- corrections in the calculations for latency and throughput points in `validate-causal-json.py`
- `omnitrace-causal` LD_PRELOAD libpthread
  - ensures omnitrace is always wrapping libpthread.so pthread symbols
- minimal experiment delay
  - always sleep 10 milliseconds before starting experiments
  - ensures ~10 samples are taken to determine the sampling rate
- fixes issue with deadlocks on condition variables
- overhaul of `causal::component::blocking_gotcha` and `causal::component::unblocking_gotcha` components
  - these components enforce the processing/crediting of delays before/after a thread is suspended
  - these components wrap functions `pthread_cond_wait`, `pthread_cond_signal`, `pthread_mutex_lock`, etc.
- Fully implemented correct handling of processing/crediting delays based on return values and arguments
  - E.g. skip crediting delay if `pthread_mutex_trylock` fail acquiring lock
  - E.g. `kill`, `sigwait`, etc. check to make sure they are only applied if the PID matches its PID
 
## Condition Variable Deadlock Fix

In parallel applications using condition variables, it was found that the causal profiling was virtually guaranteed to deadlock. Although it was difficult to prove, evidence suggested that this was due to the work that was being done while taking a sample was causing notification to the condition variable to be lost. This was alleviated by the following updates:
 
- Separate out the part of `causal::backtrace::sample(int)` which calculates the sampling rate into small `sample_rate` component
  - This component is essentially "always on"  during sampling
  - Added bundle of components invoked by `causal_sampler_t` during sampling
- Added two function calls to support disabling and re-enabling calls to `causal::backtrace::sample(int)` on a per-thread basis 
  - `causal::sampling::block_backtrace_samples()`
  - `causal::sampling::unblock_backtrace_samples()`
  - These two function now surround the wrappee functions of `blocking_gotcha` and `unblocking_gotcha`

**This solution was experimentally validated with a Geant4 application which uses a tasking model which makes _numerous_ calls to wait on a condition variables** (it was this application which exposed the bug)

* Fix validate-causal-json.py

- corrections in the calculations for latency and throughput points

* Update timemory submodule

- support for thread-local trait::runtime_enabled

* omnitrace-causal: LD_PRELOAD pthread library

- ensures omnitrace is always wrapping libpthread.so pthread symbols

* initial experiment delay

- always sleep 10 milliseconds before starting experiments
- ensures ~10 samples are taken to determine the sampling rate

* sample_rate component + block_backtrace_samples

- separate out the part of backtrace::sample which calculates the sampling rate into small sample_rate component
- add sample_rate component to causal_bundle_t used by causal_sampler_t
- causal::sampling::block_backtrace_samples() disables backtrace samples from being taken on a thread
- causal::sampling::unblock_backtrace_samples() enables backtrace samples from being taken on a thread
- above two function surround calls to function wrapped by blocking_gotcha and unblocking_gotcha
  - the work happening in backtrace::sample when within these calls
    produced deadlocks for condition variables (notifications to
    condition variables were lost)

* blocking/unblocking gotcha updates

- overhaul of blocking_gotcha and unblocking_gotcha
  - added fast_gotcha trait: replace function calls instead of wrapping
- when wrappees are called, backtrace samples are suppressed (thread-local)
- properly handle kill, sigwait, sigwaitinfo, sigtimedwait
- properly handle all instances of applying postblock based on return value

* Fix calculation of OMNITRACE_MAX_THREADS

* removed unnecessary checks in causal::delay

* Updated timemory with internal compiler error fix

[ROCm/rocprofiler-systems commit: 7c73d98125]
Этот коммит содержится в:
Jonathan R. Madsen
2023-02-09 09:47:48 -06:00
коммит произвёл GitHub
родитель ecc794276c
Коммит 5eb895fc7d
15 изменённых файлов: 429 добавлений и 216 удалений
+26 -2
Просмотреть файл
@@ -233,10 +233,19 @@ endif()
include(ProcessorCount)
processorcount(OMNITRACE_PROCESSOR_COUNT)
math(EXPR OMNITRACE_THREAD_COUNT "16 * ${OMNITRACE_PROCESSOR_COUNT}")
if(OMNITRACE_THREAD_COUNT LESS 128)
if(OMNITRACE_PROCESSOR_COUNT LESS 8)
set(OMNITRACE_THREAD_COUNT 128)
else()
math(EXPR OMNITRACE_THREAD_COUNT "16 * ${OMNITRACE_PROCESSOR_COUNT}")
compute_pow2_ceil(OMNITRACE_THREAD_COUNT "16 * ${OMNITRACE_PROCESSOR_COUNT}")
# set the default to 2048 if it could not be calculated
if(OMNITRACE_THREAD_COUNT LESS 2)
set(OMNITRACE_THREAD_COUNT 2048)
endif()
endif()
set(OMNITRACE_MAX_THREADS
"${OMNITRACE_THREAD_COUNT}"
CACHE
@@ -247,6 +256,21 @@ omnitrace_add_feature(
OMNITRACE_MAX_THREADS
"Maximum number of total threads supported in the host application (default: max of 128 or 16 * nproc)"
)
compute_pow2_ceil(_MAX_THREADS "${OMNITRACE_MAX_THREADS}")
if(_MAX_THREADS GREATER 0 AND NOT OMNITRACE_MAX_THREADS EQUAL _MAX_THREADS)
omnitrace_message(
FATAL_ERROR
"Error! OMNITRACE_MAX_THREADS must be a power of 2. Recommendation: ${_MAX_THREADS}"
)
elseif(NOT OMNITRACE_MAX_THREADS EQUAL _MAX_THREADS)
omnitrace_message(
AUTHOR_WARNING
"OMNITRACE_MAX_THREADS (=${OMNITRACE_MAX_THREADS}) must be a power of 2. We were unable to verify it so we are emitting this warning instead. Estimate resulted in: ${_MAX_THREADS}"
)
endif()
set(OMNITRACE_MAX_UNWIND_DEPTH
"64"
CACHE
+30
Просмотреть файл
@@ -892,4 +892,34 @@ function(OMNITRACE_INSTALL_TPL _TPL_TARGET _NEW_NAME _BUILD_TREE_DIR _COMPONENT)
endfunction()
function(COMPUTE_POW2_CEIL _OUTPUT _VALUE)
find_package(Python3 COMPONENTS Interpreter)
if(Python3_FOUND)
execute_process(
COMMAND
${Python3_EXECUTABLE} -c
"VALUE = ${_VALUE}; ispow2 = lambda x: x if (x and (not(x & (x - 1)))) else None; v = list(filter(ispow2, [x for x in range(VALUE, VALUE**2)])); print(v[0])"
RESULT_VARIABLE _POW2_RET
OUTPUT_VARIABLE _POW2_OUT
ERROR_VARIABLE _POW2_ERR
OUTPUT_STRIP_TRAILING_WHITESPACE)
if(_POW2_RET EQUAL 0)
set(${_OUTPUT}
${_POW2_OUT}
PARENT_SCOPE)
else()
set(${_OUTPUT}
"-1"
PARENT_SCOPE)
endif()
else()
set(${_OUTPUT}
"-1"
PARENT_SCOPE)
endif()
endfunction()
cmake_policy(POP)
Submodule projects/rocprofiler-systems/external/timemory updated: e4b06f9479...d290a33575
+4 -1
Просмотреть файл
@@ -43,6 +43,7 @@
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <gnu/lib-names.h>
#include <iostream>
#include <stdexcept>
#include <string>
@@ -340,7 +341,9 @@ prepare_environment_for_run(std::vector<char*>& _env)
if(launcher.empty())
{
update_env(_env, "LD_PRELOAD",
get_realpath(get_internal_libpath("libomnitrace-dl.so")), true);
join(":", LIBPTHREAD_SO,
get_realpath(get_internal_libpath("libomnitrace-dl.so"))),
true);
}
}
+27 -50
Просмотреть файл
@@ -70,38 +70,6 @@ get_delay_statistics()
static auto& _v = thread_data_t::instance(construct_on_init{});
return _v;
}
auto&
get_in_use()
{
using thread_data_t = thread_data<identity<bool>, category::sampling>;
static_assert(
use_placement_new_when_generating_unique_ptr<thread_data_t>::value,
"sampling is_use thread data should use placement new to allocate unique_ptr");
static auto& _v = thread_data_t::instance(construct_on_init{});
return _v;
}
struct scoped_in_use
{
scoped_in_use(int64_t _tid = utility::get_thread_index())
: value{ get_in_use()->at(_tid) }
{
value = true;
}
~scoped_in_use() { value = false; }
bool& value;
};
auto
is_in_use(int64_t _tid = threading::get_id())
{
return get_in_use()->at(_tid);
}
} // namespace
void
@@ -111,13 +79,26 @@ backtrace::start()
// before it is called in sampler or else a deadlock will occur when
// the sample interrupts a malloc call
(void) get_delay_statistics();
(void) get_in_use();
}
void
backtrace::stop()
{}
void
sample_rate::sample(int _sig)
{
if(_sig != get_realtime_signal()) return;
// update the last sample for backtrace signal(s) even when in use
static thread_local int64_t _last_sample = 0;
auto _this_sample = tracing::now();
auto& _period_stat = get_delay_statistics()->at(threading::get_id());
if(_last_sample > 0) _period_stat += (_this_sample - _last_sample);
_last_sample = _this_sample;
}
void
backtrace::sample(int _sig)
{
@@ -125,16 +106,17 @@ backtrace::sample(int _sig)
constexpr int64_t ignore_depth = ::omnitrace::causal::unwind_offset;
// update the last sample for backtrace signal(s) even when in use
static thread_local int64_t _last_sample = 0;
static thread_local size_t _protect_flag = 0;
if(is_in_use() ||
// sampling_guard _guard{};
if((_protect_flag & 1) == 1 ||
OMNITRACE_UNLIKELY(!trait::runtime_enabled<causal::component::backtrace>::get()))
{
if(_sig == get_realtime_signal()) _last_sample = tracing::now();
return;
}
scoped_in_use _in_use{};
++_protect_flag;
m_index = causal::experiment::get_index();
m_stack = get_unw_signal_frame_stack_raw<depth, ignore_depth>();
@@ -150,10 +132,8 @@ backtrace::sample(int _sig)
}
else if(_sig == get_realtime_signal())
{
auto _this_sample = tracing::now();
auto& _period_stat = get_delay_statistics()->at(threading::get_id());
if(_last_sample > 0) _period_stat += (_this_sample - _last_sample);
_last_sample = _this_sample;
static thread_local auto _tid = threading::get_id();
auto& _period_stat = get_delay_statistics()->at(_tid);
if(causal::experiment::is_active() && causal::experiment::is_selected(m_stack))
{
@@ -173,6 +153,8 @@ backtrace::sample(int _sig)
{
OMNITRACE_THROW("unhandled signal %i\n", _sig);
}
++_protect_flag;
}
template <typename Tp>
@@ -194,13 +176,10 @@ backtrace::get_period(uint64_t _units)
tim::statistics<int64_t>
backtrace::get_period_stats()
{
scoped_in_use _in_use{};
auto _data = tim::statistics<int64_t>{};
auto _data = tim::statistics<int64_t>{};
if(!get_delay_statistics()) return _data;
for(size_t i = 0; i < get_delay_statistics()->size(); ++i)
for(auto itr : *get_delay_statistics())
{
scoped_in_use _thr_in_use{ static_cast<int64_t>(i) };
const auto& itr = get_delay_statistics()->at(i);
if(itr.get_count() > 1) _data += itr;
}
return _data;
@@ -209,11 +188,9 @@ backtrace::get_period_stats()
void
backtrace::reset_period_stats()
{
scoped_in_use _in_use{};
for(size_t i = 0; i < get_delay_statistics()->size(); ++i)
for(auto& itr : *get_delay_statistics())
{
scoped_in_use _thr_in_use{ static_cast<int64_t>(i) };
get_delay_statistics()->at(i).reset();
itr.reset();
}
}
} // namespace component
+8
Просмотреть файл
@@ -45,6 +45,14 @@ namespace causal
{
namespace component
{
struct sample_rate
: tim::component::empty_base
, tim::concepts::component
{
using value_type = void;
static void sample(int = -1);
};
struct backtrace
: tim::component::empty_base
, tim::concepts::component
+160 -57
Просмотреть файл
@@ -24,18 +24,41 @@
#include "core/config.hpp"
#include "core/debug.hpp"
#include "core/state.hpp"
#include "library/causal/components/causal_gotcha.hpp"
#include "library/causal/delay.hpp"
#include "library/causal/experiment.hpp"
#include "library/causal/sampling.hpp"
#include "library/runtime.hpp"
#include <timemory/components/macros.hpp>
#include <timemory/hash/types.hpp>
#include <timemory/utility/types.hpp>
#include <atomic>
#include <csignal>
#include <cstdint>
#include <pthread.h>
#include <stdexcept>
#pragma weak pthread_join
#pragma weak pthread_mutex_lock
#pragma weak pthread_spin_lock
#pragma weak pthread_cond_wait
#pragma weak pthread_rwlock_rdlock
#pragma weak pthread_rwlock_wrlock
#pragma weak pthread_tryjoin_np
#pragma weak pthread_timedjoin_np
#pragma weak pthread_cond_timedwait
#pragma weak pthread_rwlock_timedrdlock
#pragma weak pthread_rwlock_timedwrlock
#pragma weak pthread_mutex_trylock
#pragma weak pthread_spin_trylock
#pragma weak pthread_rwlock_trywrlock
#pragma weak sigwait
#pragma weak sigwaitinfo
#pragma weak sigtimedwait
#pragma weak sigsuspend
namespace omnitrace
{
namespace causal
@@ -67,48 +90,48 @@ blocking_gotcha::configure()
blocking_gotcha_t::get_initializer() = []() {
if(!config::get_use_causal()) return;
blocking_gotcha_t::configure(
comp::gotcha_config<0, int, pthread_mutex_t*>{ "pthread_mutex_lock" });
// postblock(true)
// - pthread_join
// - pthread_mutex_lock
// - pthread_cond_wait
// - pthread_barrier_wait
// - pthread_rwlock_rdlock
// - pthread_rwlock_wrlock
blocking_gotcha_t::configure(
comp::gotcha_config<1, int, pthread_mutex_t*>{ "pthread_mutex_trylock" });
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 0, pthread_join);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 1, pthread_mutex_lock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 2, pthread_spin_lock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 3, pthread_cond_wait);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 4, pthread_rwlock_rdlock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 5, pthread_rwlock_wrlock);
blocking_gotcha_t::configure(
comp::gotcha_config<2, int, pthread_rwlock_t*>{ "pthread_rwlock_wrlock" });
// postblock(result == 0)
// - pthread_tryjoin_np
// - pthread_timedjoin_np
// - pthread_cond_timedwait
// - pthread_rwlock_timedrdlock
// - pthread_rwlock_timedwrlock
blocking_gotcha_t::configure(
comp::gotcha_config<3, int, pthread_rwlock_t*>{ "pthread_rwlock_trywrlock" });
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 6, pthread_tryjoin_np);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 7, pthread_timedjoin_np);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 8, pthread_cond_timedwait);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 9, pthread_rwlock_timedrdlock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 10, pthread_rwlock_timedwrlock);
blocking_gotcha_t::configure(
comp::gotcha_config<4, int, pthread_spinlock_t*>{ "pthread_spin_lock" });
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 11, pthread_mutex_trylock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 12, pthread_spin_trylock);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 13, pthread_rwlock_trywrlock);
blocking_gotcha_t::configure(
comp::gotcha_config<5, int, pthread_spinlock_t*>{ "pthread_spin_trylock" });
// postblock(...)
// - sigwait
// - sigwaitinfo
// - sigtimedwait
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 14, sigwait);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 15, sigwaitinfo);
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 16, sigtimedwait);
blocking_gotcha_t::configure(
comp::gotcha_config<6, int, pthread_t, void**>{ "pthread_join" });
blocking_gotcha_t::configure(
comp::gotcha_config<7, int, pthread_cond_t*, pthread_mutex_t*>{
"pthread_cond_wait" });
blocking_gotcha_t::configure(
comp::gotcha_config<8, int, pthread_cond_t*, pthread_mutex_t*,
const struct timespec*>{ "pthread_cond_timedwait" });
blocking_gotcha_t::configure(
comp::gotcha_config<9, int, const sigset_t*, int*>{ "sigwait" });
blocking_gotcha_t::configure(
comp::gotcha_config<10, int, const sigset_t*, int*, siginfo_t*>{
"sigwaitinfo" });
blocking_gotcha_t::configure(
comp::gotcha_config<11, int, const sigset_t*, int*, siginfo_t*,
const struct timespec*>{ "sigtimedwait" });
blocking_gotcha_t::configure(
comp::gotcha_config<12, int, const sigset_t*>{ "sigsuspend" });
// other
TIMEMORY_C_GOTCHA(blocking_gotcha_t, 17, sigsuspend);
};
}
@@ -118,34 +141,114 @@ blocking_gotcha::shutdown()
blocking_gotcha_t::disable();
}
void
blocking_gotcha::start()
template <typename Ret, typename... Args>
Ret
blocking_gotcha::operator()(const comp::gotcha_data& _data, Ret (*_func)(Args...),
Args... _args) const noexcept
{
if(causal::experiment::is_active() &&
get_thread_state() == ::omnitrace::ThreadState::Enabled && delay_value == 0)
delay_value = causal::delay::get_global().load();
int64_t _delay_value = causal::delay::get_global().load(std::memory_order_relaxed);
causal::sampling::block_backtrace_samples();
auto _ret = (*_func)(_args...);
causal::sampling::unblock_backtrace_samples();
if(get_thread_state() < ::omnitrace::ThreadState::Internal)
{
if(_data.index <= 5)
causal::delay::postblock(_delay_value);
else if(_ret == 0 && _data.index >= 6 && _data.index <= 13)
causal::delay::postblock(_delay_value);
else
OMNITRACE_FAIL_F("Error! unexpected index %zu ('%s')\n", _data.index,
_data.tool_id.c_str());
}
return _ret;
}
void
blocking_gotcha::audit(const comp::gotcha_data& _data, audit::outgoing, int _ret)
int
blocking_gotcha::operator()(const comp::gotcha_data&, int (*)(const sigset_t*, int*),
const sigset_t* _set_v, int* _sig) const noexcept
{
// if one of the try/timed functions did not succeed, reset the delay value to zero
if(_ret != 0 && _ret != ETIMEDOUT &&
std::set<size_t>{ 1, 3, 5, 8, 11 }.count(_data.index) > 0)
{
delay_value = 0;
}
auto _active = get_thread_state() < ::omnitrace::ThreadState::Internal;
sigset_t _set = *_set_v;
causal_gotcha::remove_signals(&_set);
siginfo_t _info;
int64_t _delay_value = (_active) ? causal::delay::get_global().load() : 0;
auto* _data = blocking_gotcha_t::at(16);
auto f_sigwaitinfo = reinterpret_cast<decltype(&sigwaitinfo)>(_data->wrappee);
causal::sampling::block_backtrace_samples();
auto _ret = (*f_sigwaitinfo)(&_set, &_info);
causal::sampling::unblock_backtrace_samples();
// Woken up by another thread if the call did not fail and this is waking process
if(_active && _ret != -1 && _info.si_pid == process::get_id())
causal::delay::postblock(_delay_value);
if(_ret == -1)
return errno; // If there was an error, return the error code
else
*_sig = _ret; // sig is declared as non-null so skip check
return 0;
}
void
blocking_gotcha::stop()
int
blocking_gotcha::operator()(const comp::gotcha_data&,
int (*_func)(const sigset_t*, siginfo_t*),
const sigset_t* _set_v, siginfo_t* _info_v) const noexcept
{
if(delay_value > 0 && causal::experiment::is_active() &&
get_thread_state() == ::omnitrace::ThreadState::Enabled)
{
causal::delay::postblock(delay_value);
delay_value = 0;
}
auto _active = get_thread_state() < ::omnitrace::ThreadState::Internal;
sigset_t _set = *_set_v;
causal_gotcha::remove_signals(&_set);
siginfo_t _info;
int64_t _delay_value = (_active) ? causal::delay::get_global().load() : 0;
causal::sampling::block_backtrace_samples();
auto _ret = (*_func)(&_set, &_info);
causal::sampling::unblock_backtrace_samples();
// Woken up by another thread if the call did not fail and this is waking process
if(_active && _ret > 0 && _info.si_pid == process::get_id())
causal::delay::postblock(_delay_value);
if(_ret > 0 && _info_v) *_info_v = _info;
return _ret;
}
int
blocking_gotcha::operator()(const comp::gotcha_data&,
int (*_func)(const sigset_t*, siginfo_t*,
const struct timespec*),
const sigset_t* _set_v, siginfo_t* _info_v,
const struct timespec* _wait_v) const noexcept
{
auto _active = get_thread_state() < ::omnitrace::ThreadState::Internal;
sigset_t _set = *_set_v;
causal_gotcha::remove_signals(&_set);
siginfo_t _info;
int64_t _delay_value = (_active) ? causal::delay::get_global().load() : 0;
causal::sampling::block_backtrace_samples();
auto _ret = (*_func)(&_set, &_info, _wait_v);
causal::sampling::unblock_backtrace_samples();
// Woken up by another thread if the call did not fail and this is waking process
if(_active && _ret > 0 && _info.si_pid == process::get_id())
causal::delay::postblock(_delay_value);
if(_ret > 0 && _info_v) *_info_v = _info;
return _ret;
}
} // namespace component
} // namespace causal
+17 -10
Просмотреть файл
@@ -40,11 +40,9 @@ namespace causal
{
namespace component
{
using timespec_t = struct timespec;
// this is used to wrap pthread_mutex()
struct blocking_gotcha : comp::base<blocking_gotcha, void>
{
static constexpr size_t gotcha_capacity = 13;
static constexpr size_t gotcha_capacity = 19;
OMNITRACE_DEFAULT_OBJECT(blocking_gotcha)
@@ -57,20 +55,29 @@ struct blocking_gotcha : comp::base<blocking_gotcha, void>
static void configure();
static void shutdown();
void start();
void audit(const comp::gotcha_data&, audit::outgoing, int);
void stop();
template <typename Ret, typename... Args>
Ret operator()(const comp::gotcha_data&, Ret (*)(Args...), Args...) const noexcept;
private:
int64_t delay_value = 0;
int operator()(const comp::gotcha_data&, int (*)(const sigset_t*, int*),
const sigset_t*, int*) const noexcept;
int operator()(const comp::gotcha_data&, int (*)(const sigset_t*, siginfo_t*),
const sigset_t*, siginfo_t*) const noexcept;
int operator()(const comp::gotcha_data&,
int (*)(const sigset_t*, siginfo_t*, const struct timespec*),
const sigset_t*, siginfo_t*, const struct timespec*) const noexcept;
};
using blocking_gotcha_t =
comp::gotcha<blocking_gotcha::gotcha_capacity,
tim::lightweight_tuple<blocking_gotcha>, category::causal>;
comp::gotcha<blocking_gotcha::gotcha_capacity, tim::type_list<>, blocking_gotcha>;
} // namespace component
} // namespace causal
} // namespace omnitrace
OMNITRACE_DEFINE_CONCRETE_TRAIT(prevent_reentry, causal::component::blocking_gotcha_t,
false_type)
OMNITRACE_DEFINE_CONCRETE_TRAIT(static_data, causal::component::blocking_gotcha_t,
false_type)
OMNITRACE_DEFINE_CONCRETE_TRAIT(fast_gotcha, causal::component::blocking_gotcha_t,
true_type)
+56 -47
Просмотреть файл
@@ -23,18 +23,31 @@
#include "library/causal/components/unblocking_gotcha.hpp"
#include "core/config.hpp"
#include "core/debug.hpp"
#include "core/state.hpp"
#include "library/causal/components/causal_gotcha.hpp"
#include "library/causal/delay.hpp"
#include "library/causal/experiment.hpp"
#include "library/causal/sampling.hpp"
#include "library/runtime.hpp"
#include <timemory/components/macros.hpp>
#include <timemory/hash/types.hpp>
#include <timemory/utility/types.hpp>
#include <csignal>
#include <cstdint>
#include <pthread.h>
#include <stdexcept>
#pragma weak pthread_mutex_unlock
#pragma weak pthread_spin_unlock
#pragma weak pthread_cond_signal
#pragma weak pthread_cond_broadcast
#pragma weak pthread_kill
#pragma weak pthread_sigqueue
#pragma weak pthread_barrier_wait
#pragma weak kill
namespace omnitrace
{
namespace causal
@@ -66,29 +79,15 @@ unblocking_gotcha::configure()
unblocking_gotcha_t::get_initializer() = []() {
if(!config::get_use_causal()) return;
unblocking_gotcha_t::configure(
comp::gotcha_config<0, int, pthread_mutex_t*>{ "pthread_mutex_unlock" });
unblocking_gotcha_t::configure(
comp::gotcha_config<1, int, pthread_rwlock_t*>{ "pthread_rwlock_unlock" });
unblocking_gotcha_t::configure(
comp::gotcha_config<2, int, pthread_spinlock_t*>{ "pthread_spin_unlock" });
unblocking_gotcha_t::configure(
comp::gotcha_config<3, int, pthread_barrier_t*>{ "pthread_barrier_wait" });
unblocking_gotcha_t::configure(
comp::gotcha_config<4, int, pthread_cond_t*>{ "pthread_cond_signal" });
unblocking_gotcha_t::configure(
comp::gotcha_config<5, int, pthread_cond_t*>{ "pthread_cond_broadcast" });
unblocking_gotcha_t::configure(
comp::gotcha_config<6, int, pthread_t, int>{ "pthread_kill" });
unblocking_gotcha_t::configure(
comp::gotcha_config<7, void, void*>{ "pthread_exit" });
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 0, pthread_mutex_unlock);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 1, pthread_spin_unlock);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 2, pthread_rwlock_unlock);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 3, pthread_cond_signal);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 4, pthread_cond_broadcast);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 5, pthread_kill);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 6, pthread_sigqueue);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 7, pthread_barrier_wait);
TIMEMORY_C_GOTCHA(unblocking_gotcha_t, 8, kill);
};
}
@@ -98,35 +97,45 @@ unblocking_gotcha::shutdown()
unblocking_gotcha_t::disable();
}
void
unblocking_gotcha::start()
template <typename Ret, typename... Args>
Ret
unblocking_gotcha::operator()(const comp::gotcha_data& _data, Ret (*_func)(Args...),
Args... _args) const noexcept
{
if(causal::experiment::is_active() &&
get_thread_state() == ::omnitrace::ThreadState::Enabled)
causal::delay::process();
auto _active = get_thread_state() < ::omnitrace::ThreadState::Internal;
if(_active) causal::delay::process();
if(_active && _data.index == 7)
{
int64_t _delay_value = (_active) ? causal::delay::get_global().load() : 0;
causal::sampling::block_backtrace_samples();
auto _ret = (*_func)(_args...);
causal::sampling::unblock_backtrace_samples();
causal::delay::postblock(_delay_value);
return _ret;
}
else
{
return (*_func)(_args...);
}
}
void
unblocking_gotcha::stop()
int
unblocking_gotcha::operator()(const comp::gotcha_data&, int (*_func)(pid_t, int),
pid_t _pid, int _sig) const noexcept
{
if(causal::experiment::is_active() &&
get_thread_state() == ::omnitrace::ThreadState::Enabled)
causal::delay::credit();
}
auto _active = get_thread_state() < ::omnitrace::ThreadState::Internal;
void
unblocking_gotcha::set_data(const comp::gotcha_data& _data)
{
OMNITRACE_SCOPED_THREAD_STATE(ThreadState::Internal);
auto _hash = tim::add_hash_id(_data.tool_id);
auto&& _ident = tim::get_hash_identifier(_hash);
if(_ident != _data.tool_id)
throw ::omnitrace::exception<std::runtime_error>(
JOIN("", "Error! resolving hash for \"", _data.tool_id, "\" (", _hash,
") returns ", _ident.c_str()));
#if defined(OMNITRACE_CI)
OMNITRACE_VERBOSE_F(3, "data set for '%s'...\n", _data.tool_id.c_str());
#endif
if(_active && _pid == process::get_id()) causal::delay::process();
causal::sampling::block_backtrace_samples();
auto _ret = (*_func)(_pid, _sig);
causal::sampling::unblock_backtrace_samples();
return _ret;
}
} // namespace component
} // namespace causal
@@ -39,11 +39,9 @@ namespace causal
{
namespace component
{
using timespec_t = struct timespec;
// this is used to wrap pthread_mutex()
struct unblocking_gotcha : comp::base<unblocking_gotcha, void>
{
static constexpr size_t gotcha_capacity = 8;
static constexpr size_t gotcha_capacity = 9;
OMNITRACE_DEFAULT_OBJECT(unblocking_gotcha)
@@ -56,15 +54,15 @@ struct unblocking_gotcha : comp::base<unblocking_gotcha, void>
static void configure();
static void shutdown();
static void start();
static void stop();
template <typename Ret, typename... Args>
Ret operator()(const comp::gotcha_data&, Ret (*)(Args...), Args...) const noexcept;
static void set_data(const comp::gotcha_data&);
int operator()(const comp::gotcha_data&, int (*)(pid_t, int), pid_t,
int) const noexcept;
};
using unblocking_gotcha_t =
comp::gotcha<unblocking_gotcha::gotcha_capacity,
tim::lightweight_tuple<unblocking_gotcha>, category::causal>;
comp::gotcha<unblocking_gotcha::gotcha_capacity, tim::type_list<>, unblocking_gotcha>;
} // namespace component
} // namespace causal
} // namespace omnitrace
@@ -72,4 +70,6 @@ using unblocking_gotcha_t =
OMNITRACE_DEFINE_CONCRETE_TRAIT(prevent_reentry, causal::component::unblocking_gotcha_t,
false_type)
OMNITRACE_DEFINE_CONCRETE_TRAIT(static_data, causal::component::unblocking_gotcha_t,
false_type)
OMNITRACE_DEFINE_CONCRETE_TRAIT(fast_gotcha, causal::component::unblocking_gotcha_t,
true_type)
+3
Просмотреть файл
@@ -467,6 +467,9 @@ perform_experiment_impl(std::shared_ptr<std::promise<void>> _started) // NOLINT
}
}
// allow ~10 samples to be collected
std::this_thread::sleep_for(std::chrono::milliseconds{ 10 });
double _delay_sec =
config::get_setting_value<double>("OMNITRACE_CAUSAL_DELAY").second;
double _duration_sec =
-12
Просмотреть файл
@@ -97,9 +97,6 @@ int64_t sleep_for_overhead = compute_sleep_for_overhead();
void
delay::process()
{
if(!trait::runtime_enabled<delay>::get()) return;
if(get_state() >= ::omnitrace::State::Finalized) return;
if(causal::experiment::is_active())
{
if(get_global() < get_local())
@@ -126,9 +123,6 @@ delay::process()
void
delay::credit()
{
if(!trait::runtime_enabled<delay>::get()) return;
if(get_state() >= ::omnitrace::State::Finalized) return;
auto _diff = get_global() - get_local();
if(_diff > 0)
{
@@ -139,9 +133,6 @@ delay::credit()
void
delay::preblock()
{
if(!trait::runtime_enabled<delay>::get()) return;
if(get_state() >= ::omnitrace::State::Finalized) return;
auto _diff = get_global() - get_local();
if(_diff > 0)
{
@@ -152,9 +143,6 @@ delay::preblock()
void
delay::postblock(int64_t _preblock_global_delay_value)
{
if(!trait::runtime_enabled<delay>::get()) return;
if(get_state() >= ::omnitrace::State::Finalized) return;
get_local() += (get_global() - _preblock_global_delay_value);
}
+16 -1
Просмотреть файл
@@ -60,7 +60,8 @@ namespace sampling
using ::tim::sampling::dynamic;
using ::tim::sampling::timer;
using causal_bundle_t = tim::lightweight_tuple<causal::component::backtrace>;
using causal_bundle_t =
tim::lightweight_tuple<causal::component::sample_rate, causal::component::backtrace>;
using causal_sampler_t = tim::sampling::sampler<causal_bundle_t, dynamic>;
} // namespace sampling
} // namespace causal
@@ -307,6 +308,20 @@ unblock_samples()
trait::runtime_enabled<causal_sampler_t>::set(true);
}
void
block_backtrace_samples()
{
trait::runtime_enabled<causal::component::backtrace>::set(scope::thread_scope{},
false);
}
void
unblock_backtrace_samples()
{
trait::runtime_enabled<causal::component::backtrace>::set(scope::thread_scope{},
true);
}
void
block_signals(std::set<int> _signals)
{
+6
Просмотреть файл
@@ -45,6 +45,12 @@ block_samples();
void
unblock_samples();
void
block_backtrace_samples();
void
unblock_backtrace_samples();
void block_signals(std::set<int> = {});
void unblock_signals(std::set<int> = {});
+67 -27
Просмотреть файл
@@ -65,13 +65,15 @@ class validation(object):
)
class experiment_data(object):
class throughput_point(object):
def __init__(self, _speedup):
self.speedup = _speedup
self.delta = []
self.duration = []
def __iadd__(self, _val):
self.duration += [float(_val)]
def __iadd__(self, _data):
self.delta += [float(_data[0])]
self.duration += [float(_data[1])]
def __len__(self):
return len(self.duration)
@@ -85,11 +87,48 @@ class experiment_data(object):
def __lt__(self, rhs):
return self.speedup < rhs.speedup
def mean(self):
return mean(self.duration)
def get_data(self):
return [x / y for x, y in zip(self.duration, self.delta)]
def stddev(self):
return stddev(self.duration)
def mean(self):
return sum(self.duration) / sum(self.delta)
class latency_point(object):
def __init__(self, _speedup):
self.speedup = _speedup
self.arrivals = []
self.departures = []
self.duration = []
def __iadd__(self, _data):
self.arrivals += [float(_data[0])]
self.departures += [float(_data[1])]
self.duration += [float(_data[2])]
def __len__(self):
return len(self.duration)
def __eq__(self, rhs):
return self.speedup == rhs.speedup
def __neq__(self, rhs):
return not self == rhs
def __lt__(self, rhs):
return self.speedup < rhs.speedup
def get_data(self):
_duration = sum(self.duration)
return [y / x for x, y in zip(self.arrivals, self.duration)]
def get_difference(self):
_duration = sum(self.duration)
return [x / _duration for x in self.duration]
def mean(self):
rate = sum(self.arrivals) / sum(self.duration)
return sum(self.get_difference()) / rate
class line_speedup(object):
@@ -114,7 +153,7 @@ class line_speedup(object):
return 0.0
_data = []
_base = self.base.mean()
for ditr in self.data.duration:
for ditr in self.data.get_data():
_data += [((_base - ditr) / _base) * 100]
return stddev(_data)
@@ -164,19 +203,18 @@ class experiment_progress(object):
self.data = _data
def get_impact(self):
"""
speedup_c = [x.compute_speedup() for x in self.data]
speedup_v = [x.virtual_speedup() for x in self.data]
impact = []
for i in range(len(self.data) - 1):
x = speedup_v[i + 1] - speedup_v[i]
y_low = speedup_c[i]
y_upp = speedup_c[i + 1]
a_low = x * min([y_low, y_upp])
a_high = 0.5 * x * (max([y_low, y_upp]) - min([y_low, y_upp]))
impact += [a_low + a_high]
"""
impact = [x.compute_speedup() for x in self.data]
y = [speedup_c[i], speedup_c[i + 1]]
y_min = min(y)
y_max = max(y)
a_low = x * y_min
a_upp = 0.5 * x * (y_max - y_min)
impact += [a_low + a_upp]
# impact = [x.compute_speedup() for x in self.data]
return [sum(impact), mean(impact), stddev(impact)]
def __len__(self):
@@ -197,9 +235,12 @@ class experiment_progress(object):
return self.get_impact()[0] < rhs.get_impact()[0]
def find_or_insert(_data, _value):
def find_or_insert(_data, _value, _type):
if _value not in _data:
_data[_value] = experiment_data(_value)
if _type == "throughput":
_data[_value] = throughput_point(_value)
elif _type == "latency":
_data[_value] = latency_point(_value)
return _data[_value]
@@ -232,19 +273,18 @@ def process_data(data, _data, args):
if "delta" in pts:
_delt = pts["delta"]
if _delt > 0:
itr = find_or_insert(data[_selected][_name], _speedup)
itr += float(_duration) / float(_delt)
else:
_diff = pts["arrival"] - pts["departure"] + 1
_rate = pts["arrival"] / float(_duration)
if _rate > 0:
itr = find_or_insert(data[_selected][_name], _speedup)
itr += float(_diff) / float(_rate)
itr = find_or_insert(
data[_selected][_name], _speedup, "throughput"
)
itr += [_delt, _duration]
elif "arrival" in pts and pts["arrival"] > 0:
itr = find_or_insert(data[_selected][_name], _speedup, "latency")
itr += [pts["arrival"], pts["departure"], _duration]
else:
_delt = pts["laps"]
if _delt > 0:
itr = find_or_insert(data[_selected][_name], _speedup)
itr += float(_duration) / float(_delt)
itr += [_delt, _duration]
return data