AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix sampling counter time scales (#33)

Browse Source 
* Fix sampling counter time scales

- All perfetto trace events have "begin_ns" and "end_ns" debug fields
- data for thread start and end timestamp in pthread_create_gotcha
- discard samples outside of thread start and end timestamps
- rename "CPU User CPU Time" perfetto counter to "CPU User Time"
- rename "CPU Kernel CPU Time" perfetto counter to "CPU Kernel Time"
- ensure CPU system samples in perfetto are set to zero at end
- backtrace uses comp::wall_clock record() for timestamps (consistency)
- "Peak Memory Usage [Thread X] (S)" renamed to "Thread Peak Memory Usage [X] (S)"
- "Context Switches [Thread X] (S)" renamed to "Thread Context Switches Usage [X] (S)"
- "Page Faults [Thread X] (S)" renamed to "Thread Page Faults Usage [X] (S)"
- "<PAPI_DESC> [Thread X] (S)" renamed to "Thread <PAPI_DESC> [X] (S)"
- samples

* Fix includes

[ROCm/rocprofiler-systems commit: 90ab7a89fc]
This commit is contained in:
Jonathan R. Madsen
2022-06-10 08:35:39 -05:00
committed by GitHub
parent 2eb4b4981c
commit ac25b485e7
9 changed files with 123 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-systems/source/lib/omnitrace/library.cpp
+8 -2
View File
@@ -26,6 +26,7 @@
#include "library/components/functors.hpp"
#include "library/components/fwd.hpp"
#include "library/components/mpi_gotcha.hpp"
#include "library/components/pthread_create_gotcha.hpp"
#include "library/components/pthread_gotcha.hpp"
#include "library/components/pthread_mutex_gotcha.hpp"
#include "library/config.hpp"
@@ -752,13 +753,13 @@ omnitrace_init_tooling_hidden()
using CategoryT = std::decay_t<decltype(_category)>;
uint64_t _ts = comp::wall_clock::record();
TRACE_EVENT_BEGIN(trait::name<CategoryT>::value, perfetto::StaticString(name),
_ts);
_ts, "begin_ns", _ts);
};
auto _pop_perfetto = [](auto _category, const char*) {
using CategoryT = std::decay_t<decltype(_category)>;
uint64_t _ts = comp::wall_clock::record();
TRACE_EVENT_END(trait::name<CategoryT>::value, _ts);
TRACE_EVENT_END(trait::name<CategoryT>::value, _ts, "end_ns", _ts);
};
if(get_use_perfetto() && get_use_timemory())
@@ -855,6 +856,8 @@ omnitrace_init_tooling_hidden()
if(dmp::rank() == 0 && get_verbose() >= 0) fprintf(stderr, "\n");
pthread_create_gotcha::get_execution_time()->first = comp::wall_clock::record();
return true;
}
@@ -949,6 +952,8 @@ omnitrace_init_hidden(const char* _mode, bool _is_binary_rewrite, const char* _a
{
get_gotcha_bundle()->start();
}
pthread_create_gotcha::get_execution_time()->first = comp::wall_clock::record();
}
//======================================================================================//
@@ -970,6 +975,7 @@ omnitrace_finalize_hidden(void)
}
OMNITRACE_VERBOSE_F(0, "finalizing...\n");
pthread_create_gotcha::get_execution_time()->second = comp::wall_clock::record();
// some functions called during finalization may alter the push/pop count so we need
// to save them here