Updates:
* Added rsmi_dev_nps_mode_set and rsmi_dev_nps_mode_get
* Added ability to set multiple SYSFS files in debug build
* Added ability to see user's env variables set for debug build
* Added tests for rsmi_dev_nps_mode_set and rsmi_dev_nps_mode_get
* Added ability to restart AMD GPU driver, used in nps_mode_set
* Updated ROCm_SMI_Manual.pdf to include new APIs
* Added progress bar for long running python_smi_tools, used
in setting nps_mode if runs longer than .1 seconds
Change-Id: I6d61bedd28d7cba6aff432ad2d127ba741b7d15a
Signed-off-by: Charis Poag <Charis.Poag@amd.com>
The Shader Engines number should be shadder array_count divided by simd_arrays_per_engine
not array_count.
Signed-off-by: Xiaogang Chen<Xiaogang.Chen@amd.com>
Change-Id: I808d1fedd6b9843500719e902ecf759f5668a7d1
ROCr internally uses the same allocation_map_ list to track memory
allocations that are both for internal allocations and allocations by
users of ROCr library. In some edge cases, the library user would call
hsa_amd_pointer_info on an invalid pointer, but ROCR would return the
pointer as valid because this pointer belongs to a memory range that
was allocated internally within ROCr. Adding a flag to differentiate
between internal and external allocations.
Change-Id: I98c52bd85f3985d1ba1b0e3101d2254b003412cf
- 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
* Update timemory submodule for OMPT
- Updated OMPT support for OpenMP 5.2
* omnitrace exe cleanup
- fixed "omnitrace --" segfault
- added nullptr checks
* CMake updates
- moved omnitrace-interface-library definition up a directory
- general cleanup
- fixed branch/tag/ref for git submodule checkouts
* Improve shutdown of causal profiling after duration limit
* Fix dyninst minimum version number
* Removed debug print from binary::get_link_map
* Remove use of thread-pool in causal
* Relax causal testing when variance is high
* causal_gotcha utilities for blocking signals
* Tweak to causal example
* Install validate-causal-json as omnitrace-causal-print
* simplify address_multirange
* improve causal line saving
* Always print PID in log messages
* omnitrace-dl updates
- omnitrace_preload does not call omnitrace_init or omnitrace_init_tooling
- omnitrace_preload will call omnitrace_set_mpi if OMNITRACE_USE_MPI
or OMNITRACE_USE_MPIP in the env is true but not call it otherwise
because doing so either overrides OMNITRACE_USE_PID (when true) or
disable mpip from initialization (when false) and the MPI
init can be caught later and override OMNITRACE_USE_PID
* config updates
- set_setting_value sets user update type
- remove volatile from get_settings_configured
- don't override settings::default_process_suffix
- don't kill process in omnitrace_exit_action
- set_state ignores updating state if >= State::Finalized
* Handle state > State::Finalized
* fork gotcha updates
- unsets LD_PRELOAD
- sets OMNITRACE_ROOT_PROCESS
- sets OMNITRACE_CHILD_PROCESS
* libomnitrace library.cpp updates
- basic_bundle for fini metrics
- handle finalization from child process
* sampling updates
- sampling::shutdown handles when child process
* Add example and test using fork
* Update run-ci script to support not submitting
* Tweak test envs
* Update build flags when codecov enabled
* remove unnecessary includes of sampling header
* Replace mpi copy/fini static lambda with free-funcs
* Update codecov job
* Fix OMPT segfaults after finalization
* Miscellaneous updates after rebase
* fixes for causal profiling
* revert some run-ci.sh changes
* Disable storing env in sampling::shutdown
* formatting fix
* Update timemory submodule
- fixed occasional synchronization issues with allocator offloading
- exclude protozero:: from internal samples
* improve root/child process detection
- avoid omnitrace_finalize in MPI when child process
- revert some testing tweaks
## Overview
This PR attempts to increase the stability of binary rewrite and runtime instrumentation.
### Improved protection against self-instrumentation
Using ~~the binary analysis capabilities added from #229~~ the Dyninst SymtabAPI, OmniTrace now does a much better job of avoiding instrumentation of functions which are internally called by OmniTrace:
- The `omnitrace` executable searches for and parses the symbols of various libraries which are known to cause problems when instrumented
- GNU libraries which are common to nearly every library, e.g., `"libc.so.6"`, `"libdl.so.2"`, etc., and thus are outside the scope of the users optimizations efforts
- Libraries which OmniTrace depends on for functionality, e.g. `"libunwind.so"`, `"libgotcha.so"`, `"libroctracer64.so"`, etc.
- OmniTrace skips instrumenting any `module_function` instance when it's member `module_name` or `function_name` variable matches the library name, source file, or function name found for that symbol (unless the user explicitly requests that it be eligible for instrumentation)
- Note: the parsing of the "internal" libraries may result in longer instrumentation time and higher memory usage. Please file an issue if either of these is found to be excessive.
### Function filters based on linkage and visibility
Added options to restrict instrumentation to certain linkage types (e.g. avoid instrumenting weak symbols) and visibility types (e.g. avoid instrumenting hidden symbols).
### Function filters based on instructions
In the past, after instrumentation, some applications instrumented by Dyninst would fail with a trap signal (e.g. #147). In several cases, it was found that this occurred whenever certain instructions were present in the function so an option was added to exclude functions based on one or more regex patterns was added.
## Details
- generates list of "internal" libraries and attempts to find the first match via:
- the library is already open, e.g. `dlopen(<libname>, RTLD_LAZY | RTLD_NOLOAD)`
- searching for the library in `LD_LIBRARY_PATH`
- searching for the library in `OMNITRACE_ROCM_PATH`, `ROCM_PATH`
- searching the folders from `/sbin/ldconfig -p`
- searching for the library in common places such as `/usr/local/lib`
- provides new `--linkage` command line option to restrict instrumentation to functions with particular type(s) of linkage
- Linkage types: `unknown`, `global`, `local`, `weak`, `unique`
- provides new `--visibility` command line option to restrict instrumentation to functions with particular type(s) of visibility
- Visibility types: `unknown`, `default`, `hidden`, `protected`, `internal`
- provides new `--internal-module-include` and `--internal-function-include` command line regex options to bypass automatic exclusion from instrumentation
- provides new `--internal-library-append` command line option to specify a library should be considered internal
- provides new `--internal-library-remove` command line option to specify a library should not be considered internal
- provides new `--instruction-exclude` command line regex option to exclude functions which contain matching instructions
- provides new `--internal-library-deps` command line option to treat libraries linked to internal libraries as internal libraries
- generally, this will only be helpful during runtime instrumentation when OmniTrace is built with an external dyninst library which is dynamically linked to boost libraries and the application is using the same boost libraries
- relaxed restrictions in `module_function::is_module_constrained()`
- relaxed restrictions in `module_function::is_routine_constrained()`
- added a few miscellaneous nullptr checks
## Miscellaneous
- Fix `LD_PRELOAD` + `OMNITRACE_DL_VERBOSE=3` issue
- Adds a sampling offload verbose message
- Improves MPI send-recv.cpp example error message
- Minor tweaks to binary library
- `binary::get_linked_path` returns `std::optional<string>`
- renamed `binary::symbol::read_bfd` to `binary::symbol::read_bfd_line_info`
- `binary::get_binary_info` has param options for reading line info and included undefined symbols
- fixed another edge case instance of resource deadlock during first call to configure_settings
- improved the error log printing in `omnitrace` (does not print repeated messages)
* fix OMNITRACE_DL_VERBOSE=3 + preload issue
- join needs to handle nullptr
* sampling offload verbose message
* mpi-send-recv error message
* binary updates
- get_linked_path returns std::optional<string>
- get_binary_info accepts include_undef flag
- renamed symbol::read_bfd to symbol::read_bfd_line_info
- get_binary_info has param options for reading line info and included undefined symbols
* config updates (initialization)
- fixed another instance of resource deadlock during first call to configure_settings
* Testing fix for HIP w/o rocprofiler support
- disable rocprofiler tests when HIP enabled but OMNITRACE_USE_ROCPROFILER=OFF
* omnitrace exe: insert_instr nullptr check
* omnitrace exe: new method for determining internal constraints
- added internal-libs.cpp
- using binary::get_binary_info on various known libs used by omnitrace
- any matching func/file from symbols found in known internal libs are excluded
- relaxed restrictions in is_module_constrained
- relaxed restrictions in is_routine_constrained
- added a few safety checks
* internal libs append/remove
- options to change which libs are considered internal libraries
* omnitrace exe instruction exclude
- regex option for excluding functions containing specific instructions
* fix is_internal_constrained
* binary link map verbose message
* support constraints on linkage and visibility of symbols
* misc fixes
- fix compiler error for Ubuntu Jammy + GCC 12
- dlopen + libtbbmalloc_proxy appears to be causing issues on OpenSUSE
* Performance details + MT
- multithread processing internal info
- report timing info
* Defer parsing internal data
- wait until after address space is created
* Performance improvement finding for get_symtab_function
* fix data race in get_binary_info
* remove set_default for linkage and visibility argparse
* Parse internal libs with Dyninst::Symtab instead of binary reader
- conflicting versions of libraries for binary analysis causes problems
- expanded whole function restrictions
- expanded module_function::is_routine_constrained regex
* internal lib updates
- include memory usage info
- option to read libraries linked against internal libs: --internal-library-deps
- defer parsing internal libs data to when processing modules
Track and report the size, in bytes, of pending unexecuted blit
commands. To be used in copy ganging.
Change-Id: Ia7453ff88571e927df771c6c819b73c17e67708e
- The primary feature of this PR is the **addition of support for scoping the collection of tracing/profiling data into one or more time-based windows**
- Closes#222
- Closes#207
- Support for a real-clock time delay and/or a duration for tracing/profiling was added, *resembling the support for this feature during sampling and process-sampling*
- However, above paradigm was enhanced for tracing
- Instead of one delay and/or one duration based on real time, ***tracing supports periodic and varying delays and durations and these delay+duration sets can be controlled with different clocks***
- At some point, this capability will be extended to sampling and process-sampling
- A secondary feature of this PR are the improvements to the handling of categories (by-product of the primary feature)
- For example, previously setting `OMNITRACE_ENABLE_CATEGORIES` to a specific set of categories only eliminated the disabled categories from the perfetto trace, now these are applied to timemory profiles too
- A new configuration variable `OMNITRACE_DISABLE_CATEGORIES` was added for when disabling only a handful of categories is easier
- There are quite a few miscellaneous modifications which pollute this PR a bit
## Multiple Tracing Windows
As noted above, tracing now supports specifying multiple delays and durations _and_ with different clocks. Consider the configuration below with two entries in the format `<DELAY>:<DURATION>:<REPEAT>:<CLOCK_TYPE>`:
```console
OMNITRACE_TRACE_PERIODS = 0.5:1.0:2:realtime 10.0:5.0:3:cputime
```
The above configuration defines:
1. `0.5:1.0:2:realtime`
- A delay of 0.5 seconds (real-time)
- Followed by a data collection duration of 1 second (real-time)
- This delay + duration is repeated 2x
- Summary: tracing data is collected for 2 out of the first 3 seconds of the application's execution
2. `10.0:5.0:3:cputime`
- A delay of 10 seconds (process _CPU-time_)
- Followed by a data collection duration of 5 seconds (process _CPU-time_)
- This delay + duration is repeated 3x
- Summary: tracing data is collected for a total of 15 seconds of process CPU-time in the ensuing 75 seconds of CPU-time during the application execution.
- Note: the elapsed CPU-time is the aggregate of the CPU-time consumed by all the threads in the process and should be scaled accordingly, e.g., 4 threads running constantly for 1 second of real-time is ~4 seconds of CPU time.
## `omnitrace-sample` Changes
Formerly, `--wait` and `--duration` command-line options only applied to sampling delay and duration. The value of these options are now applied to the tracing delay and duration. To retain the ability to control sampling delay/duration without setting tracing delay/duration or vice versa, `--sampling-wait`, `--sampling-duration`, `--trace-wait`, and `--trace-duration` options were added. `omnitrace-sample` also has new options for most of the new configuration options detailed below.
## New configuration options
| Option | Description |
| ------- | ----------- |
| `OMNITRACE_DISABLE_CATEGORIES` | inverse behavior from `OMNITRACE_ENABLE_CATEGORIES` -- populates list of all available categories and then removes the specified ones. |
| `OMNITRACE_TRACE_DELAY` | Single floating-point number specifying time to wait before starting data collection. Analagous to `OMNITRACE_SAMPLING_DELAY` and `OMNITRACE_PROCESS_SAMPLING_DELAY` |
| `OMNITRACE_TRACE_DURATION` | Single floating-point number specifying data collection duration. Analagous to `OMNITRACE_SAMPLING_DURATION` and `OMNITRACE_PROCESS_SAMPLING_DURATION` |
| `OMNITRACE_TRACE_PERIOD_CLOCK_ID` | Sets the default clock-type for tracing delay/duration. Always applied to above two options, can be overridden in below option. Accepts `CLOCK_REALTIME`, `CLOCK_MONOTONIC`, `CLOCK_PROCESS_CPUTIME_ID`, `CLOCK_MONOTONIC_RAW`, `CLOCK_REALTIME_COARSE`, `CLOCK_MONOTONIC_COARSE`, `CLOCK_BOOTTIME`. See `man 2 clock_gettime` for details on differences. |
| `OMNITRACE_TRACE_PERIODS` | More powerful version for specifying delay + duration. Supports formats: `<DELAY>`, `<DELAY>:<DURATION>`, `<DELAY>:<DURATION>:<REPEAT>`, and `<DELAY>:<DURATION>:<REPEAT>:<CLOCK_ID>`. |
## Miscellaneous Changes
- Expanded `critical_trace_categories_t` to include tracing data from MPI, pthread, HIP, HSA, RCCL, NUMA, and Python.
- Added categories `thread_wall_time` and `thread_cpu_time` (derived from sampling)
- Read DWARF info for breakpoints
- Relocated some source code
- Reason: necessary to make `libomnitrace` a bit more modular. Eventually, a large chunk will be separated into `libomnitrace-core`, `libomnitrace-binary`, etc. in order to facilitate re-usability
- Relocated some functionality from `runtime.cpp` to `config.cpp`
- Relocated code using rocm-smi library to query number of devices to `gpu.cpp` (where the code for using HIP to query number of devices is)
- Relocated code for perfetto config and perfetto session out of tracing namespace to reside with other perfetto code
- `OMNITRACE_COLORIZED_LOG` configuration option renamed to `OMNITRACE_MONOCHROME`
- Backwards compatibility via a deprecated option was not retained here since the logic changed (i.e. true in former means false in latter)
- Replaced `TIMEMORY_DEFAULT_OBJECT` macro with `OMNITRACE_DEFAULT_OBJECT` macro
- Updated some code in roctracer to use `component::category_region` instead of explicitly using `tracing::` functions
- Updated `backtrace_metrics` to better support controlling their presence in the traces/profiles via categories
- Added support for `--print` in `validate-timemory-json.py`
- Generic `OMNITRACE_ADD_VALIDATION_TEST` CMake function
## Git Log
* OMNITRACE_DEFAULT_OBJECT
- replace TIMEMORY_DEFAULT_OBJECT with TIMEMORY_DEFAULT_OBJECT
* trace-time-window example + tests
- adds cmake OMNITRACE_ADD_VALIDATION_TEST function for testing
- validate-timemory-json.py now supports printing (-p)
- update to OMNITRACE_STRIP_TARGET
* Update timemory submodule
- detailed backtrace print /proc/<PID>/maps
- operation::push_node verbosity change
- storage::insert_hierarchy use emplace + at instead of operator[]
- concepts::is_type_listing
- argparse updates for start/end group
- argparse color fixes
* perfetto updates
- Remove OMNITRACE_CUSTOM_DATA_SOURCE CMake option
- move tracing::get_perfetto_config and tracing::get_perfetto_session to perfetto.cpp
* config and runtime updates
- OMNITRACE_DISABLE_CATEGORIES option
- get_enabled_categories() + get_disabled_categories()
- config impl handles populating them
- OMNITRACE_TRACE_DELAY option
- OMNITRACE_TRACE_DURATION option
- OMNITRACE_TRACE_PERIODS option
- {get,set}_signal_handler
- removes config.cpp link dependency for omnitrace_finalize
- get_realtime_signal() + get_cputime_signal() + get_sampling_signals()
- moved from runtime.cpp to config.cpp
* utility::convert
- helper function for converting string to a type
* pthread_create_gotcha + thread_info updates
- thread_index_data::as_string()
- tweak printing info about new thread / exited thread
* binary updates
- get_binary_info has arg to disable dwarf parsing
- binary_info contains vector of breakpoint addresses
- binary_info:filename() function
- binary::get_linked_path
- binary::get_link_map has args for dlopen mode
- symbol::read_dwarf -> symbol::read_dwarf_entries
- symbol::read_dwarf_breakpoints
* library updates + categories impl
- implement config::set_signal_handler
- categories.cpp for handling trace delays
- implement trace delay/duration/periods
* concepts + debug + defines
- tuple_element in concepts
- removed runtime header from debug header
- OMNITRACE_DEFAULT_COPY_MOVE
* gpu + rocm_smi
- moved rsmi_num_monitor_devices call to gpu.cpp
- gpu::rsmi_device_count()
* roctracer updates
- roctracer_bundle_t -> roctracer_hip_bundle_t
- use category_region instead of explicit tracing push/pop calls
* sampling + backtrace_metrics
- rework backtrace_metrics to support categories
* tracing updates
- category stack counters (i.e. push vs. pop counter) for profiling and tracing
- push_timemory and pop_timemory accept string_view instead of const char*
- tweaked the pop_timemory hash search
- {push,pop}_perfetto theoretically supports same invocations as for {push,pop}_perfetto_ts and {push,pop}_perfetto_track
- mark_perfetto, mark_perfetto_ts, mark_perfetto_track
* category_region update
- expanded the critical trace categories
- use category_push_disabled
- use category_pop_disabled
- use category_mark_disabled
* constraint implementation
- This provides generic functionality for constraining data collection within a windows of time.
- E.g., delay, delay + duration, (delay + duration) * nrepeat
* COLORIZED_LOG -> MONOCHROME
* constraint + omnitrace-causal + omnitrace-sample updates
- support for using different clock IDs for constraints
- OMNITRACE_TRACE_PERIOD_CLOCK_ID option
- tweak to trace-time-window example
- tweak to trace-time-window tests
* Fix formatting
* Update time-window tests
- Fix detection of validation support for perfetto
- Using the --caller-include feature + runtime instrumentation on Ubuntu 18.04 and OpenSUSE 15.2 results in a segfault in the internals of Dyninst.
- For now, mark that these tests will fail
- Later, determine if updating Dyninst submodule fixes this problem
* Fix OMNITRACE_OUTPUT_PATH for all tests
- Provide absolute path instead of relative
* Tweak lambda for checking whether HW counters are enabled
- causing strange build errors on older GCC compilers
* Update dyninst submodule
- fix issues with using --caller-include for Ubuntu 18.04, OpenSUSE 15.x
* cmake formatting
* fix sampling compiler issue for GCC 8
* Tweak thread create message
* Increase causal validation iterations
* Move hsa_init out of rocmtools and into RDC
* Remove secondary hsa_shut_down from ROCR module
Change-Id: I57d84d41ddc51595b98e734265f10bc5129a7352
Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>
Depends-On: I2b389ee1a9ba3507b2df1fc2fe83598f67731aac
Ranjith Ramakrishnan
coleramos425
Charis Poag
David Yat Sin
Bill(Shuzhou) Liu
Karl W. Schulz
Jonathan Kim
Xiaogang Chen
Cordell Bloor
Jonathan R. Madsen
Nirmal Unnikrishnan
Sean Keely
Konstantin Zhuravlyov
Hao Zhou
Galantsev, Dmitrii