Optimization Of Spectral Calculation On Multi GPU-CPU Hybrid Heterogeneous Platform

Essentially all information about astronomical objects outside the solar system comes through the study of electromagnetic radiation(light)as it reaches us.The observed spectrum contains a multitude of important information about star temperature,age,metal abundance and stellar composition etc.Currently,quite a few tools had been developed to model,calculate and analysis the electromagnetic spectrum in astrophysics,and the widely used ones include the Interactive Spectral Interpretation System(ISIS),the XSPEC,the XSTAR,and the Astrophysical Plasma Emission Code(APEC)and so on.These kits are good ways to solve spectrum calculation,but structure of the program is still in the traditional serial mode.Until now,there is no any spectrum calculation tools which are based on parallel architecture.The core of the spectrum calculation is numerical integration.Meanwhile GPU-based high performance computers have gained popularity in scientific computing as a low cost and powerful parallel architecture in the last decades,and the use of GPUs has proven to significantly increase the performance in numerous applications,including solving large differential equations and high-dimensional numerical integrations.However,the spectral calculation has two distinct characteristics that common GPU-based numerical integration schemes seldom address.One is each single one-dimensional integral computing is very small and fast,but there are huge amounts of small integrations.The other is classical load balancing approaches for CPU-GPU hybrid architecture may be not efficient to schedule so many small tasks due to the extra overhead proportional to the frequency of schedulingIn this paper,we proposed a hybrid CPU-GPU parallel approach to accelerate spectral calculation.First,we offloaded the compute-intensive integral parts of the application to GPUs,and reduced the frequency of memory copy between device and host by combining many single integral operations within one ion into a coarse-grained task.Second,for a large number of small tasks in the spectral calculation,we developed a task scheduling scheme among multiple CPUs and GPUs via share memory that can avoid extra communication overhead in the traditional client-server architecture.Last,comprehensive theoretical analysis and experiments were conducted to verify the efficiency and accuracy of the approach,and the experiments showed that 24 CPU cores with 3 GPU devices can speed up the computation by a factor of 300 relative to the original serial implementation,and a factor of 22 relative to the 24 CPU cores parallel version.Additionally,the approach was also adapted to NEI related application involving numerous ODEs and achieved a 15-fold speedup relative to the pure MPI implementation.