System News
HPC Profiling with the Sun Studio Performance Tools
Techniques Implemented in Tools to Profile HPC Applications
November 10, 2009,
Volume 141, Issue 2

Sun Studio Performance Tools collect performance data on fully optimized and parallelized applications written in C, C++, Fortran, or Java, and any combination of these languages.
 

Marty Itzkowitz and Yukon Maruyama describe how to use the Sun Studio Performance Tools to understand the performance issues in single-threaded, multi-threaded, OpenMP and MPI applications, and the techniques used to profile them. Presented at the Third Parallel Tools Workshop held in Dresden, Germany back in September, this technical paper outlines techniques implemented in the Sun Studio Performance Tools to profile HPC applications. The 30-page PDF is hosted on Sun's developer's site.

The two authors explain the first step in understanding performance problems is to first determine if one truly exists. They recommend using a repeatable example (aka benchmark) that uses input data and problem size of a scale comparable to the production runs for which the program is being tuned. System administrators also need to ask themselves what can be changed to improve performance of the application.

Itzkowitz and Maruyama proceed by examining single-threaded application performance issues like algorithmic inefficiency, memory subsystem problems and dataspace profiling, and how to go about detecting if a problem exists within these areas through various examples. They then explore additional issues presented by multi-threaded applications that could affect performance. They specifically target lock contention and false sharing of cache lines.

Performance topics relating to the OpenMP programming model cover excess parallel overhead, insufficient parallelism, lock contention and load imbalance. In regards to MPI programs, the paper considers performance concerns relating to computation issues and parallelization.

Additionally, the paper demonstrates how to use the MPI timeline and charts to visualize job behaviors, such as message traffic and work distribution. It also explains how MPI filters can be used to pick out behaviors of interest and determine which events are responsible.

More Information

HPC Profiling with the Sun Studio Performance Tools

Sun Studio C, C++ & Fortran Compilers and Tools

Major Upgrade: Sun Studio 12 Update 1 [...read more...]

Keywords:

fullsource
 

Other articles in the HPC section of Volume 141, Issue 2:
  • HPC Profiling with the Sun Studio Performance Tools (this article)

See all archived articles in the HPC section.



News and Solutions for Users of Solaris, Java and Oracle's Sun hardware products
Just the news you need, none of what you don't – 42,000+ Members – 24,000+ Articles Published since 1998