The High Performance Computing Studies Based On FPGA

High performance computing(HPC)is useful in the national economy?national defense and scientific research fields such as geological science?life science?finance?chemical engineering?computer aided engineering?mechanical design?meteorology?biotechnology?oil exploration?theoretical physics?astrophysics?However,with the increasing complexity of application field,the gradual enlargement of application scale and the continuous improvement of precision requirement,the challenge of high performance computing is also increasing.The mainstream of high performance computing platforms which adopt the cluster CPU mode has good processing power for the computation-intensive computing tasks with complicated calculation and simple communication.However,for the communication-intensive computing tasks with simple calculation and complicated communication,the the mainstream of high-performance computing platform shows the obvious disadvantages of low computational performance and low computational efficiency.FPGA becomes a new force of high performance computing platform due to its rich parallel computing resources,reconfigurability,high performance,low power consumption and low cost.The FPGA shows superior performance than clustered CPU or GPU in dealing with communication-intensive high-performance applications.However,one needs to be careful on the problem selection,algorithm design,system implementation and testing to achieve the performance advantage of FPGA.In this thesis,the goal is to achieve the ultra high performance of FPGA computing.We design?develop and test the high performance computing tasks on FPGA,focus on the implementation of Ising model simulation system on FPGA.Our Ising model simulation system has a very fast computing speed.We use the developed Ising model simulation system to study the linear relaxation process?the generation and testing of large-scale parallel random number generators and get lots of new results.The major work and contributions of this thesis are as follows:1.A roll-your-own automatic building platform written in Python language is used to generate large Verilog programs from the simple source program files with high efficiency and low error rate.It is also easy to modify?read and maintain the files.The Quartus tcl command is used to realize the data transmission between FPGA and PC which reduces the difficulty of system development.The Modelsim do command is used to realize the automation project simulation in Modelsim software and the Quartus tcl command is used to realize the automatic project construction in Quartus? software,which is simple.fast and easy to operate.2.The MC simulation of Ising model is designed and implemented on FPGA which obtained an amazing speedup.It has a speedup of 7000 X over single CPU,54 X over single GPU.3.The effect of extremely long autocorrelation time on simulation accuracy is studied from extremely long Monte Carlo simulations of extremely large Ising lattices.It is shown that it is better to use the accurate error estimating formula from Kikuchi and Ito than the approximated error estimating formula from Muller-Krumbhaar and Binder to measure deviations of autocorrelation time comparable to simulation time.4.Linear relaxation times are measured from extremely long Monte Carlo simulations of extremely large Ising lattices.It is found that the dynamic FSS theory of two-dimension Ising lattices is still accurate for correlation lengths as long as 2048.The dynamic exponent z is determined to be 2.179(12),and the result result is compared and verified with the results from other methods,including nonequilibrium and stochastic matrix critical dynamics methods.5.It is also found that the statistical correlations between pseudorandom numbers can seriously affect the extremely long Monte Carlo simulations of critical dynamic in Ising lattices size larger than 512.This effect makes the investigation of critical dynamics in larger Ising lattices even more difficult,at the same time provides a sensitive method to evaluate the quality of the random number generator.6.We implement and test the seven kinds of RNGs for large-scale parallel Monte Carlo(LPMC)simulations on FPGA.A new parallel RNG test is proposed.Use critical dynamic to test the pseudorandom numbers.One RNG is considered as solid candidate for large-scale parallel Monte Carlo simulations.