基于总变分最小化模型的异步并行GPU加速算法

Asynchronous Parallel GPU Acceleration Method Based on Total Variation Minimization Model

相比于传统同步并行计算策略,在异步并行计算框架下,针对最常用的总变分(TV)最小化重建模型,通过将其转化为不动点迭代问题,并利用异步交替方向法(ADM)进行求解,推导出基于TV最小化模型的异步ADM迭代重建算法,即异步交替方向总变分最小化算法(Async-ADTVM)。利用消息传递接口技术将该算法在图形处理器(GPU)集群上进行测试,进一步提高了原始基于TV最小化模型的迭代重建算法的计算效率。实验表明,该算法在计算求解精度上略优于ADTVM算法,同时在GPU性能存在差异的条件下相比传统多GPU加速策略可获得更高的加速比。

Compared to the traditional synchronous parallel computing,an asynchronous parallel alternating direction method（ADM）for total variation（TV）minimization reconstruction,namely asynchronous alternating direction total variation minimization method（Async-ADTVM）,is proposed in this paper.Under the asynchronous parallel computing framework,Async-ADTVM transforms TV minimization reconstruction model to the problem of fixed-point iteration,which is solved by asynchronous parallel ADM.It is implemented on the graphics processing unit（GPU）cluster based on message passing interface technology.Experimental results show that the proposed Async-ADTVM can provide a little higher calculation accuracy than ADTVM.Meanwhile,it can provide a higher speed-up ratio than the traditional multi-GPU acceleration strategy when the performance of each GPU is different.