Multi-GPU acceleration of a DGTD method for modeling human exposure to electromagnetic waves
Multi-GPU acceleration of a DGTD method for modeling human exposure to electromagnetic waves
复制标题
DOI：
--
复制
发表时间：
2011-04
期刊：
影响因子：
--
通讯作者：
Tristan Cabel;Joseph Charles;S. Lanteri
中科院分区：
文献类型：
--
作者：
Tristan Cabel;Joseph Charles;S. Lanteri
我们提出了一种高性能的计算方法，用于模拟电磁波在生物组织中的传播，并将其应用于头部组织中的射频吸收的数值评估，因为它们暴露于来自手机的辐射。为此目的，时域麦克斯韦方程组的系统在空间中离散化的不连续伽辽金法，其制定在四面体网格上，并依赖于一个网格单元内的电磁场分量的高阶插值。半离散方程，然后时间积分的二阶蛙跳格式。所得到的数值方法是适应现代并行计算系统与多个GPU加速卡，通过采用混合策略，结合了粗粒度SPMD编程模型的GPU间并行化和细粒度SIMD编程模型的GPU内并行化。通过使用从医学图像构建的头部组织的真实异构模型在GPU集群上执行的大规模模拟，证明了多GPU加速带来的性能提升。
We present a high performance computing methodology for the simulation of electromagnetic wave propagation in biological tissues and its application to the numerical evaluation of radio frequency absorption in head tissues as they are exposed to radiation from a cellular phone. For this purpose, the system of time-domain Maxwell equations is discretized in space by a discontinuous Galerkin method which is formulated on a tetrahedral mesh and which relies on a high order interpolation of the electromagnetic field components within a mesh element. The semi-discretized equations are then time integrated by a second order leap-frog scheme. The resulting numerical methodology is adapted to modern parallel computing systems with multiple GPU acceleration cards by adopting a hybrid strategy that combines a coarse grain SPMD programming model for inter-GPU parallelization and a fine grain SIMD programming model for intra-GPU parallelization. The performance improvement thanks to multiple-GPU acceleration is demonstrated through large-scale simulations that are performed on a cluster of GPUs using realistic heterogeneous models of head tissues built from medical images.