PPubMed10 Apr 2015Cited 10×

Characterizing Task-Based OpenMP Programs Characterizing Task-Based OpenMP Programs

Ananya Muddukrishna, Peter A. Jonsson, Mats Brorsson, Vince Grolmusz

Abstract

Programmers struggle to understand performance of task-based OpenMP programs since profiling tools only report thread-based performance. Performance tuning also requires task-based performance in order to balance per-task memory hierarchy utilization against exposed task parallelism. We provide a cost-effective method to extract detailed task-based performance information from OpenMP programs. We demonstrate the utility of our method by quickly diagnosing performance problems and characterizing exposed task parallelism and per-task instruction profiles of benchmarks in the widely-used Barcelona OpenMP Tasks Suite. Programmers can tune performance faster and understand performance tradeoffs more effectively than existing tools by using our method to characterize task-based performance.

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

A low-overhead method extracts detailed task‑based performance metrics from OpenMP programs, enabling faster tuning and deeper insight than thread‑only tools.

What’s new

i.Introduced a cost‑effective method to extract detailed task‑based performance information from OpenMP programs.
ii.Implemented a stack‑based task tracking and dynamic binary instrumentation approach to capture per‑task instructions and hardware counters.
iii.Provided a comprehensive task‑graph and per‑task property characterization of the Barcelona OpenMP Tasks Suite (BOTS) benchmarks.
iv.Demonstrated the method’s utility by diagnosing performance issues in benchmarks such as Fibonacci and Sort.
v.Performed input‑sensitivity experiments and PCA‑based similarity analysis across BOTS benchmarks.

How they did it

The authors disabled inlining of task outline functions, used a contextual stack to track task execution, and collected per‑task instruction counts via dynamic binary instrumentation together with hardware performance counters at task boundaries. Post‑processing reconstructed the fork‑join task graph, computed exposed parallelism, and derived computational intensity. Profiling was performed in parallel for low‑overhead metrics and sequentially for architecture‑independent metrics.

What they found

Task‑graph profiling overhead less than 1% of execution time.
Reading cycle and cache‑stall counters adds an average 2.5% overhead for BOTS benchmarks.
Architecture‑independent per‑task profiling slows execution by 36× for BOTS benchmarks.
For BOTS Fibonacci on a 48‑core machine, a depth cutoff of 12 gave the best speedup.
Input‑sensitivity runs executed for 1, 5, and 10 seconds on a 24‑core AMD Opteron 6172 machine.
The first four principal components explain more than 85% of variation in the BOTS property dataset.

Caveats

✕The tool does not track implicit tasks, untied tasks, or rogue tasks that outlive their parent.
✕High overhead (36× slowdown) for architecture‑independent per‑task properties due to sequential profiling and binary instrumentation.
✕Untied task support is not implemented and is planned for future versions.
✕The method relies on disabling inlining, which may not be feasible for all compilation environments.

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PPubMed3 found

BbioRxiv2 found

CchemRxiv2 found

Searching the other archives…