On the analysis of computer performance data
Jack C M Wang, John M. Gary, Hari K Iyer · 1989
This paper is devoted to an analysis of the data from the Livermore loops benchmark.We will show that in a general predictive sense the dimension of this data is rather samll; perhaps between two and five.Two techniques are used to reduce the 72 loops timings for each machine to a few scores which characterize the machine.The first is based on a principal component analysis, the second on a cluster analysis of the loops.The validity of the reduction of the data to a lesser dimension is checked by various methods.