Dynamic Content Allocation for Cloud- assisted Service of Periodic - - PowerPoint PPT Presentation

▶
dynamic content allocation for cloud assisted service of
SMART_READER_LITE
LIVE PREVIEW

Dynamic Content Allocation for Cloud- assisted Service of Periodic - - PowerPoint PPT Presentation

Sep 22, 2022 182 likes •966 views

Dynamic Content Allocation for Cloud- assisted Service of Periodic Workloads Gyrgy Dn Niklas Carlsson Royal Institute of Technology (KTH) Linkping University @ IEEE INFOCOM 2014 , Toronto, Canada, April/May 2014 Internet Content Delivery

slide-1
SLIDE 1

Dynamic Content Allocation for Cloud- assisted Service of Periodic Workloads

@ IEEE INFOCOM 2014, Toronto, Canada, April/May 2014

György Dán

Royal Institute of Technology (KTH)

Niklas Carlsson

Linköping University

slide-2
SLIDE 2

Internet Content Delivery

  • Large amounts of data with varying popularity
  • Multi-billion market ($8B to $20B, 2012-2015)
  • Goal: Minimize content delivery costs
  • Migration to cloud data centers
  • From: Dan and Carlsson, “Power-laws Revisited: A Large Scale

Measurement Study of Peer-to-Peer Content Popularity”, Proc. IPTPS 2010.

slide-3
SLIDE 3

Internet Content Delivery

  • Large amounts of data with varying popularity
  • Multi-billion market ($8B to $20B, 2012-2015)
  • Goal: Minimize content delivery costs
  • Migration to cloud data centers
  • From: Dan and Carlsson, “Power-laws Revisited: A Large Scale

Measurement Study of Peer-to-Peer Content Popularity”, Proc. IPTPS 2010.

slide-4
SLIDE 4

Internet Content Delivery

  • Large amounts of data with varying popularity
  • Multi-billion market ($8B to $20B, 2012-2015)
  • Goal: Minimize content delivery costs
  • Migration to cloud data centers
  • From: Dan and Carlsson, “Power-laws Revisited: A Large Scale

Measurement Study of Peer-to-Peer Content Popularity”, Proc. IPTPS 2010.

slide-5
SLIDE 5

Internet Content Delivery

  • Large amounts of data with varying popularity
  • Multi-billion market ($8B to $20B, 2012-2015)
  • Goal: Minimize content delivery costs
  • Migration to cloud data centers
  • From: Dan and Carlsson, “Power-laws Revisited: A Large Scale

Measurement Study of Peer-to-Peer Content Popularity”, Proc. IPTPS 2010.

slide-6
SLIDE 6

Periodic Workloads

  • Characterization of Spotify traces
  • In addition to diurnal traffic volumes …
  • … we found that also the Zipf exponent vary with time-of-day
slide-7
SLIDE 7

Content Delivery

  • Cloud-based delivery
  • Dedicated infrastructure

servers cloud

slide-8
SLIDE 8

Content Delivery

  • Cloud-based delivery
  • Flexible computation, storage, and bandwidth
  • Pay per volume and access
  • Dedicated infrastructure
  • Limited storage
  • Capped unmetered bandwidth
  • Potentially closer to the user

servers cloud

slide-9
SLIDE 9

Content Delivery

  • Cloud-based delivery
  • Flexible computation, storage, and bandwidth
  • Pay per volume and access
  • Dedicated infrastructure
  • Limited storage
  • Capped unmetered bandwidth
  • Potentially closer to the user
slide-10
SLIDE 10

Content Delivery

  • Cloud-based delivery
  • Flexible computation, storage, and bandwidth
  • Pay per volume and access
  • Dedicated infrastructure
  • Limited storage
  • Capped unmetered bandwidth
  • Potentially closer to the user
slide-11
SLIDE 11

Content Delivery

  • Cloud-based delivery
  • Flexible computation, storage, and bandwidth
  • Pay per volume and access
  • Dedicated infrastructure
  • Limited storage
  • Capped unmetered bandwidth
  • Potentially closer to the user
slide-12
SLIDE 12

Content Delivery

  • Cloud-based delivery
  • Flexible computation, storage, and bandwidth
  • Pay per volume and access
  • Dedicated infrastructure
  • Limited storage
  • Capped unmetered bandwidth
  • Potentially closer to the user

Cloud bandwidth elastic; however, flexible comes at premium …

slide-13
SLIDE 13

High-level problem

  • Minimize content delivery costs
  • Bandwidth

Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

slide-14
SLIDE 14

High-level problem

  • Minimize content delivery costs
  • Bandwidth

Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

How to get the best of two worlds?

slide-15
SLIDE 15

High-level problem

  • Minimize content delivery costs
  • How to get the best out of two worlds?
  • Bandwidth

Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

slide-16
SLIDE 16

High-level problem

  • Minimize content delivery costs
  • How to get the best out of two worlds?
  • Improved workload models and prediction enables prefetching …
  • Bandwidth

Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

slide-17
SLIDE 17

High-level problem

  • Minimize content delivery costs
  • How to get the best out of two worlds?
  • Improved workload models and predcition enables prefetching …
  • Dynamic content allocation
  • Utilize capped bandwidth (and storage) as much as possible
  • Use elastic cloud-based services to serve “spillover”
  • Bandwidth

Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

slide-18
SLIDE 18

Dynamic Content Allocation Problem

18

  • Formulate as a finite horizon dynamic

decision process problem

  • Show discrete time decision process

is good approximation

  • Define exact solution as MILP
  • Provide computationally feasible

approximations (and prove properties about approximation ratios)

  • Validate model and policies using

traces from Spotify

slide-19
SLIDE 19

Cost minimization formulation

slide-20
SLIDE 20

Cost minimization formulation

Total demand

slide-21
SLIDE 21

Cost minimization formulation

Demand of files in capped BW storage

slide-22
SLIDE 22

Cost minimization formulation

Capped BW limit (U)

slide-23
SLIDE 23

Cost minimization formulation

slide-24
SLIDE 24

Cost minimization formulation

Served from capped BW storage

slide-25
SLIDE 25

Cost minimization formulation

Served using elastic cloud resources

slide-26
SLIDE 26

Cost minimization formulation

Traffic due to allocation

slide-27
SLIDE 27

Cost minimization formulation

slide-28
SLIDE 28

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-29
SLIDE 29

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-30
SLIDE 30

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-31
SLIDE 31

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-32
SLIDE 32

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-33
SLIDE 33

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-34
SLIDE 34

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-35
SLIDE 35

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
slide-36
SLIDE 36

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
  • Equivalent formulation

Utilization maximization

slide-37
SLIDE 37

Cost minimization formulation

  • Traffic of files only in cloud
  • Spillover traffic
  • Traffic due to allocation
  • Total expected cost
  • Optimal policy
  • Equivalent formulation

Utilization maximization

slide-38
SLIDE 38

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

slide-39
SLIDE 39

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

Two file example

slide-40
SLIDE 40

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

Two file example

slide-41
SLIDE 41

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

Two file example

slide-42
SLIDE 42

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

Two file example

slide-43
SLIDE 43

Cost minimization formulation

  • Equivalent formulation

Utilization maximization

Two file example

slide-44
SLIDE 44
  • Equivalent formulation

Cost minimization formulation Utilization maximization

slide-45
SLIDE 45

Discrete-time Decision Problem

  • Equivalent formulation
slide-46
SLIDE 46
  • Approximation
  • Finite horizon decision

problem

Discrete-time Decision Problem

decrease exponentially

  • Equivalent formulation
slide-47
SLIDE 47
  • Approximation
  • Finite horizon decision

problem

Discrete-time Decision Problem

decrease exponentially

  • Equivalent formulation
slide-48
SLIDE 48
  • Approximation
  • Finite horizon decision

problem

Discrete-time Decision Problem

decrease exponentially

  • Equivalent formulation
slide-49
SLIDE 49
  • Approximation
  • Finite horizon decision

problem

Discrete-time Decision Problem

decrease exponentially

slide-50
SLIDE 50
  • Approximation
  • Finite horizon decision

problem

Discrete-time Decision Problem

Theorem: Exact solution as a MILP

decrease exponentially

slide-51
SLIDE 51
  • Consider next interval only

Policy: No Download Cost (NDC)

slide-52
SLIDE 52
  • Proposition 1: Unbounded approximation ratio
  • Proposition 2: Approximation bound
  • Consider next interval only

Policy: No Download Cost (NDC)

slide-53
SLIDE 53
  • Proposition 1: Unbounded approximation ratio
  • Proposition 2: Approximation bound
  • Consider next interval only

Policy: No Download Cost (NDC)

slide-54
SLIDE 54
  • Proposition 1: Unbounded approximation ratio
  • Proposition 2: Approximation bound
  • Consider next interval only

Policy: No Download Cost (NDC)

slide-55
SLIDE 55
  • Proposition 1: Unbounded approximation ratio
  • Proposition 2: Approximation bound
  • Consider next interval only

Policy: No Download Cost (NDC)

slide-56
SLIDE 56
  • Consider k next intervals

Policy: k-Step Look Ahead (k-SLA)

slide-57
SLIDE 57
  • Proposition 3: Unbounded approximation ratio
  • Proposition 4: Approximation bound
  • Consider k next intervals

Policy: k-Step Look Ahead (k-SLA)

slide-58
SLIDE 58
  • Proposition 3: Unbounded approximation ratio
  • Proposition 4: Approximation bound
  • Consider k next intervals

Policy: k-Step Look Ahead (k-SLA)

slide-59
SLIDE 59
  • Proposition 3: Unbounded approximation ratio
  • Proposition 4: Approximation bound
  • Consider k next intervals

Policy: k-Step Look Ahead (k-SLA)

slide-60
SLIDE 60
  • Proposition 3: Unbounded approximation ratio
  • Proposition 4: Approximation bound
  • Consider k next intervals

Policy: k-Step Look Ahead (k-SLA)

slide-61
SLIDE 61

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Workload: 3 groups of 1000 files; peaks N(0,2) offset by 8h for

each group; sinusoid with 24h period; min/max ratio N(0.075,0.075), file sizes U(L/2,3L/2), bandwidth demand Bounded Pareto (Bmin, Bmax, )

Based on Spotify trace characterization

slide-62
SLIDE 62

Trace-based analysis (Synthetic)

Normalize against policy that stores the most popular files

  • Normalized traffic savings
slide-63
SLIDE 63

Trace-based analysis (Synthetic)

  • Normalized traffic savings
slide-64
SLIDE 64

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Good
slide-65
SLIDE 65

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Modest gains when Zipf-like (1) rank popularity
  • Significant gains when more uniform (10)
  • NDC fails for larges sizes (6-SLA still works well)
slide-66
SLIDE 66

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Modest gains when Zipf-like (1) rank popularity
  • Significant gains when more uniform (10)
  • NDC fails for larges sizes (6-SLA still works well)
slide-67
SLIDE 67

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Modest gains when Zipf-like (1) rank popularity
  • Significant gains when more uniform (10)
  • NDC fails for larges sizes (6-SLA still works well)
slide-68
SLIDE 68

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Modest gains when Zipf-like (1) rank popularity
  • Significant gains when more uniform (10)
  • NDC fails for larges sizes (6-SLA still works well)
slide-69
SLIDE 69

Trace-based analysis (Synthetic)

  • Normalized traffic savings
  • Modest gains when Zipf-like (1) rank popularity
  • Significant gains when more uniform (10)
  • NDC fails for larges sizes (6-SLA still works well)
slide-70
SLIDE 70

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
slide-71
SLIDE 71

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
  • NDC fails; 3-SLA works reasonably well
  • Dynamic allocation with k-SLA outperform LRU by far
slide-72
SLIDE 72

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
  • NDC fails; 3-SLA works reasonably well
  • Dynamic allocation with k-SLA outperform LRU by far

Normalize against offline “global knowledge” policy that stores most popular files

slide-73
SLIDE 73

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
  • NDC fails; 3-SLA works reasonably well
  • Dynamic allocation with k-SLA outperform LRU by far

Good Bad

slide-74
SLIDE 74

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
  • NDC fails; 3-SLA works reasonably well
  • Dynamic allocation with k-SLA outperform LRU by far
slide-75
SLIDE 75

Trace-based Analysis

  • Spotify traces (all requests for 1M random tracks; 1 week)
  • Prediction policies: (i) “oracle”, (ii) 24h, (iii) interval average
  • NDC fails; 3-SLA works reasonably well
  • Dynamic allocation with k-SLA outperform LRU by far
slide-76
SLIDE 76

Dynamic Content Allocation Problem

76

  • Finite horizon dynamic decision problem
  • Discrete mean-value approximation
  • Exact solution as MILP
  • Computationally feasible approximations

(e.g., k-SLA) with performance bounds

  • Validate model and policies using traces

from Spotify

slide-77
SLIDE 77

Thank you!

Niklas Carlsson (niklas.carlsson@liu.se)

www.ida.liu.se/~nikca/papers/infocom14.pdf

Dynamic Content Allocation for Cloud- assisted Service of Periodic Workloads

György Dan (KTH) and Niklas Carlsson (LiU)