| 韩晨吉* **,薛峰* **,吴钰轩***,汪文祥***,张福新*.处理器硅前性能评估仿真点全局贪心分配方法[J].高技术通讯(中文),2026,36(1):29~40 |
| 处理器硅前性能评估仿真点全局贪心分配方法 |
| Global greedy allocation for simulation points in processor pre-silicon performance evaluation |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 01. 003 |
| 中文关键词: 处理器硅前性能评估； 代表性采样； 程序微架构无关特征 |
| 英文关键词: processor pre-silicon performance evaluation, representative sampling, microarchitecture-independent characteristics |
| 基金项目: |
| 作者 | 单位 | | 韩晨吉* ** | （*处理器芯片全国重点实验室（中国科学院计算技术研究所）北京 100190）
（**中国科学院大学北京 100049）
（***龙芯中科技术有限公司北京 100190） | | 薛峰* ** | | | 吴钰轩*** | | | 汪文祥*** | | | 张福新* | |
|
| 摘要点击次数: 24 |
| 全文下载次数: 15 |
| 中文摘要: |
| 仿真点(simulation point，SimPoint)作为一种代表性采样技术被广泛应用于处理器硅前性能评估中。SimPoint为每个待评估的程序根据贝叶斯信息准则确定仿真点数目。然而，标准测试集内不同程序有着不同的行为复杂程度，需要不同数目的仿真点来准确刻画其程序行为。SimPoint无法识别出不同程序间的复杂度差异，无法做到在总仿真点数目一定的情况下，将更多的仿真点分配给行为复杂的程序以降低这些程序的性能评估误差，将更少的仿真点分配给行为简单的程序而不损失这些程序的性能评估精度。由于没有在测试集内合理地进行仿真点分配，SimPoint虽然可以给出比较准确的平均性能评估误差，但是某些行为复杂的测试子项的性能评估误差依然较大。针对这一问题，本文优化了SimPoint的仿真点局部分配方式，提出了一种全局贪心分配方法——贪心点(greedy point，GreedyPoint)方法。该方法将仿真点的分配问题抽象为含约束的优化问题，使用微架构无关特征计算表征误差，通过全局贪心算法来求解该优化问题。实验数据表明，在相同仿真开销下，与SimPoint相比，GreedyPoint可以将SPEC CPU 2017测试套件的平均性能评估误差由3.23%降低到2.08%，最大性能评估误差由21.22%大幅降低至7.01%。 |
| 英文摘要: |
| SimPoint, as a representative sampling technique, is widely employed in pre-silicon performance evaluation of processors. SimPoint determines the number of simulation points for each program based on the Bayesian information criterion, yet it fails to recognize the complexity variations among different programs. However, within a standard test suite, different programs exhibit varying levels of behavioral complexity, necessitating different numbers of simulation points to accurately characterize their behaviors. Given a fixed total number of simulation points, SimPoint is incapable of allocating more simulation points to more complex programs to reduce their performance evaluation errors while assigning fewer simulation points to simpler programs without compromising their performance evaluation accuracy. Due to the lack of a global view of the test suite, although SimPoint can provide relatively accurate average performance evaluation errors, certain behaviorally complex programs still exhibit significant performance evaluation errors. To address this issue, this paper optimizes the local allocation of simulation points within the SimPoint method and proposes a global greedy allocation method, GreedyPoint. This method abstracts the simulation point allocation problem as a constrained optimization problem, utilizes microarchitecture independent characteristics to quantify represent errors, and employs a global greedy algorithm to solve the optimization problem. Experimental data indicate that, under the same simulation overhead, compared to SimPoint, GreedyPoint reduces the average performance evaluation error of the SPEC CPU 2017 test suite from 3.23% to 2.08%, and significantly decreases the maximum performance evaluation error from 21.22% to 7.01%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
