运行时缓存缺失率曲线生成与共享缓存划分

发布时间：2018-04-05 18:44

本文选题：共享缓存　切入点：缓存缺失率曲线　出处：《华中科技大学》2012年硕士论文

【摘要】：共享缓存结构加速了核与核之间的通讯速度，在多核处理器中有着重要作用。然而，多个核竞争使用共享缓存，互相污染对方的缓存数据，降低了系统的整体性能。为了解决这个问题，研究者提出了基于缓存缺失率曲线划分缓存的方法划分共享缓存，从而消除共享缓存竞争。但是，运行时生成缺失率曲线需要特殊的硬件支持，而用纯软件的方法在运行时生成缺失率曲线开销较高。针对上述问题，提出了一种纯软件、低开销和高准确度的缓存缺失率曲线生成方法，基于生成的缺失率曲线划分缓存。该方法根据程序访问共享缓存的行为分形特点，基于分形模型在运行时预测程序的缓存缺失率曲线，降低了曲线生成的开销；利用现代处理器上普遍存在的硬件计数器单元记录程序的运行时信息，极大地减少了记录运行时信息的开销，并且不需要修改应用程序，也不需添加额外的硬件；利用运行时信息计算预测模型的参数，，将参数代入分形预测模型计算程序的缓存缺失率曲线，该模型简单，计算量小；基于生成的缺失率曲线有效地指导操作系统制定缓存划分策略，在操作系统层实现了共享缓存划分，提升了系统的整体性能。为了验证该预测模型的准确度与该预测方法的开销，选用了SPEC CPU2006基准测试集中的26个程序进行了测试。预测的缺失率曲线与真实的缺失率曲线之间的平均误差为3.13%。使用该预测方法生成缺失率曲线所需的时间为420ms。基于生成的缺失率曲线划分共享缓存后，系统的平均IPC最高提升了20%。
[Abstract]:Shared buffer architecture speeds up the communication between cores and plays an important role in multi-core processors.However, multiple core competition uses shared cache, which pollutes each other's cache data and reduces the overall performance of the system.In order to solve this problem, the researchers proposed a method to partition the shared cache based on the cache missing rate curve, so as to eliminate the competition of the shared cache.However, it needs special hardware support to generate the missing rate curve at run time, and it is expensive to generate the missing rate curve at run time by using pure software.In order to solve the above problems, a pure software, low overhead and high accuracy method for generating cache loss rate curves is proposed, which is based on the generated loss rate curves to partition the cache.According to the behavior fractal characteristic of program accessing shared cache, this method predicts the missing rate curve of program at run time based on fractal model, and reduces the cost of curve generation.The runtime information of the program is recorded by using the common hardware counter unit on the modern processor, which greatly reduces the overhead of recording the run-time information, and does not need to modify the application program or add additional hardware.The parameters of the prediction model are calculated by using the run-time information, and the parameters are substituted into the cache missing rate curve of the calculation program of the fractal prediction model. The model is simple and the calculation amount is small.Based on the generated loss rate curve, the cache partition strategy is effectively directed to the operating system, and the shared cache partition is implemented in the operating system layer, which improves the overall performance of the system.In order to verify the accuracy of the prediction model and the cost of the prediction method, 26 programs in the SPEC CPU2006 benchmark set were selected for testing.The average error between the predicted loss rate curve and the true loss rate curve is 3.13.The time required to generate the defect rate curve using this prediction method is 420 Ms.After the shared cache is partitioned based on the generated loss rate curve, the average IPC of the system increases by 20%.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333

【参考文献】