Pseudorapidity distribution and decorrelation of anisotropic flow within the open-computing-language implementation CLVisc hydrodynamics

Studies of fluctuations and correlations of soft hadrons and hard and electromagnetic probes of the dense and strongly interacting medium require event-by-event hydrodynamic simulations of high-energy heavy-ion collisions that are computing intensive. We develop a (3+1)-dimensional viscous hydrodynamic model-CLVisc that is parallelized on a graphics processing unit (GPU) by using the open computing language (OpenCL) with 60 times performance increase for spacetime evolution and more than 120 times for the Cooper-Frye particlization relative to that without GPU parallelization. The model is validated with comparisons with different analytic solutions, other existing numerical solutions of hydrodynamics, and experimental data on hadron spectra in high-energy heavy-ion collisions. The pseudorapidity dependence of anisotropic flow v(n)(eta) are then computed in CLVisc with initial conditions given by the a multiphase transport (ampt) model, with energy density fluctuations both in the transverse plane and along the longitudinal direction. Although the magnitude of v(n)(eta) and the ratios between v(2)(eta) and v(3)(eta) are sensitive to the effective shear viscosity over entropy density ratio eta(v)/s, the shape of the v(n)(eta) distributions in eta do not depend on the value of eta(v)/s. The decorrelation of v(n) along the pseudorapidity direction due to the twist and fluctuation of the event planes in the initial parton density distributions is also studied. The decorrelation observable r(n)(eta(a),eta(b)) between v(n){-eta(a)} and v(n){eta(a)} with the auxiliary reference window eta(b) is found not to be sensitive to eta(v)/s when there is no initial fluid velocity. For small eta(v)/s, the initial fluid velocity from mini-jet partons introduces sizable splitting of r(n)(eta(a),eta(b)) between the two reference rapidity windows eta(b)is an element of[3,4] and eta(b)is an element of[4.4,5.0], as has been observed in experiment. The implementation of CLVisc and guidelines on how to efficiently parallelize scientific programs on GPUs are also provided.