dc drf adaptive multi resource sharing at public cloud
play

DC-DRF : Adaptive Multi- Resource Sharing at Public Cloud Scale - PowerPoint PPT Presentation

Dec 18, 2023 •319 likes •1.21k views

DC-DRF : Adaptive Multi- Resource Sharing at Public Cloud Scale ACM Symposium on Cloud Computing 2018 Ian A Kash, Greg OShea, Stavros Volos 1 Public Cloud DC hosting enterprise customers O(100K) servers, mostly small tenants 2

  1. DC-DRF : Adaptive Multi- Resource Sharing at Public Cloud Scale 
 ACM Symposium on Cloud Computing 2018 Ian A Kash, Greg O’Shea, Stavros Volos 1

  2. Public Cloud DC hosting enterprise customers 
 O(100K) servers, mostly small tenants 2

  3. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 3

  4. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 One VM in compute V T O R1 V T O R2 VTORa VTORb server in compute rack T X1 R X1 TXb R X b S S D b compute storage 4

  5. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 One VM in compute V T O R1 V T O R2 VTORa VTORb server in compute rack T X1 R X1 TXb R X b One VHD in storage server in storage rack S S D b compute storage 5

  6. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 6

  7. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 7

  8. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 8

  9. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 9

  10. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 10

  11. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b 11

  12. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 12

  13. Small customer : one VM accessing storage T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 13

  14. Result: a multi-resource “demand vector” T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 14

  15. Encodes resource id and proportions T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 V T O R1 V T O R2 VTORa VTORb T X1 R X1 TXb R X b S S D b compute storage 15

  16. Encodes resource id and proportions T X1 VTOR1 VTORb R X b S S D b T X b VTORa VTOR2 R X1 Any element could be a V T O R1 V T O R2 VTORa VTORb bottleneck to performance T X1 R X1 TXb R X b S S D b compute storage 16

  17. Demand vectors form a sparse demand matrix r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - - - - - - - - - n 1 - - - - - - - - - - n 2 - - - - - - - - - - - - - - - - - - - - n 3 - - - - - - - - - - n 4 - - - - - - - - - - n 5 n 6 - - - - - - - - - - - - - - - - - - - - n 7 n 8 - - - - - - - - - - n 9 - - - - - - - - - - 17

  18. Columns are shared physical resources r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - - - - - - - - - n 1 - - - - - - - - - - n 2 - - - - - - - - - - - - - - - - - - - - n 3 - - - - - - - - - - n 4 - - - - - - - - - - n 5 n 6 - - - - - - - - - - - - - - - - - - - - n 7 n 8 - - - - - - - - - - n 9 - - - - - - - - - - 18

  19. Rows are tenants’ demand vectors r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - - - - - - - - - n 1 - - - - - - - - - - n 2 - - - - - - - - - - - - - - - - - - - - n 3 - - - - - - - - - - n 4 - - - - - - - - - - n 5 n 6 - - - - - - - - - - - - - - - - - - - - n 7 n 8 - - - - - - - - - - n 9 - - - - - - - - - - 19

  20. Shown as fractions of a resource r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - 1.0 - - - - - - .92 n 1 - - - - - - - - - - n 2 - - - - - - - - - - - - - - - - - - - - n 3 - - - - - - - - - - n 4 - - - - - - - - - - n 5 n 6 - - - - - - - - - - - - - - - - - - - - n 7 n 8 - - - - - - - - - - n 9 - - - - - - - - - - 20

  21. Large and very sparse matrix r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - 1.0 - - - - - - .92 n 1 .95 - .47 - - - - - 1.0 - n 2 .54 1.0 - .30 .33 .23 .55 - .56 .31 - .41 .20 .12 .13 .09 .23 1.0 .23 .13 n 3 DC matrix 100K by 100K - 1.0 - .30 - .23 .55 - - .31 n 4 Rows mostly empty - .41 .09 .12 1.0 .64 .23 .20 .13 .13 n 5 n 6 .32 - .09 .12 1.0 .64 .23 .20 .13 .13 - - - - - - 1.0 - - .57 n 7 n 8 - - .56 .64 .20 .32 .13 .09 1.0 .23 n 9 .90 .27 .45 .64 .20 .32 .13 .09 1.0 .56 21

  22. Provider has multi-resource allocation problem • Goal: maintain acceptable service level for all tenants • Acceptable means always “willing to pay” • Avoid abrupt performance collapse for any tenant • Assuming aggressive (noisy) neighbors and oversubscription • DC-DRF builds on existing multi-resource algorithms • DRF [Ghodsi et al, NSDI’11] • EDRF [Parkes et al., EC2012] • Challenging at DC scale: EDRF iterates and is 22

  23. Systems aspects 23

  24. Systems challenges • How to capture multi-resource demand vectors? • How to enforce multi-resource allocations? • DRF implies central SDN-like controller – good or bad? • Good: Simpler algorithm and global view • Bad: EDRF at Public Cloud DC scale 24

  25. SIGCOMM 2015 demonstration 25

  26. SIGCOMM 2015 demonstration Central controller running EDRF Pass1: reservation-based SLAs Pass2: work conservation of residual 26

  27. SIGCOMM 2015 demonstration 4 tenants, 30 VMs each Spread over 10 servers R/W to 2X storage servers 40Gb RDMA switch 27

  28. SIGCOMM 2015 demonstration Demand estimation and enforcement in HyperV 28

  29. SIGCOMM 2015 demonstration Aggressive red tenant Perf. collapses for blue,yellow,green 29

  30. video 30

  31. SIGCOMM 2015 demonstration What did we learn from prototype? Potentially very powerful. But EDRF algorithm not scaling well. 31

  32. The algorithms • to understand DC-DRF first understand EDRF • to understand DRF first understand max-min 32

  33. Max-Min fairness : mice before elephants • Maximize the minimum allocation across competing tenants • Allocate fractions of a single shared resource based on demand • No tenant gets a larger fraction than its demand • Tenants with unsatisfiable demand obtain equal share Residual resource = 1.0 Residual resource = 0.7 Tenants remaining = 4 Tenants remaining = 2 .35 D Current share = 1.0/4 Current share = 0.7/2 x t = 0.25 x t = 0.6 .35 C Demand 0.35 0.5 x t =0.35 x t =0.25 B 0.2 0.2 0.1 0.1 A Allocated D A B C Tenant 33

  34. How to handle multiple resources? r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - 1.0 - - - - - - .92 n 1 .95 - .47 - - - - - 1.0 - n 2 .54 1.0 - .30 .33 .23 .55 - .56 .31 - .41 .20 .12 .13 .09 .23 1.0 .23 .13 n 3 - 1.0 - .30 - .23 .55 - - .31 n 4 - .41 .09 .12 1.0 .64 .23 .20 .13 .13 n 5 n 6 .32 - .09 .12 1.0 .64 .23 .20 .13 .13 - - - - - - 1.0 - - .57 n 7 n 8 - - .56 .64 .20 .32 .13 .09 1.0 .23 n 9 .90 .27 .45 .64 .20 .32 .13 .09 1.0 .56 34

  35. Dominant Resource Fairness (DRF) • For each tenant identifies its Dominant Resource • The resource of which it demands the largest fraction • Apply max-min fairness across dominant shares • Maximize smallest dominant share in system • Then second smallest, and so on… • Think : find the smallest mouse across all columns 35

  36. Demand vectors normalized by Dominant Resource r 0 r 1 r 2 r 3 r 4 r 5 r 6 r 7 r 8 r 9 n 0 - - 1.0 - - - - - - .92 n 1 .95 - .47 - - - - - 1.0 - n 2 .54 1.0 - .30 .33 .23 .55 - .56 .31 - .41 .20 .12 .13 .09 .23 1.0 .23 .13 n 3 - 1.0 - .30 - .23 .55 - - .31 n 4 - .41 .09 .12 1.0 .64 .23 .20 .13 .13 n 5 n 6 .32 - .09 .12 1.0 .64 .23 .20 .13 .13 - - - - - - 1.0 - - .57 n 7 n 8 - - .56 .64 .20 .32 .13 .09 1.0 .23 n 9 .90 .27 .45 .64 .20 .32 .13 .09 1.0 .56 36

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

Adaptive Knowledge Sharing in Multi-Task Learning: Improving Low-Resource Neural Machine

Adaptive Knowledge Sharing in Multi-Task Learning: Improving Low-Resource Neural Machine

Adaptive Knowledge Sharing in Multi-Task Learning: Improving Low-Resource Neural Machine Translation Poorya ZareMoodi , Wray Buntine, Gholamreza (Reza) Haffari Monash University Slides : Roadmap ! 2 Introduction & background Adaptive

360 views • 17 slides
UTSA Community-Based Secure Information and Resource Sharing in AWS Public Cloud Cyber Incident

UTSA Community-Based Secure Information and Resource Sharing in AWS Public Cloud Cyber Incident

UTSA Community-Based Secure Information and Resource Sharing in AWS Public Cloud Cyber Incident Response A Model for Information and Resource Sharing Amy(Yun) Zhang, Farhan Patwa, Ravi Sandhu Institute for Cyber Security University of Texas

578 views • 19 slides
MULTI-LEVEL MECHANISMS TO SUPPORT SPORADIC CLOUD COMPUTING MOBILE SERVICES BY RESOURCE-SHARING IN

MULTI-LEVEL MECHANISMS TO SUPPORT SPORADIC CLOUD COMPUTING MOBILE SERVICES BY RESOURCE-SHARING IN

MULTI-LEVEL MECHANISMS TO SUPPORT SPORADIC CLOUD COMPUTING MOBILE SERVICES BY RESOURCE-SHARING IN VANET We want to develop a new concept of sporadic CMA to deploy MOTIVATION contex-aware services in VANET. Intelligent Transport Systems (

367 views • 11 slides
UTSA Hierarchical Secure Information and Resource Sharing in OpenStack Community Cloud Cyber

UTSA Hierarchical Secure Information and Resource Sharing in OpenStack Community Cloud Cyber

UTSA Hierarchical Secure Information and Resource Sharing in OpenStack Community Cloud Cyber Incident Response An Model for Information and Resource Sharing Amy(Yun) Zhang, Farhan Patwa, Ravi Sandhu, Bo Tang Institute for Cyber Security

522 views • 17 slides
UTSA Secure Information and Resource Sharing in Cloud Infrastructure as a Service Cyber Incident

UTSA Secure Information and Resource Sharing in Cloud Infrastructure as a Service Cyber Incident

UTSA Secure Information and Resource Sharing in Cloud Infrastructure as a Service Cyber Incident Response Models for Information and Resource Sharing Amy(Yun) Zhang, Ravi Sandhu Institute for Cyber Security University of Texas at San

704 views • 42 slides
UTSA Community-Based Secure Information and Resource Sharing in Azure Cloud IaaS Cyber Incident

UTSA Community-Based Secure Information and Resource Sharing in Azure Cloud IaaS Cyber Incident

UTSA Community-Based Secure Information and Resource Sharing in Azure Cloud IaaS Cyber Incident Response Models for Information and Resource Sharing Yun Zhang, Farhan Patwa , Ravi Sandhu Institute for Cyber Security University of Texas at

711 views • 20 slides
UTSA Information Sharing and Coordination Initiatives collaboration and coordination to

UTSA Information Sharing and Coordination Initiatives collaboration and coordination to

Secure Information and Resource Sharing in Cloud Infrastructure as a Service Cyber Incident Response Models for Information and Resource Sharing Amy(Yun) Zhang, Ram Krishnan, Ravi Sandhu Institute for Cyber Security University of Texas at San

219 views • 17 slides
Network and Computing Resource Sharing in Federated Cloud Systems Walter Cerroni Dept. of

Network and Computing Resource Sharing in Federated Cloud Systems Walter Cerroni Dept. of

Performance of Network and Computing Resource Sharing in Federated Cloud Systems Walter Cerroni Dept. of Electrical, Electronic and Information Engineering University of Bologna, Italy walter.cerroni@unibo.it Motivations Success of cloud

485 views • 35 slides
A Resource Allocation Controller for Cloud-based Adaptive Video Streaming Luca De Cicco , Saverio

A Resource Allocation Controller for Cloud-based Adaptive Video Streaming Luca De Cicco , Saverio

1/16 A Resource Allocation Controller for Cloud-based Adaptive Video Streaming Luca De Cicco , Saverio Mascolo, Dario Calamita Politecnico di Bari, Dipartimento di Ingegneria Elettrica e dellInformazione MCN 2013 - Budapest, Hungary 13 June

564 views • 27 slides
OVERSEE: Outsourcing Verification to Enable Resource Sharing in Edge Environment Reporter:

OVERSEE: Outsourcing Verification to Enable Resource Sharing in Edge Environment Reporter:

OVERSEE: Outsourcing Verification to Enable Resource Sharing in Edge Environment Reporter: Xiaoqing Cai August 2020 From Cloud to Edge Cloud Computing Edge Computing Low Latency Lack of Resources Cloud Computing Edge Computing No More

582 views • 17 slides
Infrastructure of Local Public Health: A Study of the Effects of Resource Sharing Among Local

Infrastructure of Local Public Health: A Study of the Effects of Resource Sharing Among Local

Building Evidence to Improve the Infrastructure of Local Public Health: A Study of the Effects of Resource Sharing Among Local Public Health Jurisdictions by the CT and MA Public Health Practice Based Research Networks Elaine OKeefe, MS,

799 views • 40 slides
Multi-Armed Bandits: Non-adaptive and Adaptive Sampling Instructor: Sham Kakade 1 The

Multi-Armed Bandits: Non-adaptive and Adaptive Sampling Instructor: Sham Kakade 1 The

CSE 547/Stat 548: Machine Learning for Big Data Lecture Multi-Armed Bandits: Non-adaptive and Adaptive Sampling Instructor: Sham Kakade 1 The (stochastic) multi-armed bandit problem The basic paradigm is as follows: K Independent Arms: a

273 views • 5 slides
Stout An Adaptive Interface to Scalable Cloud Storage John Dunagan John C. McCullough Alec

Stout An Adaptive Interface to Scalable Cloud Storage John Dunagan John C. McCullough Alec

Stout An Adaptive Interface to Scalable Cloud Storage John Dunagan John C. McCullough Alec Wolman Alex C. Snoeren UC San Diego Microsoft Research June 23, 2010 Scalable Multi-tiered Services 2 Scalable Multi-tiered Services app

1.45k views • 107 slides
Facilitating Communal Data Sharing in Public Clouds Roxana Geambasu Steve Gribble Hank Levy

Facilitating Communal Data Sharing in Public Clouds Roxana Geambasu Steve Gribble Hank Levy

Facilitating Communal Data Sharing in Public Clouds Roxana Geambasu Steve Gribble Hank Levy University of Washington Outline Vision: cloud as a platform for sharing code and data Why now: favorable cloud technology trends

531 views • 18 slides
Optimal p Sequential Resource Sharing and Exchange i in Multi Agent Systems l i S Yuanzhang

Optimal p Sequential Resource Sharing and Exchange i in Multi Agent Systems l i S Yuanzhang

Optimal p Sequential Resource Sharing and Exchange i in Multi Agent Systems l i S Yuanzhang Xiao Advisor: Prof. Mihaela van der Schaar Electrical Engineering Department UCLA Electrical Engineering Department, UCLA Ph.D. defense, March 3,

892 views • 61 slides
Chapter 6 Cloud Resource Management and Scheduling Contents Resource management and

Chapter 6 Cloud Resource Management and Scheduling Contents Resource management and

Chapter 6 Cloud Resource Management and Scheduling Contents Resource management and scheduling. Policies and mechanisms. Applications of control theory to cloud resource allocation. Stability of a two-level resource allocation

1.47k views • 39 slides
Object Oriented Programming COP3330 / CGS5409 Intro to Data Structures Vectors Linked

Object Oriented Programming COP3330 / CGS5409 Intro to Data Structures Vectors Linked

Object Oriented Programming COP3330 / CGS5409 Intro to Data Structures Vectors Linked Lists Queues Stacks C++ has some built-in methods of storing compound data in useful ways, like arrays and structs. Collectively

464 views • 28 slides
Week 15 - Monday What did we talk about last time? Timsort Tries Lab hours

Week 15 - Monday What did we talk about last time? Timsort Tries Lab hours

Week 15 - Monday What did we talk about last time? Timsort Tries Lab hours Wednesdays at 5 p.m. in The Point 113 Saturdays at noon in The Point 113 CS Club Tuesdays at 5 p.m. in The Point 113 (or next door in The

226 views • 21 slides
1 Peter Series Lesson #132 May 31, 2018 Dean Bible Ministries www.deanbibleministries.org Dr.

1 Peter Series Lesson #132 May 31, 2018 Dean Bible Ministries www.deanbibleministries.org Dr.

1 Peter Series Lesson #132 May 31, 2018 Dean Bible Ministries www.deanbibleministries.org Dr. Robert L. Dean, Jr. Spiritual Gifts 1 Peter 4:1011; 1 Cor. 12; Eph. 4:11, 12 1 Pet. 4:8787, And above all things have fervent love for one

198 views • 19 slides
Phylogeny and Sequence Simulation with Hybridization Michael Woodhams School of Mathematics and

Phylogeny and Sequence Simulation with Hybridization Michael Woodhams School of Mathematics and

Phylogeny and Sequence Simulation with Hybridization Michael Woodhams School of Mathematics and Physics, University of Tasmania Image credits: witchwarrior101.blogspot.com.au thedrawinghands.deviantart.com eduardoleon.deviantart.com

424 views • 16 slides
Announcements Reminder: Candidates on campus in next few weeks Day-by-day schedule may change,

Announcements Reminder: Candidates on campus in next few weeks Day-by-day schedule may change,

Announcements Reminder: Candidates on campus in next few weeks Day-by-day schedule may change, so check often Thursday Extra: Today @ 4:00 Tenure-track candidate #1 Kevin Angstadt, U. of Michigan All Computers Great and Small: Supporting

86 views • 8 slides
Sor$ng, Stacks, Queues Bryce Boe 2013/11/06 CS24, Fall 2013

Sor$ng, Stacks, Queues Bryce Boe 2013/11/06 CS24, Fall 2013

Sor$ng, Stacks, Queues Bryce Boe 2013/11/06 CS24, Fall 2013 Outline Finish Lab 5 Part 2 Review Lab 6 Review / Sor$ng / C++ Templates

193 views • 17 slides
The HISPACAT comparative database of syntactic constructions and its applications to syntactic

The HISPACAT comparative database of syntactic constructions and its applications to syntactic

Goals Theory Architecture Results Conclusions The HISPACAT comparative database of syntactic constructions and its applications to syntactic variation research Workshop on Research Infrastructure for Linguistic Variation Xavier Villalba

509 views • 20 slides
Data Structures and Object-Oriented Design VIII Spring 2014 Carola Wenk Collections and Maps

Data Structures and Object-Oriented Design VIII Spring 2014 Carola Wenk Collections and Maps

Data Structures and Object-Oriented Design VIII Spring 2014 Carola Wenk Collections and Maps The Collection interface is for storage and access, while a Map interface is geared towards associating keys with objects. Student database

644 views • 22 slides

More recommend