TOD-Tree: Task-Overlapped Direct Send Tree Image Compositing for Hybrid MPI Parallelism and GPUs.

Area of Science:

• High-performance computing
• Computer graphics
• Parallel algorithms

Background:

• Supercomputers utilize thousands of nodes with powerful CPUs/GPUs, but network speeds limit performance for time-critical tasks.
• Communication bottlenecks, not computation, hinder interactive visualization and other demanding workloads.
• Existing algorithms often struggle to balance workload and communication efficiency.

Purpose of the Study:

• To develop an image compositing algorithm optimized for communication avoidance and overlapping communication with computation.
• To create a solution applicable to both CPU-only and GPU-accelerated supercomputers.
• To improve performance for time-critical scientific workloads.

Main Methods:

• A three-stage algorithm: parallel direct send, tree compositing, and gather.
• Implementation on CPU-only and GPU-accelerated supercomputers (Stampede, Edison, Piz Daint).
• Comparison with IceT library algorithms (radix-k, binary-swap) using hybrid OpenMP/MPI.
• Development of a GPU-based algorithm using CUDA kernels and GPU Direct RDMA.

Main Results:

• The proposed algorithm demonstrates strong scaling performance, outperforming existing methods like radix-k and binary-swap.
• The GPU-based algorithm achieves performance comparable to CPU-based methods.
• A novel workflow integrating GPU rendering and compositing was introduced.

Conclusions:

• The communication-avoiding image compositing algorithm offers significant performance gains on supercomputers.
• GPU acceleration and direct inter-node GPU communication are effective for high-performance image compositing.
• Integrating rendering and compositing on the GPU presents a promising direction for future visualization workflows.