Multi-Tenant Data Centers Mohammad A. Islam, Xiaoqi Ren, Shaolei Ren, - - PowerPoint PPT Presentation

▶
multi tenant data centers
SMART_READER_LITE
LIVE PREVIEW

Multi-Tenant Data Centers Mohammad A. Islam, Xiaoqi Ren, Shaolei Ren, - - PowerPoint PPT Presentation

Dec 07, 2022 19 likes •773 views

A Spot Capacity Market to Increase Power Infrastructure Utilization in Multi-Tenant Data Centers Mohammad A. Islam, Xiaoqi Ren, Shaolei Ren, and Adam Wierman This work was supported in part by the U.S. NSF under grants CNS-1551661, CNS-1565474,

slide-1
SLIDE 1

A Spot Capacity Market to Increase Power Infrastructure Utilization in Multi-Tenant Data Centers

Mohammad A. Islam, Xiaoqi Ren, Shaolei Ren, and Adam Wierman

This work was supported in part by the U.S. NSF under grants CNS-1551661, CNS-1565474, CNS-1518941, CPS154471, ECCS-1610471, and AitF-1637598.

slide-2
SLIDE 2

Multi-tenant data centers

Utility (Primary) Generator UPS ATS P D U P D U

2

slide-3
SLIDE 3

Managed by

  • perator

Multi-tenant data centers

Utility (Primary) Generator UPS ATS P D U P D U

2

slide-4
SLIDE 4

Managed by individual tenants Managed by

  • perator

Multi-tenant data centers

Utility (Primary) Generator UPS ATS P D U P D U

2

slide-5
SLIDE 5

3

Multi-tenant data centers are everywhere

slide-6
SLIDE 6

3

Google, Amazon, MS, Fb… :7.8% Enterprise: 53% Multi-tenant: 37%

Multi-tenant data centers are everywhere

slide-7
SLIDE 7

4

Who are using multi-tenant data centers?

slide-8
SLIDE 8

4

25% of Apple’s servers ate in multi-tenant data centers

Who are using multi-tenant data centers?

slide-9
SLIDE 9

4

25% of Apple’s servers ate in multi-tenant data centers

Who are using multi-tenant data centers?

slide-10
SLIDE 10

4

25% of Apple’s servers ate in multi-tenant data centers

Who are using multi-tenant data centers?

slide-11
SLIDE 11

Data center costs breakdown

5

Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev.

slide-12
SLIDE 12

Data center costs breakdown

5

Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev.

slide-13
SLIDE 13

Data center costs breakdown

5

Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev.

Capital Expenditure (CapEx)

slide-14
SLIDE 14

Data center costs breakdown

5

Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev.

Capital Expenditure (CapEx) Operational Expenditure (OpEx)

slide-15
SLIDE 15

Data center costs breakdown

5

Source: A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel. 2008. The cost of a cloud: research problems in data center networks. SIGCOMM Comput. Commun. Rev.

Capital Expenditure (CapEx) Operational Expenditure (OpEx)

slide-16
SLIDE 16

Cost of infrastructure

Utility (Primary) Generator UPS ATS P D U P D U

6

slide-17
SLIDE 17

Underutilization in data centers

7

slide-18
SLIDE 18

Underutilization in data centers

7

slide-19
SLIDE 19

Underutilization in data centers

7

slide-20
SLIDE 20

Increase infrastructure utilization

8

slide-21
SLIDE 21

Exploit the “spot capacity” Increase infrastructure utilization

8

slide-22
SLIDE 22

Some inspirations

  • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09

9

slide-23
SLIDE 23

Some inspirations

  • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09
  • “Spot instances” from Amazon

9

slide-24
SLIDE 24

Some inspirations

  • “Power routing” in ASPLOS’10 and “soft fuse” in EuroSys’09
  • “Spot instances” from Amazon
  • “Preemptible VM” from Google Cloud

9

slide-25
SLIDE 25

Spot capacity in multi-tenant data centers

10

slide-26
SLIDE 26

No centralized control

Spot capacity in multi-tenant data centers

10

slide-27
SLIDE 27

No centralized control  Power routing,…

Spot capacity in multi-tenant data centers

10

slide-28
SLIDE 28

A market for spot capacity

No centralized control  Power routing,…

Spot capacity in multi-tenant data centers

10

slide-29
SLIDE 29

A market for spot capacity

No centralized control  Power routing,… Tenants buy spot capacity from the data center operator

Spot capacity in multi-tenant data centers

10

slide-30
SLIDE 30

Spot capacity in multi-tenant data centers

  • Flexibility for cost conscious tenants

11

slide-31
SLIDE 31

Spot capacity in multi-tenant data centers

  • Flexibility for cost conscious tenants

Peak-based subscription

11

slide-32
SLIDE 32

Spot capacity in multi-tenant data centers

  • Flexibility for cost conscious tenants

Peak-based subscription Conservative subscription

11

slide-33
SLIDE 33

Spot capacity in multi-tenant data centers

  • Flexibility for cost conscious tenants

Peak-based subscription Conservative subscription Spot capacity

11

slide-34
SLIDE 34
  • Tenants:
  • tenants with insufficient capacity reservations can temporarily

process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise).

Spot capacity in multi-tenant data centers

12

slide-35
SLIDE 35
  • Tenants:
  • tenants with insufficient capacity reservations can temporarily

process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise).

  • Operator:
  • Revenue from guaranteed capacity: not affected
  • Extra revenue from spot capacity

Spot capacity in multi-tenant data centers

12

slide-36
SLIDE 36
  • Tenants:
  • tenants with insufficient capacity reservations can temporarily

process its workloads without power capping (or cap power less frequently/aggressively than it would otherwise).

  • Operator:
  • Revenue from guaranteed capacity: not affected
  • Extra revenue from spot capacity

Spot capacity market is a win-win for both tenants and operator

Spot capacity in multi-tenant data centers

12

slide-37
SLIDE 37

Challenges

  • Spot capacity is limited and intermittent

13

slide-38
SLIDE 38

Challenges

  • Spot capacity is limited and intermittent
  • Tenants’ spot capacity need is dynamic and invisible to the

data center operator

13

slide-39
SLIDE 39

Challenges

  • Spot capacity is limited and intermittent
  • Tenants’ spot capacity need is dynamic and invisible to the

data center operator

  • Infrastructure constraints require fine granularity in spot

capacity allocation (e.g., rack level)

13

slide-40
SLIDE 40

Goal: A scalable and runtime design for spot capacity allocation

14

slide-41
SLIDE 41

Problem formulation

  • Goal: operator profit maximization

15

slide-42
SLIDE 42

Rack level demand

Problem formulation

  • Goal: operator profit maximization

15

slide-43
SLIDE 43

Price of spot capacity Rack level demand

Problem formulation

  • Goal: operator profit maximization

15

slide-44
SLIDE 44

Infrastructure constraints Price of spot capacity Rack level demand

Problem formulation

  • Goal: operator profit maximization

15

slide-45
SLIDE 45

How to solve it?

16

slide-46
SLIDE 46

Unknown

How to solve it?

16

slide-47
SLIDE 47

Unknown

How to solve it?

  • Soliciting the demand curve  privacy and overhead

16

slide-48
SLIDE 48

Unknown

How to solve it?

  • Soliciting the demand curve  privacy and overhead
  • Pre-set price  low level demand prediction

16

slide-49
SLIDE 49

Unknown

How to solve it?

  • Soliciting the demand curve  privacy and overhead
  • Pre-set price  low level demand prediction
  • Market approach  an in-between solution

16

slide-50
SLIDE 50

SpotDC: spot capacity management

17

Operator Tenants

slide-51
SLIDE 51

SpotDC: spot capacity management

17

Spot capacity predictions

Operator Tenants

slide-52
SLIDE 52

SpotDC: spot capacity management

17

Spot capacity predictions

Operator Tenants

Response (bids)

slide-53
SLIDE 53

SpotDC: spot capacity management

17

Spot capacity predictions

Operator Tenants

Response (bids) Price and actual spot power allocation

slide-54
SLIDE 54

SpotDC: spot capacity management

17

Spot capacity predictions

Operator Tenants

Response (bids) Price and actual spot power allocation Gain spot power

slide-55
SLIDE 55

Timings in SpotDC

18

slide-56
SLIDE 56

Demand bidding

  • A piece-wise-linear bid

19

slide-57
SLIDE 57

Demand bidding

  • A piece-wise-linear bid
  • Tenants only submit four parameters

19

slide-58
SLIDE 58

Demand bidding

  • A piece-wise-linear bid
  • Tenants only submit four parameters
  • Captures tenants’ demand elasticity

19

slide-59
SLIDE 59

Spot capacity prediction

  • Available spot capacity prediction: max - predicted
  • UPS and PDU level predictions: Use previous time slot usage

as references.

20

slide-60
SLIDE 60

Spot capacity prediction

  • Available spot capacity prediction: max - predicted
  • UPS and PDU level predictions: Use previous time slot usage

as references.

20

slide-61
SLIDE 61

Spot capacity prediction

  • Available spot capacity prediction: max - predicted
  • UPS and PDU level predictions: Use previous time slot usage

as references.

20

Less than ±2.5% change 99% of the time

slide-62
SLIDE 62

Evaluation methodology

  • 10 tenants with sprinting (delay sensitive) and opportunistic

(delay tolerance) workloads

  • Using Dynamic voltage and frequency scaling (DVFS) for

power scaling.

21

slide-63
SLIDE 63

Evaluation methodology

  • 10 tenants with sprinting (delay sensitive) and opportunistic

(delay tolerance) workloads

  • Using Dynamic voltage and frequency scaling (DVFS) for

power scaling.

21

slide-64
SLIDE 64

Evaluation methodology

  • 10 tenants with sprinting (delay sensitive) and opportunistic

(delay tolerance) workloads

  • Using Dynamic voltage and frequency scaling (DVFS) for

power scaling.

21

slide-65
SLIDE 65

Performance evaluation

22

slide-66
SLIDE 66

Performance evaluation

22

Sprinting tenants drive up the price

slide-67
SLIDE 67

Performance evaluation

23

slide-68
SLIDE 68

Performance evaluation

23

Sprinting tenants avoid SLO violations

slide-69
SLIDE 69

Performance evaluation

23

Sprinting tenants avoid SLO violations Opportunistic tenants gain throughput boost

slide-70
SLIDE 70

Tenants’ benefit from SpotDC

24

slide-71
SLIDE 71

Tenants’ benefit from SpotDC

24

Performance boosts with SpotDC

slide-72
SLIDE 72

Operator’s extra profit

25

slide-73
SLIDE 73

Operator’s extra profit

25

SpotDC is close to optimal allocation with full information

slide-74
SLIDE 74

SpotDC: Spot capacity management

A market-based approach for providing spot capacity to tenants and helping operator further increase data center utilization

26

slide-75
SLIDE 75

SpotDC: Spot capacity management

A market-based approach for providing spot capacity to tenants and helping operator further increase data center utilization

26

Simple, Scalable & Efficient