Performance analysis and modeling for quantum computing simulation on distributed GPU platforms

Quantum computing holds great promise for accelerating computational tasks, but they are still not accessible. To fill this gap, quantum computing simulators have been widely used for the developing of quantum circuits and algorithms. Simulating quantum algorithms on classical computers also poses challenges due to the need for exponential memory and computational requirements. Many researchers attempted to address such challenges on different single-core, multi-core, and many-core systems, especially graphics processing units (GPUs). The diversity of CPU and GPU simulation of quantum circuits, including various CPU–GPU combinations and multiple parameters, including qubit size, memory capacity, circuit depth, GPU performance, resource heterogeneity, and load imbalance, makes it even more challenging. Finding the best configuration requires an exhaustive search in the design space, which is not possible in an acceptable time frame. Therefore, given the multitude of parameters and the analysis of influential factors, having an analytical model for selecting the proper configuration is desirable and even essential for large systems. This paper proposes a novel analytical performance model for quantum circuit simulation on a hybrid CPU–GPU platform of various sizes and parameters such as number of CPUs/GPUs, qubit size, memory capacity, quantum circuit depth, CPU/GPU performance, resource heterogeneity, and processing load. To do so, we focus on evaluating a scalable and adaptive hybrid quantum simulator in a hybrid platform with some CPUs and GPUs across multiple hosts. The model analyzes the execution time of individual GPU kernels and the impact of major micro-architecture features on performance. By employing dynamic load partitioning (DLP) and the heterogeneous multi-GPU kernel, performance bottlenecks are accurately identified, and execution time is estimated. The proposed model shows high accuracy, with a 94% accuracy compared to the experimental results on a hybrid multi-node cluster. Therefore, the proposed model provides insights into scalability, efficiency, and load balancing in hybrid parallel systems, hence supporting code optimization and development of efficient quantum algorithms and advanced quantum circuit simulation on hybrid parallel architectures.