A Hybrid Task Graph Scheduler for High Performance Image Processing Workflows.

Area of Science:

• Computer Science
• High-Performance Computing
• Parallel Computing

Background:

• Scalability is crucial for performance in hybrid and cluster computing environments.
• Scheduling parallel code presents challenges due to data dependencies, memory management, data motion, and processor occupancy.
• Existing methods often require complex manual optimization for multi-core and multi-GPU systems.

Purpose of the Study:

• To introduce the Hybrid Task Graph Scheduler (HTGS) as an abstract execution model, framework, and API.
• To improve programmer productivity in developing hybrid workflows for multi-core and multi-GPU systems.
• To demonstrate efficient management of tasks, memory, and data movement for enhanced computational performance.

Main Methods:

• Developed HTGS as an abstract execution model and API for hybrid workflows.
• Implemented HTGS to manage task dependencies, independent CPU/GPU memory representations, and overlapping computations with I/O and data transfers.
• Demonstrated HTGS with matrix multiplication and microscopy image stitching algorithms.

Main Results:

• HTGS implementation for image stitching reduced code size by approximately 43% with minimal overhead.
• HTGS-based matrix multiplication achieved 1.3x and 1.8x speedups for large matrices compared to OpenBLAS.
• HTGS effectively keeps multiple GPUs occupied and utilizes all available compute resources.

Conclusions:

• HTGS significantly enhances programmer productivity for hybrid computing workflows.
• The explicit data motion and memory abstractions in HTGS facilitate data locality decisions.
• HTGS provides a performant and efficient solution for scalable applications on modern multi-core and multi-GPU architectures.