[rocprofiler-systems]: Add documentation for communication API tracing (#2478)
Add documentation for communication runtime tracing for MPI, UCX, RCCL. --------- Co-authored-by: David Galiffi <David.Galiffi@amd.com>
このコミットが含まれているのは:
@@ -141,6 +141,11 @@ Optional third-party packages
|
||||
(By default, if ROCm Systems Profiler cannot find an OpenMPI MPI distribution, it uses a local copy
|
||||
of the OpenMPI ``mpi.h``.)
|
||||
|
||||
* UCX
|
||||
|
||||
* ``ROCPROFSYS_USE_UCX`` enables UCX (Unified Communication X) support for tracing UCX communication functions
|
||||
used as a transport layer for MPI and other communication libraries
|
||||
|
||||
.. csv-table::
|
||||
:header: "Third-Party Library", "CMake Enable Option"
|
||||
:widths: 15, 45
|
||||
@@ -149,6 +154,7 @@ Optional third-party packages
|
||||
"PAPI", "``ROCPROFSYS_USE_PAPI`` (default: ON)"
|
||||
"MPI", "``ROCPROFSYS_USE_MPI`` (default: OFF)"
|
||||
"MPI (header-only)", "``ROCPROFSYS_USE_MPI_HEADERS`` (default: ON)"
|
||||
"UCX", "``ROCPROFSYS_USE_UCX`` (default: ON)"
|
||||
|
||||
Installing Dyninst
|
||||
-----------------------------------
|
||||
@@ -290,6 +296,20 @@ ROCm Systems Profiler on an application built against OpenMPI causes a segmentat
|
||||
This happens because the value of the ``MPI_COMM_WORLD`` is truncated
|
||||
during the function wrapping before being passed along to the underlying MPI function.
|
||||
|
||||
UCX support within ROCm Systems Profiler
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
ROCm Systems Profiler supports tracing UCX (Unified Communication X) communication functions
|
||||
when ``ROCPROFSYS_USE_UCX`` is enabled (default: ON). UCX is a high-performance communication
|
||||
framework that serves as a transport layer for MPI and other communication libraries, providing
|
||||
optimized point-to-point and collective communication operations.
|
||||
|
||||
When UCX support is enabled, ROCm Systems Profiler can automatically intercept and trace UCX
|
||||
communication functions such as ``ucp_tag_send``, ``ucp_tag_recv``, and other UCP (Unified
|
||||
Communication Protocol) and UCT (Unified Communication Transport) layer operations. This allows
|
||||
for detailed analysis of communication patterns and performance in applications using UCX as
|
||||
their underlying transport mechanism.
|
||||
|
||||
Python support within ROCm Systems Profiler
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
|
||||
新しいイシューから参照
ユーザーをブロックする