TAU

TAU (Tuning and Analysis Utilities) is a program and performance analysis tool framework for high-performance parallel and distributed computing.

TAU provides a suite of tools for static and dynamic analysis of programs written in C, C++, FORTRAN 77/90, Python, High Performance FORTRAN, and Java.

In particular, TAU offers a state-of-the-art performance profiling and tracing facility that supports a general scalable parallel execution model based on nodes, contexts, and threads.

The goal of the TAU project is to develop program and performance analysis technology that meets both the challenges of evolving scalable parallel computing systems and the needs of programming methodologies used for next generation scientific applications.

TAU should be able to target the diversity of computing paradigms and machines while offering a framework of portable and reconfigurable measurement and analysis components that can be optimized and extended.

What is new in this release:

• This version adds a port to ARM Linux, Intel MIC platforms, CUDA 5.0 support, UPC Runtime instrumentation for Cray CCE and BUPC, improvements in the MPI wrapper, support to ParaProf 3D for IBM BG/Q, and ARM Linux.

What is new in version 2.21.2:

• Major changes include SHMEM profiling for Cray and SGI, NVIDIA OpenCL and AMD OpenCL support, a tau_rewrite tool based on MAQAO, OpenSHMEM Profiling, Score-P Atomic/Context Events, OpenMP 3.0 instrumentation with Opari2, NVIDIA CUPTI v4.1, H2 database in PerfDMF, debugging with TAU_TRACK_SIGNALS, MinGW compiler support for Windows, UPC source level instrumentation using the Rose parser to support the Cray CCE compiler, and PDT 3.17 with Rose and MAQAO binary instrumentor.

What is new in version 2.19.1:

• ParaProf:
• ParaProf has a new expression window that allows you to easily create
• derived metrics (Options -> Show Derived Metric Panel. You may also pass
• a list of expressions in a file to paraprof and apply the expressions at
• any level (application, experiment, trial in a profile file or in a
• database. This window also appears in perfexplorer.
• Paraprof has an updated 3D communication window that features sliders
• and cross hairs for accurately evaluating the extent of point-to-point
• communication between sender and receiver nodes. To use the communication
• matrix displays in paraprof, please set the TAU_COMM_MATRIX environment
• variable to 1 prior to executing the instrumented application.
• A new group changer window is available for changing the group
• membership of events during the analysis phase. This window (accessible
• via Windows -> Group Changer allows you to select a subset of routines
• using multiple selection or wildcards (e.g., HDF5 and associate it with a
• new group (e.g., IO. You may save the file, upload it in a database and
• use it in PerfExplorer's windows (e.g., 'Charts -> Group % of total
• runtime' to see the scaling of IO operations.
• The filter box allows you to use regular expressions to filter which
• functions/regions you want to change. There is an implicit * added to
• the end of the filter, so that you can just type MPI_ and you will see
• all the MPI routines for example.
• Once you select one or more function/regions, the current and available
• groups will be displayed. If the functions you've selected do not have
• the same set of groups, then the intersection of their groups will be
• black, and the rest will be grey, so that you can either add that group
• to the rest, or remove it from all of them.
• Fixed tree selection model after the changes made for multi-selection.
• Now if you right-click on an unselected object, it becomes the select
• object and the pop-up appears only for this.
• 3D visualization now works on IBM BG/P using ppc64 JOGL.
• Better Perfsuite support for parsing perfsuite 1.0.0a4 profiles.
• PerfExplorer:
• Perfexplorer's code has been updated to Weka 3.6.1. You will need to
• run perfexplorer_configure to use an older PerfDMF installation in the new
• perfexplorer. DBSCAN clustering is now operational, and you can launch the
• Weka Explorer directly from a dataset. Initially, the top 10
• "interesting" functions are exported to a CSV file, and automatically
• loaded into Weka. From there, you can do all kinds of fun data mining.
• The DBSCAN method will first normalize the data, then find the k-distances
• for each process, where k is the minimum number of points to be found in
• the cluster. Then the k-distances are sorted descending, and the
• k-distances are tried as various neighborhood distances for the
• clustering. Any clustering which finds a new number of clusters is saved.
• Updated scripts to support the new version of Jython ( 2.5.1 that
• supports an actual up-to-date standard for Python, version 2.5. The new
• version requires a registry file, and a user-writable location to put the
• cachefiles when the jars are pre-processed to find symbols.
• Created a utility which will reconstruct a Paraver trace from TAU EBS samples.
• tau_instrumentor:
• The tau_instrumentor tool now supports automatic loop level
• instrumentation in the SPEC mode. The spec mode allows tau_instrumentor to
• insert instrumentation calls in a copy of the original source code for any
• tool. Earlier this was used by tools such as Scalasca for routine level
• instrumentation only. Now, it supports loop level instrumentation for
• external tools as well.
• Hybrid source and compiler-based instrumentation:
• When the automatic source level instrumentation (using PDT and
• compilation does not produce a .o file (due to a possible parsing error,
• TAU used to revert to the original compilation command line. Now, it
• invokes the compiler to use compiler-based instrumentation to insert
• routine level instrumentation in the code. This leads to a better
• instrumentation coverage of the code. This change should be transparent to
• the user. To disable this new feature, the user must explicitly pass the
• -optNoCompInst option in the TAU_OPTIONS environment variable. Care must
• be taken while defining exclude or include lists of routines and files in
• selective instrumentation files. The routine names may differ (for PDT and
• compiler-based instrumentation and TAU's compiler scripts are case
• sensitive, but they allow the use of wildcards in names.
• Support for external packages.
• TAU now supports the recent PAPI-C 4.0.0 release as well as Scalasca 1.3.0 and DyninstAPI 6.1 release.
• New Platforms:
• TAU supports the Cray CCE compilers (for C, C++ and Fortran on the Cray
• XT systems (simply invoke 'module load PrgEnv-cray' before configuring
• TAU. Compiler-based instrumentation is not yet supported. TAU also
• features IBM BG/P configuration enhancements.
• Outreach:
• We will be giving a full day POINT tutorial at the LCI'10 conference in
• Pittsburgh (March 8. We will be distributing the POINT VI-HPS LiveDVDs
• for training.

What is new in version 2.18.1:

• Support was added for GPGPUs in PGI Compiler and for PGI compilers on Mac OS X.
• Workflow support under Eclipse, Charm++ support, and OpenMP support were improved.
• The Python API was accelerated.
• The documentation was updated.
• Assorted bugs were fixed.

What is new in version 2.17.3:

• Updates include compiler-based instrumentation for Intel, GNU, and PathScale compilers, a new Python API for memory tracking, fixes for IBM BG/P configuration, and support for CQoS analysis and drawing charts from script files in PerfExplorer.