Accelerator Programming Model Performance Measurements for Conjugate Gradient

Performance measurements from a study comparing CUDA, OpenCL, OpenACC, and OpenMP implementations of the Conjugate Gradient method on NVIDIA Pascal GPUs and Intel Xeon Phi coprocessors. The data originates from research at RWTH Aachen University evaluating asynchronous offloading capabilities for multi-device systems. The archive includes modifications to the liboffload library, binaries, and raw measurement data.

Use Cases

• Benchmarking the performance of CUDA, OpenCL, and OpenACC on NVIDIA GPUs based on the described Conjugate Gradient implementation.
• Comparing directive-based programming models (OpenACC, OpenMP) against lower-level paradigms (CUDA, OpenCL) for performance analysis.
• Evaluating asynchronous kernel execution strategies for keeping multiple accelerator devices busy concurrently.
• Studying the performance portability of the Conjugate Gradient algorithm across NVIDIA GPU and Intel Xeon Phi architectures.

Strengths

• Includes raw measurement data from a controlled performance study.
• Compares multiple major accelerator programming models (CUDA, OpenCL, OpenACC, OpenMP).
• Covers two distinct hardware platforms: NVIDIA Pascal GPUs and Intel Xeon Phi coprocessors.
• Associated software artifacts (liboffload modifications, binaries) are included with commit IDs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

RWTH Aachen University (author Jonas Hahnfeld)

Collection Method

Experimental measurements from running benchmark implementations.