The Other HPC: Profiling Enterprise-scale Applications

Marty Itzkowitz Senior Principal Engineer, Oracle

ABSTRACT

In this talk we will discuss the difference between traditional HPC applications and what I am calling "The Other HPC". Where traditional HPC applications tend to be either OpenMP or MPI, the other HPC applications use a variety of inter-thread and inter-process communications and synchronizations; where traditional HPC applications are written in Fortran, C, and C++, the other HPC application are primarily written in Java and C++. While the other HPC applications are not quite a petascale today, they are moving in that direction.

Typical Enterprise-class applications have many threads and processes, and have a complex startup, often taking many minutes or more. Their steady-state behavior is transaction-driven. Profiling such applications requires the ability to enable and disable data collection to get the clean mid-stream sample that is desired for analyzing the steady-state behavior. It also requires techniques for drilling down into the data, and filtering for specific time ranges, and for processes or threads.

We will then show performance measurements on a variety of these applications, primarily benchmarks, and discuss the issues in understanding their performance.