A pattern fusion model for multi-step-ahead CPU load prediction Systems and Software Dingyu Yanga, Jian Caoa,∗, Jiwen Fua, Jie Wangb, Jianmei Guoc ιΎθη½ 1 Introduction • Resource availability often changes from time to time and schedulers needs a global view of all resources in a distributed system. • Sometimes high CPU load means performance anomalies that may result in system collapse. 2 WNN and PSF • Weighted Nearest Neighbors (WNN) algorithm • Pattern Sequence-based Forecasting (PSF) algorithm • Euclidean distance – not enough to match the similar patterns. 3 WNN and PSF (cont.) • One-step-ahead prediction strategy – if the one-step-ahead value is not accurate, the predictions for the following points in time will become more and more inaccurate. 4 A pattern fusion model for multi-stepahead CPU load prediction 5 Pattern extraction 6 Pattern extraction (cont.) 7 Pattern extraction (cont.) CPU load 9 8 7 6 5 CPU load 4 3 2 1 0 0 1 2 3 4 5 6 7 8 8 Pattern extraction (cont.) • Pattern filtering ππ’ππππ(ππππ‘ππ(ππ )) ≤1−πΌ ππ’ππππ(ππ ) ππππ‘ππ ππ : π ππ’πππ‘πππ π‘βπ πππ‘π’πππ π πππ‘π‘πππ π π’ππ ππ‘ π€ππ‘β βππβππ ππππ’ππππππ πππππ’ππππ¦ ππ’ππππ ππ : πππ‘π‘πππ ππ’ππππ ππ’ππππ ππππ‘ππ ππ : ππ’ππππ ππ πππππππππ πππ‘π‘ππππ • Pattern merging ↑↑↑↑↑↑↑↑ ↑↑↑∗↑↑↑↑ ↑↑↑↓↑↑↑↑ 9 Pattern matching 10 Pattern matching (cont.) • Pattern similarity matching – Euclidean distance matching πππ π‘ π, π = ππ − ππ πβπ π‘ππ π ππππππ π‘ πππ‘π‘ππππ πππ ππ πβππ ππ – Fluctuation pattern set matching ππ′ − ππ′ π ≥1− πππππ‘β(ππ′ ) 11 Pattern weighting strategy • Average rule strategy 12 Pattern weighting strategy (cont.) • Uniform decline strategy ππ = 1 π π π ππ′ = π π π=1 ππ 13 Pattern weighting strategy (cont.) 14 Pattern weighting strategy (cont.) 15 Prediction results function • Adaboost algorithm π‘ π€ π 1. π ππ‘ πππ‘ = π π‘ π=1 π€π 2. ππππ‘ = π π‘ π=1 ππ ∗ πΌ ππ ≠ βπ π₯π π π‘ π=1 ππ πΌ= 1 ππ ≠ βπ π₯π 0 ππ = βπ π₯π 1 1 − ππππ‘ 3. πΌπ‘ = ln 2 ππππ‘ 4. πππ‘ = πππ‘ ∗ ππ₯π πΌπ‘ ∗ πΌ ππ ≠ βπ π₯π 16 Prediction results function (cont.) • Prediction values from different values of various length patterns: 17 Experiment • Experiment settings – Mean absolute error (MAE)= 1 π π π=1 – Mean relative error (MRE)= 100 ∗ 1 π π₯π − π₯ π π₯π −π₯ π=1 π₯ π 18 Experiment (cont.) 19 Experiment (cont.) • Pattern similarity measurement • The influences of filter parameter ππ’ππππ(ππππ‘ππ(ππ )) ≤1−πΌ ππ’ππππ(ππ ) 20 Experiment (cont.) • Different pattern weighting strategies • Multi-step-ahead predictions 21 Experiment (cont.) • Comparison with iterative one-step-ahead prediction 22 Experiment (cont.) • Comparisons with other prediction approaches • Run time 23 Conclusion • The contribution of this paper is a fusion model for multi-step-ahead CPU load prediction. • We plan to consider longer steps and improve the prediction accuracy in future. • In addition, multivariate prediction, such as memory usage and disk usage predictions are also interesting topics. 24