Ge Gene nerative and nd Mul ulti-phase se Learning fo for Computer Syst stems s Optimization Ding , Nikita Mishra, Henry Hoffmann Yi Yi Di
Co Compu puter Syst ystems Op Opti timizati tion on • Optimizing modern computer systems requires tradeoffs: Deliver reliable performance Minimize energy consumption 1
Co Compu puter Syst ystems Op Opti timizati tion on • Optimizing modern computer systems requires tradeoffs: Deliver reliable performance Minimize energy consumption • Resource management via system configuration: Resources have complex, non-linear effects on performance and energy Resource interactions create local optima 2
Co Compu puter Syst ystems Op Opti timizati tion on • Optimizing modern computer systems requires tradeoffs: Deliver reliable performance Minimize energy consumption • Resource management via system configuration: Resources have complex, non-linear effects on performance and energy Resource interactions create local optima • How to find the optimal system configuration? 3
Ex Example ample of f a a Config nfigur urat atio ion n Spac ace 𝐃 2.26 Hz Clock Speed Memory Memory Cores Controller 1 Controller 2 Memory controller 4
Mac Machine hine Le Learning to to the the Re Rescue 5
Mac Machine hine Lear earning ning to to the the Re Rescue • However… ta : expensive collection, limited range behavior Sc Scarce da data 6
Mac Machine hine Le Learning to to the the Re Rescue • However … Sc Scarce da data ta : expensive collection, limited range behavior ier useful As Asym ymmetri ric be benefits : only configs on op optim imal l fron ontie 7
Machine Mac hine Le Learning to to the the Re Rescue • However … Sc Scarce da data ta : expensive collection, limited range behavior à Generative model ier useful As Asym ymmetri ric be benefits : only configs on op optim imal l fron ontie à Multi-phase sampling 8
Mac Machine hine Le Learning to to the the Re Rescue • However … Sc Scarce da data ta : expensive collection, limited range behavior à Generative model ier useful Asym As ymmetri ric be benefits : only configs on op optim imal l fron ontie à Multi-phase sampling We We advocate: • De De-em empha hasiz izing ing predic edictio tion accu accuracy acy • In Incorp orpora orating s system s stru ructure i into o lea learner ner 9
Pr Problem blem Formulat mulatio ion Meet Meet la laten ency con onstra rain ints wit with min minima imal l en ener ergy via via sy syst stem co configs Co Confi fig: : an an al allocat ation of of ha hardware resourc re rces to an an ap applicat ation Re Require th the po power an and pe perfo formance pr profi file fo for ap applicat ations Expensiv Ex ive Le Learn to estimate these values 10
SR SRAD AD on AR ARM bi big.LITTLE syst system em Model A Model B Optimal points Just far enough True data Non-optimal points True data Very far Goodness of fit 99% 0 22% ❌ 0 ✅ Energy over optimal 11
SR SRAD AD on AR ARM bi big.LITTLE syst system em Key Insight: Ke High accuracy ≠ good syst Hi good system re results Model A Model B Optimal points Just far enough True data Non-optimal points True data Just far enough Goodness of fit 99% 0 22% ❌ 0 ✅ Energy over optimal 12
Re Recommender Systems -> > Le Learning by Examples https://www.muvi.com/blogs/deciphering-the-unstoppable-netflix-and-the-role-of-big-data.html https://datajobs.com/data-science-repo/Recommender-Systems-[Netflix].pdf 13 1. Paragon: QoS-Aware Scheduling for Heterogeneous Datacenters. Christina Delimitrou and Christos Kozyrakis. (ASPLOS 2013) 2. Quasar: Resource-Efficient and QoS-Aware Cluster Management. Christina Delimitrou and Christos Kozyrakis (ASPLOS 2014)
An An An Analo alogy Users ? 1 3 5 5 4 5 4 4 2 1 2 2 4 1 2 3 4 3 5 2 4 5 Movies 4 2 4 2 4 1 2 4 1 3 2 3 3 4 2 2 5 3 1 3 3 2 4 2 Ratings between 1 to 5 Unknown ratings 14
An An An Analo alogy Users ? 1 3 5 5 4 5 4 4 2 1 2 2 4 1 2 3 4 3 5 2 4 5 Movies 4 2 4 2 4 1 2 4 1 3 2 3 3 4 2 2 5 3 1 3 3 2 4 2 Performance/Power Ratings between 1 to 5 Unknown value Unknown ratings 15
Outline • Motivation • Methods • Experimental Results • Conclusion 16
Ge Genera rati ting D g Data ta f for A r Accura racy • Goal: different enough but still realistic to be plausible 17
Ge Genera rati ting D g Data ta f for A r Accura racy • Goal: different enough but still realistic to be plausible • How: Random number generator à different but not plausible 18
Ge Genera rati ting D g Data ta f for A r Accura racy • Goal: different enough but still realistic to be plausible • How: à different but not plausible Random number generator Gaussian Mixture Model (GMM) à plausible but not different Component 1 K: number of components x i : data points, i=1,…,N Component 2 w k : weight of k-th component Component 3 Density Probability that x i belongs to k-th comp: Data 19
Ge Genera rati ting D g Data ta wi with th a a GM GMM vide Known Data Di Divi Known Applications Computer System Configurations 20
Ge Genera rati ting D g Data ta wi with th a a GM GMM Learn GMMs vide Known Data Le Di Divi Known Applications Computer System Configurations Density Behavior Density Behavior 21
Ge Genera rati ting D g Data ta wi with th a a GM GMM Learn GMMs Swap Max and Min Divi Di vide Known Data Le Sw Known Applications Computer System Configurations Density Density Behavior Behavior Density Density Behavior Behavior 22
Ge Genera rati ting D g Data ta wi with th a a GM GMM Generate new data Ge Swap Max and Min Sw Le Learn GMMs vide Known Data Divi Di Known Applications Computer System Configurations Density Density Behavior Behavior Density Density Behavior Behavior 23
Ge Genera rati ting D g Data ta wi with th a a GM GMM Generate new data Ge Swap Max and Min Sw Concatenate Co Learn GMMs Le Divi Di vide Known Data Known Computer System Configurations Known Applications Density Density Applications New Application Behavior Behavior Density Density Behavior Behavior 24
Mult Multi-ph phase e Sa Sampl pling Input: Configuration-Application data matrix, Sampling budget N Matrix Completion with Sample Size N/ N/2 Known Applications Computer System Configurations New Application 25
Mult Multi-ph phase e Sa Sampl pling Input: Configuration-Application data matrix, Sampling budget N Matrix Completion Estimated with Sample Size N/ N/2 Behavior for New Application Known Applications Computer System Configurations New Application 26
Mult Multi-ph phase e Sa Sampl pling Input: Configuration-Application data matrix, Sampling budget N Estimated Matrix Completion Select N/ N/2 Behavior for New with Sample Size N/ N/2 Best Configs Application Known Applications 3 Computer System Configurations 4 New Application 1 5 2 6 8 7 27
Mult Multi-ph phase e Sa Sampl pling Input: Configuration-Application data matrix, Sampling budget N Estimated Matrix Completion Matrix Completion with N/ N/2 original Select N/ N/2 Behavior for New with Sample Size N/ N/2 samples and N/ N/2 estimated be best co configs Best Configs Application Known Applications Known Applications 3 Computer System Configurations Computer System Configurations 4 New Application New Application 1 5 2 6 8 7 28
Outline • Motivation • Methods • Experimental Results • Conclusion 29
Ex Exper periment imental al Set etup up Mobile Server System Ubuntu 14.04 Linux 3.2.0 system Architecture ARM big.LITTLE Intel Xeon E5-2690 # Applications 21 22 # Configurations 128 1024 30
Le Learning Models and Fr Frameworks ks Learning Models Category MCGD MC Fi First st co comprehe hens nsive stud udy of mat atrix MCMF MC completion co n (MC) al algorithm hms for Nuclear MC syst sy stem ems s op optim imiz ization ion ta task WNNM MC HBM Bayesian 31
Le Learning Models and Fr Frameworks ks Learning Models Category MCGD MC First Fi st co comprehe hens nsive stud udy of mat atrix MCMF MC completion co n (MC) al algorithm hms for Nuclear MC syst sy stem ems s op optim imiz ization ion ta task WNNM MC HBM Bayesian Frameworks Definitions Vanilla Basic learners GM Generative model MP Multi-phase sampling MP-GM Combine GM and MP 32
Im Improve Predict ction Accu ccuracy cy w/ GM High is Better Mobile Server Average percentage points of accuracy improvement 33
Im Improve Ener nergy gy Sav avings ings w/ / MP MP Lower is Better Server Mobile Average energy improvement 34
Co Conclusion Clock Speed Memo Memo Cores ry ry Contr Contr oller 1 oller 2 Memory controller 35
Co Conclusion Clock Speed Memo Memo Cores ry ry Contr Contr oller 1 oller 2 Memory controller 36
Co Conclusion Clock Speed Memo Memo Cores ry ry Contr Contr oller 1 oller 2 Memory controller 37
Recommend
More recommend
