摘要
在评价多核CPU加速比已知模型的基础上,基于第一性计算原理融合理解阿姆达尔定律和兰特法则,提出描述多核CPU加速比的一个新模型.研究方法是从传统的阿姆达尔定律切入,论述的逻辑顺序分别基于约束固定任务,固定时间,存储器和互连复杂性;兼顾了举例论述同构多核的NoC带宽性质和最大温度特性.计算表明:基于固定时间模型与存储器模型预测多核的加速能力,容易得到估计结果的乐观上限;我们提出的基于兰特法则的模型计算结果,在并行比例较大时稍小于但接近前述模型估计值,而比固定任务模型的保守结果要好;NoC带宽和最大温度的结果提示,多(同构)核CPU期盼相对高的并行度架构.
We made Amdahl' s Law and Rent's Rule understood and proposed one novel model to describe the multicore CPU speedup based on evaluating the known speedup models. Our methodology starts with traditional Amdahl' s Law, in the steps of fixed-size,fixed-fime,memory-bounded, and interconnection-bounded. Also the NoC-bandwidth and max-temperature of multicore features ale discussed. The results showed, fixed-time model and memory-bounded model grasp speedups' up-limits, our model touches the middle gap between formers and Amdahl' s Law, at high parallel ratio, and NoC bandwidth and Max temperature show that homogeneous multicore architectures long for higher ratio of parallelization.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2012年第2期230-234,共5页
Acta Electronica Sinica
基金
江苏省高校自然科学基金(No.09KJB510017)