From 2e6ce620fa140a90f6f5b6df1345f3e5d70005bb Mon Sep 17 00:00:00 2001
From: Rene Milk <koshi@springlobby.info>
Date: Wed, 12 Aug 2009 22:20:24 +0200
Subject: [PATCH] better profiling output
---
stuff/profiler.hh | 110 ++++++++++++++++++++++++++++++++++++----------
1 file changed, 86 insertions(+), 24 deletions(-)
diff --git a/stuff/profiler.hh b/stuff/profiler.hh
index 9ba391163..0535d28fe 100644
--- a/stuff/profiler.hh
+++ b/stuff/profiler.hh
@@ -7,6 +7,8 @@
#include <vector>
#include <ctime>
+#include "misc.hh"
+
struct TimingData
{
clock_t start;
@@ -76,8 +78,8 @@ public:
* \param comm used to gather and average the runtime data over all processes
* \tparam CollectiveCommunication should be Dune::CollectiveCommunication< MPI_Comm / double >
**/
- template <class CollectiveCommunication>
- long Output(CollectiveCommunication& comm, const int refineLevel, const long numDofs);
+ template <class CollectiveCommunication, class InfoContainer>
+ long Output(CollectiveCommunication& comm, InfoContainer& run_infos);
/** call this with correct numRuns <b> before </b> starting any profiling
* if you're planning on doing more than one iteration of your code
@@ -89,10 +91,7 @@ public:
m_timings = MapVector(numRuns, DataMap());
m_total_runs = numRuns;
m_cur_run_num = 0;
- // DataMap td;
- // m_timings.push_back( td );
- m_l2_error = 0;
- init_time_ = clock();
+ init_time_ = clock();
}
//! simple counter, usable to count how often a single piece of code is called
@@ -103,19 +102,15 @@ public:
//! call this after one iteration of your code has finished. increments current run number and puts new timing data
//! into the vector
- void NextRun(const double error)
+ void NextRun()
{
- m_cur_run_num++; // += m_cur_run_num < m_total_runs - 1 ? 1 : 0 ;
- // DataMap td;
- // m_timings.push_back( td );
- m_l2_error += error;
+ m_cur_run_num++;
}
protected:
MapVector m_timings;
int m_cur_run_num;
int m_total_runs;
- double m_l2_error;
// debug counter, only outputted in debug mode
std::map<int, int> m_count;
clock_t init_time_;
@@ -162,8 +157,7 @@ long Profiler::OutputAveraged(CollectiveCommunication& comm, const int refineLev
csv << "Speedup (total); Speedup (ohne Solver)" << std::endl;
// outputs column values
- csv << refineLevel << "," << comm.size() << "," << numDofs << ","
- << comm.sum(m_l2_error) / double(m_total_runs * numProce) << ",";
+ csv << refineLevel << "," << comm.size() << "," << numDofs << "," << 0 << ","; //! FIXME
for (AvgMap::const_iterator it = averages.begin(); it != averages.end(); ++it) {
long clock_count = it->second;
clock_count = long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
@@ -178,13 +172,13 @@ long Profiler::OutputAveraged(CollectiveCommunication& comm, const int refineLev
return long((clock() - init_time_) / double(CLOCKS_PER_SEC * 0.001));
}
-template <class CollectiveCommunication>
-long Profiler::Output(CollectiveCommunication& comm, const int refineLevel, const long numDofs)
+template <class CollectiveCommunication, class InfoContainer>
+long Profiler::Output(CollectiveCommunication& comm, InfoContainer& run_infos)
{
const int numProce = comm.size();
std::ostringstream filename;
- filename << "p" << numProce << "_refinelvl_" << refineLevel << ".csv";
+ filename << "prof_p" << numProce << ".csv";
filename.flush();
if (comm.rank() == 0)
@@ -206,24 +200,92 @@ long Profiler::Output(CollectiveCommunication& comm, const int refineLevel, cons
for (DataMap::const_iterator it = m_timings[0].begin(); it != m_timings[0].end(); ++it) {
csv << it->first << ",";
}
- csv << "Speedup (total); Speedup (ohne Solver)" << std::endl;
+ csv << "Relative total time" << std::endl;
// outputs column values
- csv << refineLevel << "," << comm.size() << "," << numDofs << ","
- << comm.sum(m_l2_error) / double(m_total_runs * numProce) << ",";
MapVector::const_iterator ti_it = m_timings.begin();
+ int idx = 0;
for (; ti_it != m_timings.end(); ++ti_it) {
+ RunInfo info = run_infos[idx];
+ csv << info.refine_level << "," << comm.size() << "," << info.codim0 << "," << info.L2Errors[0] << ",";
const DataMap& data_map = *ti_it;
for (DataMap::const_iterator it = data_map.begin(); it != data_map.end(); ++it) {
TimingData data = it->second;
long clock_count = data.end - data.start;
- clock_count = long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
- csv << clock_count / double(m_total_runs) << ",";
+ clock_count = long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
+ csv << clock_count << ",";
+ }
+ csv << "=1/I$2*I" << Stuff::toString(idx + 2) << std::endl;
+
+ idx++;
+ }
+
+ csv << "\n Linear estimate \n";
+
+ {
+ RunInfo info = run_infos[0];
+ csv << info.refine_level << "," << comm.size() << "," << info.codim0 << "," << info.L2Errors[0] << ",";
+
+ const DataMap& data_map = m_timings[0];
+ for (DataMap::const_iterator it = data_map.begin(); it != data_map.end(); ++it) {
+ TimingData data = it->second;
+ long clock_count = data.end - data.start;
+ clock_count = long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
+ csv << clock_count << ",";
+ }
+ csv << 1 << std::endl;
+
+ double factor = 1;
+ for (int i = 1; i < run_infos.size(); ++i) {
+ RunInfo info = run_infos[i];
+ RunInfo info_last = run_infos[i - 1];
+ factor *= info.codim0 / double(info_last.codim0);
+ csv << info.refine_level << "," << comm.size() << "," << info.codim0 << "," << info.L2Errors[0] / factor << ",";
+
+ const DataMap& data_map = m_timings[0];
+ for (DataMap::const_iterator it = data_map.begin(); it != data_map.end(); ++it) {
+ TimingData data = it->second;
+ long clock_count = data.end - data.start;
+ clock_count = factor * long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
+ csv << clock_count << ",";
+ }
+ csv << factor << std::endl;
+ }
+ }
+
+ csv << "\n Quadratic estimate \n";
+
+ {
+ RunInfo info = run_infos[0];
+ csv << info.refine_level << "," << comm.size() << "," << info.codim0 << "," << info.L2Errors[0] << ",";
+
+ const DataMap& data_map = m_timings[0];
+ for (DataMap::const_iterator it = data_map.begin(); it != data_map.end(); ++it) {
+ TimingData data = it->second;
+ long clock_count = data.end - data.start;
+ clock_count = long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
+ csv << clock_count << ",";
+ }
+ csv << 1 << std::endl;
+
+ double factor = 1;
+ for (int i = 1; i < run_infos.size(); ++i) {
+ RunInfo info = run_infos[i];
+ RunInfo info_last = run_infos[i - 1];
+ factor *= std::pow(info.codim0 / double(info_last.codim0), 2);
+ csv << info.refine_level << "," << comm.size() << "," << info.codim0 << "," << info.L2Errors[0] / factor << ",";
+
+ const DataMap& data_map = m_timings[0];
+ for (DataMap::const_iterator it = data_map.begin(); it != data_map.end(); ++it) {
+ TimingData data = it->second;
+ long clock_count = data.end - data.start;
+ clock_count = factor * long(comm.sum(clock_count) / double(CLOCKS_PER_SEC * 0.001 * numProce));
+ csv << clock_count << ",";
+ }
+ csv << factor << std::endl;
}
- csv << "=I$2/I2,"
- << "=SUM(E$2:G$2)/SUM(E2:G2)" << std::endl;
}
csv.close();
--
GitLab