[Tau-announcements] SC'13: TAU, PDT, HPC Linux released

Sameer Shende sameer at cs.uoregon.edu
Fri Nov 15 18:36:03 PST 2013


	We are pleased to announce the release of TAU v2.23, PDT v3.20, and the HPC Linux LiveDVD:


Our SC'13 demo schedule is at the end of this e-mail.
The following new features have been added since TAU v2.22.2 released on May 27, 2013:

1. OpenMP Tools API (OMPT)
TAU supports the OMPT interface for Intel, IBM, and OpenUH compilers. For Intel, please configure TAU using the -ompt=download option without using -opari. This allows tracking of OpenMP constructs without rewriting the source code. It uses hooks in the OpenMP runtime layer to track performance data and attribute it to the callsite using libunwind (where available). 

2. Support for Power Profiling.
To use power profiling, TAU may be configured with PAPI (v5.2+) using RAPL. Then, a user may insert the following source instrumentation call:

#include <TAU.h>
to trigger recording of power (CPU, integrated GPU, socket, DRAM) periodically. The sampling period can be specified and tracking events can be enabled or disabled at runtime.  
To use this, all user DSOs must be linked in using -Wl,-rpath,dir and the instrumented binary must be granted special privileges as root:

% setcap cap_sys_rawio=ep a.out
and /dev/cpu/*/msr files should be readable by the user. 

3. Support for special marker events
Marker events are triggered by TAU when an atomic or a context event varies significantly. When such an event is triggered with a value significantly above or below the observed maxima or minima by a user specified threshold, a marker context event is triggered. For instance, if the CPU power consumption is being tracked at ~65W and suddenly it jumps up to 115W, a max marker event will be triggered with the callstack where this jump takes place. Similarly, if it falls down to 15W, it will record a min marker. The threshold is controlled by TAU_EVENT_THRESHOLD environment variable (e.g., TAU_EVENT_THRESHOLD=0.25 uses 25% higher or lower values).

4. Lower OpenMP runtime overhead
TAU's OpenMP thread layer has been re-engineered and it has a significantly reduced overhead. TAU also features improvements in its callsite tracking module to work more smoothly with multi-threaded applications.

5. Offline address lookup
TAU supports a new TAU_BFD_LOOKUP=0 option along with tau_resolve_addresses.py script for offline BFD lookups for event based sampling (TAU_SAMPLING=1). 

6. Improved heap memory usage tracking
TAU includes events to track increase and decrease of heap memory for each instrumented routine. This is recorded when TAU_TRACK_HEAP=1 is used. Coupled  with marker events, we can now see which routines show sudden jumps in heap memory increase or decrease. 

7. CUPTI enhancements
CUPTI counters are now assigned to the GPU kernels that produced them. If they cannot be attributed to a single kernel the values are automatically estimated given the structure of the application.

8. Improvements in ParaProf
Updated Cube reader allows reading metadata and atomic events from Cube profiles in ParaProf. TAU features a more intelligent management of ParaProf/PerfExplorer jar downloads. It allows editing of existing TauDB views in ParaProf. ParaProf can now launch an event window from a node view. It supports adding/removing metadata fields for TauDB trials in ParaProf.

9. Improvements in PerfExplorer
TAU adds a new UI element (DrawMultiGraph) to the PerfExplorer Jython scripting API. TAU adds an option to hide the legend in PerfExplorer charts. It supports sorted event lists in the PerfExplorer custom chart. 

10. Improvements to TAU's MPC support
TAU supports tracking MPC MPI layer (http://mpc.sourceforge.net) when TAU is configured with -c++=mpc_cxx -cc=mpc_cc. It tracks performance data on a per user-level thread basis. 

PDT v3.20
Supports binary rewriting for Intel Xeon-Phi Co-processor (MIC) binaries on x86_64 Linux systems using MAQAO (http://www.maqao.org). Updates for Multiarch to better support Ubuntu and Debian distributions. It is available for download from:
http://tau.uoregon.edu/pdt.tgz or from the PDT webpage. 

HPC Linux 
TAU, PDT, and VI-HPS (www.vi-hps.org) and ACTS (acts.nersc.gov) tools are featured on the HPC Linux LiveDVD and virtual images available for download from:


Our team will be at SC'13. Please stop by the PGAS Booth (#432) at SC to pick up the LiveDVDs or talk with one of our team members. Also, please join us at:

a. Full day Tutorial
M. Geimer, S. Shende, B. Wesarg, B. Wylie, Full day tutorial, Sunday, Nov. 17
Hands-On Practical Hybrid Parallel Application Performance Engineering
8:30AM - 5:00PM
Room 301

b. Broader Engagement, Sunday Nov. 17
S. Shende, "Performance Evaluation Using the TAU Performance System"
11:00AM - 11:30AM

c. Poster:
D. Rouson, K. Morris, M. Haveraaen, J. Xia, S. Shende, Tuesday Nov. 19
High-Performance Design Patterns for Modern Fortran
5:15PM - 7:00PM
Mile High Pre-Function

d. Poster:
D. Rouson, H. Radhakrishnan, K. Morris, S. Shende, S. Kassinos, Tuesday, Nov. 19th
Test-Driven Parallelization of a Legacy Fortran Program
5:15PM - 7:00PM
Mile High Pre-Function

e. Workshops:
1st International Workshop on Software Engineering for HPC in Computational Science and Engineering (SE-HPCCSE), Fri, Nov. 22

H. Radhakrishnan, D. Rouson, K. Morris, S. Shende, S. Kassinos,
Test-Driven Parallelization of a Legacy Fortran Program
Friday, Nov. 22, 10:30am, 

f. Workshops:
Workshop on Extreme-Scale Programming Tools
Engineering Performance Tools for Extreme Scale, Monday, Nov. 18
S. Shende, A. Malony, K. Huck, J. Linford, W. Spear, S. Biersdorff

g. BoF
S. Shende, TAU Performance System, and Welcome to PGAS 2014, Eugene, OR

h. PGAS Booth, TAU demos,  #432

i. OpenSHMEM BoF
TAU Demo, DOE Booth 1327, 11AM, Wed. Nov. 20th

j. OpenSHMEM Tools TU Dresden Booth 3905, Wed. Nov. 20th 4-5PM. 

   Please let us know if we may assist you with our tools in any way.
   - Sameer
  (for tau-team@ cs.uoregon.edu)

More information about the Tau-announcements mailing list