Scalable performance measurement infrastructure for parallel codes (Score-P)

The instrumentation and measurement framework Score-P, together with analysis tools build on top of its output formats, provides insight into massively parallel HPC applications, their communication, synchronization, I/O, and scaling behaviour to pinpoint performance bottlenecks and their causes.

Score-P is a highly scalable and easy-to-use tool suite for profiling (summarizing program execution) and event tracing (capturing events in chronological order) of HPC applications.

The scorep instrumentation command adds instrumentation hooks into a user's application by either prepending or replacing the compile and link commands. C, C++, Fortran, and Python codes as well as contemporary HPC programming models (MPI, threading, GPUs, I/O) are supported.

When running an instrumented application, measurement event data is provided by the instrumentation hooks to the measurement core. There, the events are augmented with high-accuracy timestamps and potentially hardware counters (a plugin-API allows querying additional metric sources). The augmented events are then passed to one or both of the built-in event consumers, profiling and tracing (a plugin-API allows creation of additional event consumers) which finally provide output in the formats CUBE4 and OTF2, respectively. These open and backwards-compatible output formats can be consumed by established analysis tools, e.g., like

• CubeGUI, the performance report explorer for Scalasca and Score-P, a generic tool for displaying a multidimensional performance space,
• Extra-P, an automatic performance-modelling tool that supports the user in the identification of scalability bugs,
• TAU's ParaProf, a portable, scalable performance analysis tool, and PerfExplorer, a framework for parallel performance data mining and knowledge discovery,
• Scalasca Trace Tools, a collection of trace-based performance analysis tools that have been specifically designed for use on large-scale systems featuring hundreds of thousands of CPU cores, automatically identifying potential communication and synchronization bottlenecks and offering guidance in exploring their causes, and
• Vampir, a trace-based framework that enables users to quickly display and analyse arbitrary program behaviour.

Score-P is available under the 3-clause BSD Open Source license.

Version 8.4 is a bugfix release for version 8.0. For features/changes/improvements introduced in the latest version, please see the Changelog file.