The Tau Parallel Performance System

Abstract

The ability of performance technology to keep pace with the growing complexity of parallel and distributed systems depends on robust performance frameworks that can at once provide system-specific performance capabilities and support high-level performance problem solving. Flexibility and portability in empirical methods and processes are influenced primarily by the strategies available for instrumentation and measurement, and how effectively they are integrated and composed. This paper presents the TAU (Tuning and Analysis Utilities) parallel performance system and describe how it addresses diverse requirements for performance observation and analysis.

Performance Analysis Tool for HPC and Big Data Applications on Scientific Clusters

“…However, such information is generally about the whole job, and more finegrained information would be helpful to understand the individual steps of a large parallel workflow. Alternatively, the workflow management system could record the performance information of each step of a workflow [24], a profiler may be used to automatically capture detailed performance information [31], or the user may instrument selected operations with some library functions [34]. In these cases, the performance data is typically captured into log files.…”

“…These performance tools lack distributing and parallelizing the computations of the analysis to large number of machines. Some tools such as Tau [31] and Vampir [9] can parallelize computational loads MPI processes, and potentially these MPI processes can be extended to distribute multiple loads. However, this extension involves significant implementation challenges due to synchronization and inter-process communication complexities and lack of fault tolerance support.…”

“…Shende et al [31] designed Tau to support monitoring parallel applications by automatically inserting instrumentation routines. Böhme et al [7] presented an automatic mechanism which performs instrumentation during compilation in order to identify the causes of waiting periods for MPI applications.…”

A massively parallel tensor contraction framework for coupled-cluster computations

“…Outside of special network topology considerations on BlueGene/Q, CTF does not employ any optimizations which are specific to an architecture or an instruction set. Performance profiling is done by hand and with TAU [48].…”

Model-Driven Approach for Supporting the Mapping of Parallel Algorithms to Parallel Computing Platforms

“…These tools very often support the optimization of the mapping process by tracing or profiling the applications during run-time. Examples of these tools are TAU [24], HPC Toolkit [25], Open|Speedshop [26], and Scalasca [27]. Since the optimization is done at run-time these tools usually require that the implementation of the algorithm is completed before the analysis.…”