Pipeline SpectroscopyThomas R. Puzak, A. HartsteinP. G. Emma, V. SrinivasanIBM - T. J. Watson Research CenterYorktown Heights, NY 10598

Workshop on Experimental Computer Science

1 ; Arthur Nadas

Others  :  http://www.usenix.org/events/expcs07/papers/15-puzak.pdf
PID  :  6279

来源: CEUR

【 摘 要 】

Pipeline Spectroscopy is a new technique that allows us tomeasure the cost of each cache miss.The cost of a miss isdisplayed (graphed) as a histogram, which represents a precisereadout showing a detailed visualization of the cost of each cachemiss throughout all levels of the memory hierarchy.We call thegraphs ‘spectrograms’ because they reveal certain signaturecharacteristics of the processor’s memory hierarchy, the pipeline,and the miss pattern itself.We show that in a memory hierarchywith N cache levels (L1, L2, ..., LN, and memory) and a misscluster of size C, there areC + N Cpossible miss penalties. This represent all possible sums from all possible combinations ofthe miss latencies from each level of the memory hierarchy (L2,L3, ... Memory) for a given cluster size.Additionally, a theory ispresented that describes the shape of a spectrogram, and we usethis theory to predict the shape of spectrograms for larger missclusters.Detailed analysis of a spectrograph leads to muchgreater insight in pipeline dynamics, including effects due toprefetching, and miss queueing delays.Categories and Subject DescriptorsD.3.3 [Programming Languages]: Design studies, Measurementtechniques, Modeling techniques, Performance attributesGeneral TermsAlgorithms, Measurement, Performance, Theory