp.tag_);
    int child_idx = find_child(cp);
    if (child_idx < 0) {
        //DebugOut().fmt("Adding timer: {}\n", cp.tag_);
        // tag not present - create new timer
        child_idx=timers_.size();
        timers_.push_back( Timer(cp, actual_node) );
        timers_[actual_node].add_child(child_idx , timers_.back() );
    }
    actual_node=child_idx;
    
    // pause current timer
    timers_[parent_node].pause();
    
    timers_[actual_node].start();
    
    return actual_node;
}
 
 
 
int Profiler::find_child(const CodePoint &cp) {
    Timer &timer =timers_[actual_node];
    unsigned int idx = cp.hash_idx_;
    unsigned int child_idx;
    do {
        if (timer.child_timers[idx] == timer_no_child) break; // tag is not there
 
        child_idx=timer.child_timers[idx];
        ASSERT_PERMANENT_LT(child_idx, timers_.size()).error();
        if (timers_[child_idx].full_hash_ == cp.hash_) return child_idx;
        idx = ( (unsigned int)(idx)==(Timer::max_n_childs - 1) ? 0 : idx+1 );
    } while ( (unsigned int)(idx) != cp.hash_idx_ ); // passed through whole array
    return -1;
}
 
 
 
void Profiler::stop_timer(const CodePoint &cp) {
#ifdef FLOW123D_DEBUG
    // check that all childrens are closed
    Timer &timer=timers_[actual_node];
    for(unsigned int i=0; i < Timer::max_n_childs; i++)
        if (timer.child_timers[i] != timer_no_child)
            ASSERT_PERMANENT(! timers_[timer.child_timers[i]].running())(timers_[timer.child_timers[i]].tag())(timer.tag())
                .error("Child timer running while closing timer.");
#endif
    unsigned int child_timer = actual_node;
    if ( cp.hash_ != timers_[actual_node].full_hash_) {
        // timer to close is not actual - we search for it above actual
        for(unsigned int node=actual_node; node != 0; node=timers_[node].parent_timer) {
            if ( cp.hash_ == timers_[node].full_hash_) {
                // found above - close all nodes between
                for(; (unsigned int)(actual_node) != node; actual_node=timers_[actual_node].parent_timer) {
                    WarningOut() << "Timer to close '" << cp.tag_ << "' do not match actual timer '"
                            << timers_[actual_node].tag() << "'. Force closing actual." << std::endl;
                    timers_[actual_node].stop(true);
                }
                // close 'node' itself
                timers_[actual_node].stop(false);
                actual_node = timers_[actual_node].parent_timer;
                
                // actual_node == child_timer indicates this is root
                if (actual_node == child_timer)
                    return;
                
                // resume current timer
                timers_[actual_node].resume();
                return;
            }
        }
        // node not found - do nothing
        return;
    }
    // node to close match the actual
    timers_[actual_node].stop(false);
    actual_node = timers_[actual_node].parent_timer;
    
    // actual_node == child_timer indicates this is root
    if (actual_node == child_timer)
        return;
    
    // resume current timer
    timers_[actual_node].resume();
}
 
 
 
void Profiler::stop_timer(int timer_index) {
    // stop_timer with CodePoint type
    // timer which is still running MUST be the same as actual_node index
    // if timer is not running index will differ
    if (timers_[timer_index].running()) {
        ASSERT_PERMANENT_EQ(timer_index, (int)actual_node).error();
        stop_timer(*timers_[timer_index].code_point_);
    }
    
}
 
 
 
void Profiler::add_calls(unsigned int n_calls) {
    timers_[actual_node].call_count += n_calls-1;
}
 
 
 
void Profiler::notify_malloc(const size_t size, const long p) {
    if (!global_monitor_memory)
        return;
 
    MemoryAlloc::malloc_map()[p] = static_cast<int>(size);
    timers_[actual_node].total_allocated_ += size;
    timers_[actual_node].current_allocated_ += size;
    timers_[actual_node].alloc_called++;
        
    if (timers_[actual_node].current_allocated_ > timers_[actual_node].max_allocated_)
        timers_[actual_node].max_allocated_ = timers_[actual_node].current_allocated_;
}
 
 
 
void Profiler::notify_free(const long p) {
    if (!global_monitor_memory)
        return;
    
    int size = sizeof(p);
    if (MemoryAlloc::malloc_map()[(long)p] > 0) {
        size = MemoryAlloc::malloc_map()[(long)p];
        MemoryAlloc::malloc_map().erase((long)p);
    }
    timers_[actual_node].total_deallocated_ += size;
    timers_[actual_node].current_allocated_ -= size;
    timers_[actual_node].dealloc_called++;
}
 
 
double Profiler::get_resolution () {
    const int measurements = 100;
    double result = 0;
 
    // perform 100 measurements
    for (unsigned int i = 1; i < measurements; i++) {
        TimePoint t1 = TimePoint ();
        TimePoint t2 = TimePoint ();
 
        // double comparison should be avoided
        while ((t2 - t1) == 0) t2 = TimePoint ();
        // while ((t2.ticks - t1.ticks) == 0) t2 = TimePoint ();
 
        result += t2 - t1;
    }
 
    return (result / measurements) * 1000; // ticks to seconds to microseconds conversion
}
 
 
Timer Profiler::find_timer(string tag) {
    for(auto t : timers_) {
        if (t.tag() == tag) return t;
    }
    return none_timer_;
}
 
 
std::string common_prefix( std::string a, std::string b ) {
    if( a.size() > b.size() ) std::swap(a,b) ;
    return std::string( a.begin(), std::mismatch( a.begin(), a.end(), b.begin() ).first ) ;
}
 
 
 
template<typename ReduceFunctor>
void Profiler::add_timer_info(ReduceFunctor reduce, nlohmann::json* holder, int timer_idx, double parent_time) {
 
    // get timer and check preconditions
    Timer &timer = timers_[timer_idx];
    ASSERT_PERMANENT(timer_idx >=0)(timer_idx).error("Wrong timer index.");
    ASSERT_PERMANENT(timer.parent_timer >=0).error("Inconsistent tree.");
 
    // fix path
    string filepath = timer.code_point_->file_;
 
    // if constant FLOW123D_SOURCE_DIR is defined, we try to erase it from beginning of each CodePoint's filepath
    #ifdef FLOW123D_SOURCE_DIR
        string common_path = common_prefix (string(FLOW123D_SOURCE_DIR), filepath);
        filepath.erase (0, common_path.size());
    #endif
 
 
    // generate node representing this timer
    // add basic information
    nlohmann::json node;
    double cumul_time_sum;
    node["tag"] = timer.tag();
    node["file-path"] = filepath;
    node["file-line"] =  timer.code_point_->line_;
    node["function"] = timer.code_point_->func_;
    cumul_time_sum = reduce(timer, node);
 
 
    // statistical info
    if (timer_idx == 0) parent_time = cumul_time_sum;
    double percent = parent_time > 1.0e-10 ? cumul_time_sum / parent_time * 100.0 : 0.0;
    node["percent"] = percent;
 
    // write times children timers using secured child_timers array
    nlohmann::json children;
 
    // sort childs by its index in Profiler::timers_ in order to preserve call order
    std::vector<int> child_timers_values;
    for (unsigned int i = 0; i < Timer::max_n_childs; i++) {
            if (timer.child_timers[i] != timer_no_child)
                child_timers_values.push_back(timer.child_timers[i]);
    }
    std::sort(child_timers_values.begin(), child_timers_values.end());
 
    for(int idx : child_timers_values)
            add_timer_info(reduce, &children, idx, cumul_time_sum);
 
    // add children tag and other info if present
    if (child_timers_values.size() > 0) {
        node["children"] = children;
    }
 
    // add to the array
    holder->push_back(node);
}
 
 
template <class T>
void save_nonmpi_metric (nlohmann::json &node, T * ptr, string name) {
    node[name + "-min"] = *ptr;
    node[name + "-max"] = *ptr;
    node[name + "-sum"] = *ptr;
}
 
std::shared_ptr<std::ostream> Profiler::get_default_output_stream() {
     json_filepath = FilePath("profiler_info.log.json", FilePath::output_file);
 
    //LogOut() << "output into: " << json_filepath << std::endl;
    return make_shared<ofstream>(json_filepath.c_str());
}
 
 
#ifdef FLOW123D_HAVE_MPI
template <class T>
void save_mpi_metric (nlohmann::json &node, MPI_Comm comm, T * ptr, string name) {
    node[name + "-min"] = MPI_Functions::min(ptr, comm);
    node[name + "-max"] = MPI_Functions::max(ptr, comm);
    node[name + "-sum"] = MPI_Functions::sum(ptr, comm);
}
 
void Profiler::output(MPI_Comm comm, ostream &os) {
    int mpi_rank, mpi_size;
    //wait until profiling on all processors is finished
    MPI_Barrier(comm);
    stop_timer(0);
    propagate_timers();
    
    // stop monitoring memory
    bool temp_memory_monitoring = global_monitor_memory;
    set_memory_monitoring(false, petsc_monitor_memory);
 
    chkerr( MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank) );
    MPI_Comm_size(comm, &mpi_size);
 
    // output header
    nlohmann::json jsonRoot, jsonChildren;
 
    // recursively add all timers info
    // define lambda function which reduces timer from multiple processors
    // MPI implementation uses MPI call to reduce values
    auto reduce = [=] (Timer &timer, nlohmann::json &node) -> double {
        int call_count = timer.call_count;
        double cumul_time = timer.cumulative_time ();
        
        long memory_allocated = (long)timer.total_allocated_;
        long memory_deallocated = (long)timer.total_deallocated_;
        long memory_peak = (long)timer.max_allocated_;
        
        int alloc_called = timer.alloc_called;
        int dealloc_called = timer.dealloc_called;
        
        
        save_mpi_metric<double>(node, comm, &cumul_time, "cumul-time");
        save_mpi_metric<int>(node, comm, &call_count, "call-count");
        
        save_mpi_metric<long>(node, comm, &memory_allocated, "memory-alloc");
        save_mpi_metric<long>(node, comm, &memory_deallocated, "memory-dealloc");
        save_mpi_metric<long>(node, comm, &memory_peak, "memory-peak");
        // 
        save_mpi_metric<int>(node, comm, &alloc_called, "memory-alloc-called");
        save_mpi_metric<int>(node, comm, &dealloc_called, "memory-dealloc-called");
        
#ifdef FLOW123D_HAVE_PETSC
        long petsc_memory_difference = (long)timer.petsc_memory_difference;
        long petsc_peak_memory = (long)timer.petsc_peak_memory;
        save_mpi_metric<long>(node, comm, &petsc_memory_difference, "memory-petsc-diff");
        save_mpi_metric<long>(node, comm, &petsc_peak_memory, "memory-petsc-peak");
#endif // FLOW123D_HAVE_PETSC
        
        return MPI_Functions::sum(&cumul_time, comm);
    };
 
    add_timer_info (reduce, &jsonChildren, 0, 0.0);
    jsonRoot["children"] = jsonChildren;
    output_header(jsonRoot, mpi_size);
 
 
    // create profiler output only once (on the first processor)
    // only active communicator should be the one with mpi_rank 0
    if (mpi_rank == 0) {
        try {
            /**
             * indent size
             * results in json human readable format (indents, newlines)
             */
            const int FLOW123D_JSON_HUMAN_READABLE = 2;
            // write result to stream
            os << jsonRoot.dump(FLOW123D_JSON_HUMAN_READABLE) << endl;
 
        } catch (exception & e) {
            stringstream ss;
            ss << "nlohmann::json::dump error: " << e.what() << "\n";
            THROW( ExcMessage() << EI_Message(ss.str()) );
        }
    }
    // restore memory monitoring
    set_memory_monitoring(temp_memory_monitoring, petsc_monitor_memory);
}
 
 
void Profiler::output(MPI_Comm comm, string profiler_path /* = "" */) {
  int mpi_rank;
  chkerr(MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank));
 
    if (mpi_rank == 0) {
        if (profiler_path == "") {
          output(comm, *get_default_output_stream());
        } else {
          json_filepath = profiler_path;
          std::shared_ptr<std::ostream> os = make_shared<ofstream>(profiler_path.c_str());
          output(comm, *os);
        }
    } else {
      ostringstream os;
      output(comm, os);
    }
}
 
#endif /* FLOW123D_HAVE_MPI */
 
void Profiler::output(ostream &os) {
    // last update
    stop_timer(0);
    propagate_timers();
 
    // output header
    nlohmann::json jsonRoot, jsonChildren;
    /**
     * Constant representing number of MPI processes
     * where there is no MPI to work with (so 1 process)
     */
    const int FLOW123D_MPI_SINGLE_PROCESS = 1;
    output_header(jsonRoot, FLOW123D_MPI_SINGLE_PROCESS);
 
 
    // recursively add all timers info
    // define lambda function which reduces timer from multiple processors
    // non-MPI implementation is just dummy repetition of initial value
    auto reduce = [=] (Timer &timer, nlohmann::json &node) -> double {
        int call_count = timer.call_count;
        double cumul_time = timer.cumulative_time ();
        
        long memory_allocated = (long)timer.total_allocated_;
        long memory_deallocated = (long)timer.total_deallocated_;
        long memory_peak = (long)timer.max_allocated_;
        
        int alloc_called = timer.alloc_called;
        int dealloc_called = timer.dealloc_called;
        
        save_nonmpi_metric<double>(node, &cumul_time, "cumul-time");
        save_nonmpi_metric<int>(node, &call_count, "call-count");
        
        save_nonmpi_metric<long>(node, &memory_allocated, "memory-alloc");
        save_nonmpi_metric<long>(node, &memory_deallocated, "memory-dealloc");
        save_nonmpi_metric<long>(node, &memory_peak, "memory-peak");
        
        save_nonmpi_metric<int>(node, &alloc_called, "memory-alloc-called");
        save_nonmpi_metric<int>(node, &dealloc_called, "memory-dealloc-called");
        
#ifdef FLOW123D_HAVE_PETSC
        long petsc_memory_difference = (long)timer.petsc_memory_difference;
        long petsc_peak_memory = (long)timer.petsc_peak_memory;
        save_nonmpi_metric<long>(node, &petsc_memory_difference, "memory-petsc-diff");
        save_nonmpi_metric<long>(node, &petsc_peak_memory, "memory-petsc-peak");
#endif // FLOW123D_HAVE_PETSC
        
        return cumul_time;
    };
 
    add_timer_info(reduce, &jsonChildren, 0, 0.0);
    jsonRoot["children"] = jsonChildren;
 
    try {
        /**
         * indent size
         * results in json human readable format (indents, newlines)
         */
        const int FLOW123D_JSON_HUMAN_READABLE = 2;
        // write result to stream
        os << jsonRoot.dump(FLOW123D_JSON_HUMAN_READABLE) << endl;
 
    } catch (exception & e) {
        stringstream ss;
        ss << "nlohmann::json::dump error: " << e.what() << "\n";
        THROW( ExcMessage() << EI_Message(ss.str()) );
    }
}
 
 
void Profiler::output(string profiler_path /* = "" */) {
    if(profiler_path == "") {
        output(*get_default_output_stream());
    } else {
        json_filepath = profiler_path;
        std::shared_ptr<std::ostream> os = make_shared<ofstream>(profiler_path.c_str());
        output(*os);
    }
}
 
void Profiler::output_header (nlohmann::json &root, int mpi_size) {
    time_t end_time = time(NULL);
 
    const char format[] = "%x %X";
    char start_time_string[BUFSIZ] = {0};
    strftime(start_time_string, sizeof (start_time_string) - 1, format, localtime(&start_time));
 
    char end_time_string[BUFSIZ] = {0};
    strftime(end_time_string, sizeof (end_time_string) - 1, format, localtime(&end_time));
 
    if (timers_[0].cumul_time > 60) {
        calibrate();
    }
    // generate current run details
    root["program-name"] =        flow_name_;
    root["program-version"] =     flow_version_;
    root["program-branch"] =      flow_branch_;
    root["program-revision"] =    flow_revision_;
    root["program-build"] =       flow_build_;
    root["timer-resolution"] =    Profiler::get_resolution();
    root["timer-calibration"] =    calibration_time_;
 
    // print some information about the task at the beginning
    root["task-description"] =    task_description_;
    root["task-size"] =           task_size_;
 
    //print some information about the task at the beginning
    root["run-process-count"] =   mpi_size;
    root["run-started-at"] =      start_time_string;
    root["run-finished-at"] =     end_time_string;
}
 
#ifdef FLOW123D_HAVE_PYTHON
void Profiler::transform_profiler_data (const string &output_file_suffix, const string &formatter) {
    
    if (json_filepath == "") return;
 
    // error under CYGWIN environment : more details in this repo 
    // https://github.com/x3mSpeedy/cygwin-python-issue
    // 
    // For now we only support profiler report conversion in UNIX environment
    // Windows users will have to use a python script located in bin folder
    // 
 
    #ifndef FLOW123D_HAVE_CYGWIN
    // grab module and function by importing module profiler_formatter_module.py
    PyObject * python_module = PythonLoader::load_module_by_name ("profiler.profiler_formatter_module");
    //
    // def convert (json_location, output_file, formatter):
    //
    PyObject * convert_method  = PythonLoader::get_callable (python_module, "convert" );
 
    int argument_index = 0;
    PyObject * arguments = PyTuple_New (3);
 
    // set json path location as first argument
    PyObject * tmp = PyUnicode_FromString (json_filepath.c_str());
    PyTuple_SetItem (arguments, argument_index++, tmp);
 
    // set output path location as second argument
    tmp = PyUnicode_FromString ((json_filepath + output_file_suffix).c_str());
    PyTuple_SetItem (arguments, argument_index++, tmp);
 
    // set Formatter class as third value
    tmp = PyUnicode_FromString (formatter.c_str());
    PyTuple_SetItem (arguments, argument_index++, tmp);
 
    // execute method with arguments
    PyObject_CallObject (convert_method, arguments);
 
    PythonLoader::check_error();
 
    #else
 
    // print information about windows-cygwin issue and offer manual solution
    MessageOut() << "# Note: converting json profiler reports is not"
                 << " supported under Windows or Cygwin environment for now.\n"
                 << "# You can use python script located in bin/python folder"
                 << " in order to convert json report to txt or csv format.\n"
                 << "python profiler_formatter_script.py --input \"" << json_filepath
                 << "\" --output \"profiler.txt\"" << std::endl;
    #endif // FLOW123D_HAVE_CYGWIN
}
#else
void Profiler::transform_profiler_data (const string &, const string &) {
}
 
#endif // FLOW123D_HAVE_PYTHON
 
 
void Profiler::uninitialize() {
    if (Profiler::instance()) {
        ASSERT_PERMANENT(Profiler::instance()->actual_node==0)(Profiler::instance()->timers_[Profiler::instance()->actual_node].tag())
                .error("Forbidden to uninitialize the Profiler when actual timer is not zero.");
        Profiler::instance()->stop_timer(0);
        set_memory_monitoring(false, false);
        Profiler::instance(true);
    }
}
bool Profiler::global_monitor_memory = false;
bool Profiler::petsc_monitor_memory = true;
void Profiler::set_memory_monitoring(const bool global_monitor, const bool petsc_monitor) {
    global_monitor_memory = global_monitor;
    petsc_monitor_memory = petsc_monitor;
}
 
bool Profiler::get_global_memory_monitoring() {
    return global_monitor_memory;
}
 
bool Profiler::get_petsc_memory_monitoring() {
    return petsc_monitor_memory;
}
 
unordered_map_with_alloc & MemoryAlloc::malloc_map() {
    static unordered_map_with_alloc static_malloc_map;
    return static_malloc_map;
}
 
void * Profiler::operator new (size_t size) {
    return malloc (size);
}
 
void Profiler::operator delete (void* p) {
    free(p);
}
 
void *operator new (std::size_t size) OPERATOR_NEW_THROW_EXCEPTION {
    void * p = malloc(size);
    Profiler::instance()->notify_malloc(size, (long)p);
    return p;
}
 
void *operator new[] (std::size_t size) OPERATOR_NEW_THROW_EXCEPTION {
    void * p = malloc(size);
    Profiler::instance()->notify_malloc(size, (long)p);
    return p;
}
 
void *operator new[] (std::size_t size, const std::nothrow_t&) throw() {
    void * p = malloc(size);
    Profiler::instance()->notify_malloc(size, (long)p);
    return p;
}
 
void operator delete( void *p) throw() {
    Profiler::instance()->notify_free((long)p);
    free(p);
}
 
void operator delete( void *p, std::size_t) throw() {
    Profiler::instance()->notify_free((long)p);
    free(p);
}
 
void operator delete[]( void *p) throw() {
    Profiler::instance()->notify_free((long)p);
    free(p);
}
 
void operator delete[]( void *p, std::size_t) throw() {
    Profiler::instance()->notify_free((long)p);
    free(p);
}
 
#else // def FLOW123D_DEBUG_PROFILER
 
Profiler * Profiler::instance(bool) { 
    static Profiler * _instance = new Profiler();
    return _instance;
} 
 
 
// void Profiler::initialize() {
//     if (_instance == NULL) {
//         _instance = new Profiler();
//         set_memory_monitoring(true, true);
//     }
// }
 
void Profiler::uninitialize() {}
 
 
#endif // def FLOW123D_DEBUG_PROFILER