Ge Gene nerative and nd Mul ulti-phase se Learning fo for - - PowerPoint PPT Presentation

▶
ge gene nerative and nd mul ulti phase se learning fo for
SMART_READER_LITE
LIVE PREVIEW

Ge Gene nerative and nd Mul ulti-phase se Learning fo for - - PowerPoint PPT Presentation

Aug 18, 2023 239 likes •642 views

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

slide-1
SLIDE 1

Ge Gene nerative and nd Mul ulti-phase se Learning fo for Computer Syst stems s Optimization

Yi Yi Di Ding, Nikita Mishra, Henry Hoffmann

slide-2
SLIDE 2

Co Compu puter Syst ystems Op Opti timizati tion

  • n

1

  • Optimizing modern computer systems requires tradeoffs:

Deliver reliable performance Minimize energy consumption

slide-3
SLIDE 3

Co Compu puter Syst ystems Op Opti timizati tion

  • n

2

  • 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

slide-4
SLIDE 4

Co Compu puter Syst ystems Op Opti timizati tion

  • n

3

  • 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?
slide-5
SLIDE 5

Ex Example ample of f a a Config nfigur urat atio ion n Spac ace 𝐃

4

2.26 Hz

Memory Controller 1 Memory Controller 2

Clock Speed Cores Memory controller

slide-6
SLIDE 6

Mac Machine hine Le Learning to to the the Re Rescue

5

slide-7
SLIDE 7

Mac Machine hine Lear earning ning to to the the Re Rescue

  • However…

Sc Scarce da data ta: expensive collection, limited range behavior

6

slide-8
SLIDE 8

Mac Machine hine Le Learning to to the the Re Rescue

  • However …

Sc Scarce da data ta: expensive collection, limited range behavior As Asym ymmetri ric be benefits: only configs on op

  • ptim

imal l fron

  • ntie

ier useful

7

slide-9
SLIDE 9

Mac Machine hine Le Learning to to the the Re Rescue

  • However …

Sc Scarce da data ta: expensive collection, limited range behavior As Asym ymmetri ric be benefits: only configs on op

  • ptim

imal l fron

  • ntie

ier useful

8

àGenerative model àMulti-phase sampling

slide-10
SLIDE 10

Mac Machine hine Le Learning to to the the Re Rescue

  • However …

Sc Scarce da data ta: expensive collection, limited range behavior As Asym ymmetri ric be benefits: only configs on op

  • ptim

imal l fron

  • ntie

ier useful

9

àGenerative model àMulti-phase sampling

We We advocate:

  • De

De-em empha hasiz izing ing predic edictio tion accu accuracy acy

  • In

Incorp

  • rpora
  • rating s

system s stru ructure i into

  • lea

learner ner

slide-11
SLIDE 11

Pr Problem blem Formulat mulatio ion

10

Meet Meet la laten ency con

  • nstra

rain ints wit with min minima imal l en ener ergy via via sy syst stem co configs Re Require th the po power an and pe perfo formance pr profi file fo for ap applicat ations Le Learn to estimate these values

Ex Expensiv ive Co Confi fig: : an an al allocat ation of

  • f

ha hardware re resourc rces to an an ap applicat ation

slide-12
SLIDE 12

SR SRAD AD on AR ARM bi big.LITTLE syst system em

11

Model A Model B Optimal points Just far enough True data Non-optimal points True data Very far Goodness of fit 99% Energy over optimal 22% ❌ 0 ✅

slide-13
SLIDE 13

SR SRAD AD on AR ARM bi big.LITTLE syst system em

12

Model A Model B Optimal points Just far enough True data Non-optimal points True data Just far enough Goodness of fit 99% Energy over optimal 22% ❌ 0 ✅

Ke Key Insight: Hi High accuracy ≠ good syst good system re results

slide-14
SLIDE 14

Re Recommender Systems -> > Le Learning by Examples

13

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

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)

slide-15
SLIDE 15

An An An Analo alogy

14

1 3 5 5 4 5 4 4 2 1 2 2 4 1 2 3 4 3 5 2 4 5 4 2 4 2 4 1 2 4 1 3 2 3 3 4 2 2 5 3 1 3 3 2 4 2 Movies Users

?

Ratings between 1 to 5 Unknown ratings

slide-16
SLIDE 16

An An An Analo alogy

15

1 3 5 5 4 5 4 4 2 1 2 2 4 1 2 3 4 3 5 2 4 5 4 2 4 2 4 1 2 4 1 3 2 3 3 4 2 2 5 3 1 3 3 2 4 2 Movies Users

?

Performance/Power Unknown value Ratings between 1 to 5 Unknown ratings

slide-17
SLIDE 17

Outline

  • Motivation
  • Methods
  • Experimental Results
  • Conclusion

16

slide-18
SLIDE 18

Ge Genera rati ting D g Data ta f for A r Accura racy

  • Goal: different enough but still realistic to be plausible

17

slide-19
SLIDE 19

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

18

àdifferent but not plausible

slide-20
SLIDE 20

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 Gaussian Mixture Model (GMM)

Data Density Component 1 Component 2 Component 3 K: number of components xi : data points, i=1,…,N wk: weight of k-th component Probability that xi belongs to k-th comp:

19

àdifferent but not plausible àplausible but not different

slide-21
SLIDE 21

Ge Genera rati ting D g Data ta wi with th a a GM GMM

Computer System Configurations

Known Applications Di Divi vide Known Data

20

slide-22
SLIDE 22

Ge Genera rati ting D g Data ta wi with th a a GM GMM

Computer System Configurations

Known Applications Di Divi vide Known Data Le Learn GMMs

Behavior Density Behavior Density 21

slide-23
SLIDE 23

Ge Genera rati ting D g Data ta wi with th a a GM GMM

Computer System Configurations

Known Applications Di Divi vide Known Data Le Learn GMMs

Behavior Density Behavior Density Behavior Density Behavior Density

Sw Swap Max and Min

22

slide-24
SLIDE 24

Ge Genera rati ting D g Data ta wi with th a a GM GMM

Computer System Configurations

Known Applications Di Divi vide Known Data Le Learn GMMs

Behavior Density Behavior Density

Sw Swap Max and Min

Behavior Density Behavior Density

Ge Generate new data

23

slide-25
SLIDE 25

Ge Genera rati ting D g Data ta wi with th a a GM GMM

24 Computer System Configurations

Known Applications Di Divi vide Known Data Le Learn GMMs

Behavior Density Behavior Density

Sw Swap Max and Min

Behavior Density Behavior Density

Ge Generate new data Known Applications Co Concatenate New Application

slide-26
SLIDE 26

Mult Multi-ph phase e Sa Sampl pling

Computer System Configurations Known Applications New Application Matrix Completion with Sample Size N/ N/2

25

Input: Configuration-Application data matrix, Sampling budget N

slide-27
SLIDE 27

Mult Multi-ph phase e Sa Sampl pling

Computer System Configurations Known Applications New Application Matrix Completion with Sample Size N/ N/2 Estimated Behavior for New Application

26

Input: Configuration-Application data matrix, Sampling budget N

slide-28
SLIDE 28

Mult Multi-ph phase e Sa Sampl pling

Computer System Configurations Known Applications New Application Matrix Completion with Sample Size N/ N/2 Estimated Behavior for New Application 3 4 1 5 2 6 8 7 Select N/ N/2 Best Configs

27

Input: Configuration-Application data matrix, Sampling budget N

slide-29
SLIDE 29

Mult Multi-ph phase e Sa Sampl pling

Computer System Configurations Known Applications New Application Matrix Completion with Sample Size N/ N/2 Estimated Behavior for New Application 3 4 1 5 2 6 8 7 Select N/ N/2 Best Configs New Application Known Applications Computer System Configurations Matrix Completion with N/ N/2 original samples and N/ N/2 estimated be best co configs

28

Input: Configuration-Application data matrix, Sampling budget N

slide-30
SLIDE 30

Outline

  • Motivation
  • Methods
  • Experimental Results
  • Conclusion

29

slide-31
SLIDE 31

Ex Exper periment imental al Set etup up

30

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

slide-32
SLIDE 32

Le Learning Models and Fr Frameworks ks

31

Learning Models Category MCGD MC MCMF MC Nuclear MC WNNM MC HBM Bayesian

Fi First st co comprehe hens nsive stud udy of mat atrix co completion n (MC) al algorithm hms for sy syst stem ems s op

  • ptim

imiz ization ion ta task

slide-33
SLIDE 33

Le Learning Models and Fr Frameworks ks

32

Learning Models Category MCGD MC MCMF MC Nuclear MC WNNM MC HBM Bayesian Frameworks Definitions Vanilla Basic learners GM Generative model MP Multi-phase sampling MP-GM Combine GM and MP

Fi First st co comprehe hens nsive stud udy of mat atrix co completion n (MC) al algorithm hms for sy syst stem ems s op

  • ptim

imiz ization ion ta task

slide-34
SLIDE 34

Im Improve Predict ction Accu ccuracy cy w/ GM

Mobile Server Average percentage points of accuracy improvement

33

High is Better

slide-35
SLIDE 35

Im Improve Ener nergy gy Sav avings ings w/ / MP MP

Mobile Server Average energy improvement

34

Lower is Better

slide-36
SLIDE 36

Co Conclusion

35

Memo ry Contr

  • ller 1

Memo ry Contr

  • ller 2

Clock Speed Cores Memory controller

slide-37
SLIDE 37

Co Conclusion

36

Memo ry Contr

  • ller 1

Memo ry Contr

  • ller 2

Clock Speed Cores Memory controller

slide-38
SLIDE 38

Co Conclusion

37

Memo ry Contr

  • ller 1

Memo ry Contr

  • ller 2

Clock Speed Cores Memory controller

slide-39
SLIDE 39

Co Conclusion

38

Yi Yi Di Ding, Nikita Mishra, and Henry Hoffmann. 2019. Generative and Multiphase Learning for Computer Systems Optimization. In The 46th Annual International Symposium on Computer Architecture (ISCA ’19)

We We advocate:

  • De

De-em empha hasiz izing ing predic edictio tion accu accuracy acy

  • In

Incorp

  • rpora
  • rating s

system s stru ructure i into

  • lea

learner ner

Memo ry Contr

  • ller 1

Memo ry Contr

  • ller 2

Clock Speed Cores Memory controller