SCALING GYSELA CODE BEYOND 32K-CORES ON BLUE GENE/Q ;
SCALING GYSELA CODE BEYOND 32K-CORES ON BLUE GENE/Q ;
复制标题
在 BLUE GENE/Q 上将 GYSELA 代码扩展至超过 32K 核心；
DOI：
10.1051/proc/201343008
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
O. Thomine
中科院分区：
文献类型：
--
作者：
Julien Bigot;V. Grandgirard;G. Latu;C. Passeron;F. Rozar;O. Thomine
回旋运动模拟导致巨大的计算需求。到目前为止，半拉格朗日代码 Gysela 使用数千个核心（通常为 8k 核心）执行大型模拟。更高分辨率和动电子的模拟预计将大大增加这些需求，为需要百亿亿次机器的应用提供了一个很好的例子。本文介绍了我们改进 Gysela 的工作，以期找到一种架构，该架构提供了实现百亿亿级计算的一种可能途径：Blue Gene/Q。在分析了该架构上代码的局限性后，我们实现了三种改进：计算性能改进、内存消耗改进和磁盘 I/O 改进。结果，我们表明代码现在可以扩展到超过 32k 内核，并且性能大大提高。这将使得能够瞄准最强大的可用机器，从而处理更大的物理案例。
Gyrokinetic simulations lead to huge computational needs. Up to now, the semi- Lagrangian code Gysela performed large simulations using a few thousands cores (8k cores typically). Simulation with finer resolutions and with kinetic electrons are expected to increase those needs by a huge factor, providing a good example of applications requiring Exascale machines. This paper presents our work to improve Gysela in order to target an architecture that presents one possible way towards Exascale: the Blue Gene/Q. After analyzing the limitations of the code on this architecture, we have implemented three kinds of improvement: computational performance improvements, memory consumption improvements and disk i/o improvements. As a result, we show that the code now scales beyond 32k cores with much improved performances. This will make it possible to target the most powerful machines available and thus handle much larger physical cases.