OpenMP5 Offloaded Quantum-Inspired Evolutionary Optimization Performance Across GPU Ecosystem

Overview

A recent investigation explored the viability of a single OpenMP 5 source for Quantum-Inspired Evolutionary Optimization (QIEO) when offloaded to diverse GPU platforms. The core inquiry focused on whether such an implementation represents a practical production pathway across various computational settings, ranging from laboratory servers to rented cloud workstations and leadership-class accelerators. This evaluation specifically assessed the portability and performance of QIEO when utilizing the #pragma omp target directive for offloading.

Quantum-inspired evolutionary optimization (QIEO) is characterized as a population-based metaheuristic algorithm. It functions by representing design variables as a collection of qubits and navigates a continuous, multi-dimensional landscape by rotating the amplitude pairs of these qubits. The optimization process involves rotating amplitudes towards a single elite individual in each generation, corresponding to the best solution identified within that generation. The computational cost per generation for QIEO scales as $O(N_p N_g)$, where $N_p$ denotes the number of chromosomes and $N_g$ represents the number of genes (decision variables).

Research Context

Production applications of computational solvers, such as QIEO, frequently transition across different machine classes. Initial development and prototyping often occur on laboratory servers. More extensive computational campaigns typically move to rented cloud workstations. The most demanding problems are typically allocated to leadership-class accelerator systems. The research specifically addressed the question of whether a unified OpenMP 5 source, leveraging #pragma omp target for offloading, could serve as a consistent and effective production approach across this spectrum of hardware environments.

Approach

The study executed three independent campaigns focusing on the 0/1 knapsack problem. Performance was benchmarked against a multi-core Intel CPU baseline, utilizing the same source code. The investigation encompassed approximately 3,000 individual runs, systematically varying both chromosome and gene counts. Two distinct offload strategies were evaluated: chromosome-level and gene-level offload.

The hardware platforms utilized for the GPU evaluations included the NVIDIA Tesla V100 SXM2, NVIDIA A100 80GB, and AMD Instinct MI300X GPUs. To ensure high performance on these diverse platforms, specific deployment nuances were addressed. These included considerations for Volta's constant-memory cliffs, Ampere's L2 persistence and cp.async functionality, and CDNA 3's Infinity Cache and XCD occupancy characteristics.

Findings

• Gene-parallel offload strategy yielded notable performance improvements across all tested GPUs.
• Geometric-mean speedups for gene-parallel offload over a single CPU core were observed as follows:
  • 90$\times$ on the NVIDIA Tesla V100 SXM2.
  • 136$\times$ on the NVIDIA A100 80GB.
  • 155$\times$ on the AMD Instinct MI300X.
• When compared to 72 host threads, the gene-parallel offload strategy demonstrated further speedups:
  • 12$\times$ on the NVIDIA Tesla V100 SXM2.
  • 17$\times$ on the NVIDIA A100 80GB.
  • 16.6$\times$ on the AMD Instinct MI300X.
• The DetermineElite function, which involves the $O(N_p)$ selection of the generation's best chromosome, was found to be better suited for execution on the host CPU rather than the GPU device.

Why This Matters

The findings indicate that a single OpenMP 5 source for Quantum-Inspired Evolutionary Optimization, when effectively offloaded using #pragma omp target, presents a viable production pathway across varied GPU architectures. This suggests potential for optimizing the deployment and performance of QIEO algorithms in environments ranging from laboratory servers to cloud and leadership-class computing systems, provided that specific hardware characteristics and function allocation (host vs. device) are considered.