| 引用本文: | 裴颂文,张俊格,宁静.梯度学习的参数控制帮助线程预取模型.[J].国防科技大学学报,2016,38(5):59-63.[点击复制] |
| PEI Songwen,ZHANG Junge,NING Jing.Helper thread pre-fetching model based on learning gradients of control parameters[J].Journal of National University of Defense Technology,2016,38(5):59-63[点击复制] |
|
| |
|
|
| 本文已被：浏览 7634次 下载 6344次 |
| 梯度学习的参数控制帮助线程预取模型 |
| 裴颂文1,2, 张俊格1, 宁静1 |
|
(1.上海理工大学 光电信息与计算机工程学院， 上海 200093;2.
2.上海理工大学 上海市现代光学系统重点实验室， 上海 200093)
|
| 摘要: |
| 对于非规则访存的应用程序，当某个应用程序的访存开销大于计算开销时，传统帮助线程的访存开销会高于主线程的计算开销，从而导致帮助线程落后于主线程。于是提出一种改进的基于参数控制的帮助线程预取模型，该模型采用梯度下降算法对控制参数求解最优值，从而有效地控制帮助线程与主线程的访存任务量，使帮助线程领先于主线程。实验结果表明，基于参数选择的线程预取模型能获得1.1～1.5倍的系统性能加速比。 |
| 关键词: 数据预取 帮助线程 多核系统 访存延迟 梯度下降 |
| DOI：10.11887/j.cn.201605010 |
| 投稿日期：2015-11-16 |
| 基金项目:上海市自然科学基金资助项目（15ZR1428600）；计算机体系结构国家重点实验室开放资助项目（CARCH201206）；上海市浦江人才计划资助项目(16PJ1407600) |
|
| Helper thread pre-fetching model based on learning gradients of control parameters |
| PEI Songwen1,2, ZHANG Junge1, NING Jing1 |
|
(1. School of Optical Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China;2.
2. Shanghai Key Laboratory of Modern Optical System, University of Shanghai for Science and Technology, Shanghai 200093, China)
|
| Abstract: |
| To the applications with irregular accessing memory, if the overhead of accessing memory for a given application is much greater than that of computation, it will make the helper thread lag behind the main thread. Hereby, an improved helper thread pre fetching model by adding control parameters was proposed. The gradient descent algorithm is one of the most popular machine learning algorithms, which was adopted to determine the optimal control parameters. The amount of the memory access tasks was controlled by the control parameters effectively, which makes the helper thread be finished ahead of the main thread. The experiment results show that the speedup of system performance is achieved by 1.1 times to 1.5 times. |
| Keywords: data pre-fetch helper thread multi-core system memory latency gradient descent |
|
|
