HybriD-GM: parallel model for quantum computing targeted to hybrid architectures

Abstract

A huge amount of scientific research and technological developments still depends on quantum computing simulations, since quantum computers still limited by physical implementations. Currently, the development of quantum computing algorithms has been carried out by analytic or simulation procedures while quantum computers are not widely available. Although quantum computing simulation is parallel by nature, spatial and temporal complexity are major performance hazards, as quantum states and quantum transformations increase exponentially with the number of qubits simulated. This proposal contributes from the conception to the consolidation of the HybriD-GM model, as well as introduces the extension of D-GM environment, providing efficient parallel executions for quantum computing, which in this work is targeted to hybrid architectures, considering both CPU and GPU architectures. The HybriD-GM model explores the potentialities of High Performance Computing, providing functionalities and exploring projection operators which are acting on quantum structures, state and transformations, to manipulate the granularity and distribution of computations. In this context, the distribution of computations are based on tree data-structures, where intermediate and final nodes corresponding to projection and execution layers are configured to optimize the hardware resources. Simulations of Shor’s and Grover’s algorithms were performed in order to evaluate the HybriD-GM model, and the results achieved significant performance improvements for executions in both CPU and GPU. When compared to the D-GM previous version, they presented speedups of 21 and 9:5 for parallel CPU simulations and of 61:9 and 38:32 for GPU simulations. In addition, related to LIQUiji and ProjectQ simulators, the parallel simulation with 23 qubits was 4:64 faster for Shor’s and 32 faster for Grover’s. Results for hybrid simulations showed that is possible to increase performance for some classes of algorithms, improving the performance of Grover’s algorithm up to 3:18 in comparison to only using GPU approach.