Design of GPU-based parallel computation architecture of Thomson scattering diagnostic in KSTAR

Abstract

The Thomson scattering (TS) System is a diagnostic system to measure electron temperature and density profiles of tokamak plasma. The TS system requires measurement of many input signals, and the amount of raw data has significantly increased since the TS data acquisition (DAQ) system was upgraded to a fast digitizer. Research has been done on applying artificial neural network (ANN) to TS data analysis for reducing calculation time. In this paper, we propose a design of computation architecture to effectively process the increased amount of data caused by the fast digitizer and to maximize the computation performance of the ANN. In the design, the intensity values of each input signal and the ANNs can be computed in parallel by utilizing a graphical processing unit (GPU). Furthermore, we integrate the data analysis task into the TS DAQ program for real time operation. Considering stability of the integration, we separate tasks of the data acquisition and the data analysis into each thread operation, and make tasks of each digitizer board to be conducted in parallel. In the feasibility test of the design, the calculation time is shown to be appropriate for real time operation.

Introduction

The Thomson scattering (TS) System is a diagnostic system to measure electron temperature and density profiles of tokamak plasma [1]. As the whole procedure of the diagnostic, a laser beam with a single wave length of light penetrates the tokamak plasma, and is scattered to a wide range of wave-length. The scattered light is collected by lenses, and is transmitted to a polycromator which divides the light in several areas of the wave-length. It also converts the optical signal to the electrical voltage. The TS data acquisition (DAQ) system converts the analog signals into digital signals, which are used in TS data analysis.

The TS system can provide absolute measurements and high resolution spatial profiles; thus, it becomes one of the major diagnostics in most tokamak devices such as JET [2], DIII-D [3], ASDEX [4], JT-60T [5], etc. In KSTAR, the TS system was introduced in 2010 [6], and has been consistently upgraded in terms of the DAQ system [7], [8], [9], the polycromator [10], [11], an alignment device [12], and research of data analysis algorithm [13], [14]. Especially in the TS DAQ system, a fast digitizer with 5 GS/s has been adopted [9]. The previous digitizer in [6] is a charge-to-digital conversion (QDC) system which obtains only one value of integration of the electrical pulse signal. The fast digitizer acquires time series data of the pulse signal; thus, we can analyze the shape of the pulse signal and its noise level, which can be applied to signal processing. However, the amount of the raw data to be processed significantly increases. In the research of the data analysis algorithm, the authors of [13] analyzed the noise effects of the χ-square method which had been used for calculation of the electron temperature profiles. This method requires much computation time because the χ-square should be calculated with the full range of the electron temperature. In [14], an artificial neural network (ANN) was designed to improve the computation time.

In this paper, we propose a computation architecture design based on graphical processing unit (GPU) to effectively process the increased amount of data from the fast digitizer and to maximize the computation performance of the ANN. Usage of the GPU for general purposes has been enabled since NVIDIA developed CUDA™ (Compute Unified Device Architecture). Enhancement of data processing performance using the GPU had been researched in fusion engineering such as fast magnetic field computation [15], real time equilibrium reconstruction [16], 2-D microwave image [17], reflectometer [18], etc. The accelerated calculation speed is mainly due to the parallel computation architecture inherent in the GPU, and the TS diagnostic algorithm can be calculated in parallel in the following two ways: First, the intensity of the pulse signal in the fast digitizer can be effectively calculated by using multi-threads of GPU, which can be conducted in parallel by assigning a GPU block for each input signal. Second, the ANN can be effectively computed by the GPU because most of the computation of the ANN is composed of the matrix operation, and the multi-ANNs can be computed in parallel by assigning a GPU block for each ANN.

A suitable object for this proposed design is aimed at analyzing TS data in real time. In order to achieve this, both the data acquisition and the data analysis should cooperate in the TS DAQ program; thus, we design integration of the data analysis task into the program. In the design, the stability of the integration becomes important because an abnormal status of the data analysis task could cause operational failure of the whole system. Therefore, we separate both the tasks into each thread operation so that the data acquisition task is maintained even if the data analysis task fails. In addition, the DAQ system consists of several digitizer boards, and we make the tasks of each board conduct independently so that the operational status of one board does not affect the others.