Dynamic Content Allocation for Cloudassisted Service of Periodic Workloads

Transcription

1 Dynamic Content Allocation for Cloudassisted Service of Periodic Workloads György Dán Royal Institute of Technology (KTH) Niklas Carlsson Linköping IEEE INFOCOM 2014, Toronto, Canada, April/May 2014

2 Internet Content Delivery Large amounts of data with varying popularity Multi-billion market ($8B to $20B, ) 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.

3 Internet Content Delivery Large amounts of data with varying popularity Multi-billion market ($8B to $20B, ) 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.

4 Internet Content Delivery Large amounts of data with varying popularity Multi-billion market ($8B to $20B, ) 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.

5 Internet Content Delivery Large amounts of data with varying popularity Multi-billion market ($8B to $20B, ) 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.

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

7 Content Delivery Cloud-based delivery Dedicated infrastructure cloud servers

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

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

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

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

12 Content Delivery Cloud-based delivery Flexible computation, storage, and bandwidth Pay per volume and access Dedicated infrastructure Limited storage Cloud bandwidth elastic; however, flexible comes at premium Capped unmetered bandwidth Potentially closer to the user

13 High-level problem Minimize content delivery costs Bandwidth Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $

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?

15 High-level problem Minimize content delivery costs Bandwidth Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $ How to get the best out of two worlds?

16 High-level problem Minimize content delivery costs Bandwidth Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $ How to get the best out of two worlds? Improved workload models and prediction enables prefetching

17 High-level problem Minimize content delivery costs Bandwidth Cost Cloud-based Elastic/flexible $$$ Dedicated servers Capped $ 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

18 Dynamic Content Allocation Problem 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 18

19 Cost minimization formulation

20 Cost minimization formulation Total demand

21 Cost minimization formulation Demand of files in capped BW storage

22 Cost minimization formulation Capped BW limit (U)

23 Cost minimization formulation

24 Cost minimization formulation Served from capped BW storage

25 Cost minimization formulation Served using elastic cloud resources

26 Cost minimization formulation Traffic due to allocation

27 Cost minimization formulation

28 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

29 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

30 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

31 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

32 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

33 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

34 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

35 Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Total expected cost Optimal policy

36 Utilization maximization Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Equivalent formulation Total expected cost Optimal policy

37 Utilization maximization Cost minimization formulation Traffic of files only in cloud Spillover traffic Traffic due to allocation Equivalent formulation Total expected cost Optimal policy

38 Utilization maximization Cost minimization formulation Equivalent formulation

39 Utilization maximization Cost minimization formulation Two file example Equivalent formulation

40 Utilization maximization Cost minimization formulation Two file example Equivalent formulation

41 Utilization maximization Cost minimization formulation Two file example Equivalent formulation

42 Utilization maximization Cost minimization formulation Two file example Equivalent formulation

43 Utilization maximization Cost minimization formulation Two file example Equivalent formulation

44 Utilization maximization Cost minimization formulation Equivalent formulation

45 Discrete-time Decision Problem Equivalent formulation

46 Discrete-time Decision Problem Approximation decrease exponentially Finite horizon decision problem Equivalent formulation

47 Discrete-time Decision Problem Approximation decrease exponentially Finite horizon decision problem Equivalent formulation

48 Discrete-time Decision Problem Approximation decrease exponentially Finite horizon decision problem Equivalent formulation

49 Discrete-time Decision Problem Approximation decrease exponentially Finite horizon decision problem

50 Discrete-time Decision Problem Approximation decrease exponentially Finite horizon decision problem Theorem: Exact solution as a MILP

51 Policy: No Download Cost (NDC) Consider next interval only

52 Policy: No Download Cost (NDC) Consider next interval only Proposition 1: Unbounded approximation ratio Proposition 2: Approximation bound

53 Policy: No Download Cost (NDC) Consider next interval only Proposition 1: Unbounded approximation ratio Proposition 2: Approximation bound

54 Policy: No Download Cost (NDC) Consider next interval only Proposition 1: Unbounded approximation ratio Proposition 2: Approximation bound

55 Policy: No Download Cost (NDC) Consider next interval only Proposition 1: Unbounded approximation ratio Proposition 2: Approximation bound

56 Policy: k-step Look Ahead (k-sla) Consider k next intervals

57 Policy: k-step Look Ahead (k-sla) Consider k next intervals Proposition 3: Unbounded approximation ratio Proposition 4: Approximation bound

58 Policy: k-step Look Ahead (k-sla) Consider k next intervals Proposition 3: Unbounded approximation ratio Proposition 4: Approximation bound

59 Policy: k-step Look Ahead (k-sla) Consider k next intervals Proposition 3: Unbounded approximation ratio Proposition 4: Approximation bound

60 Policy: k-step Look Ahead (k-sla) Consider k next intervals Proposition 3: Unbounded approximation ratio Proposition 4: Approximation bound

61 Trace-based analysis (Synthetic) Normalized traffic savings Based on Spotify trace characterization 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 (B min, B max, )

62 Trace-based analysis (Synthetic) Normalized traffic savings Normalize against policy that stores the most popular files

63 Trace-based analysis (Synthetic) Normalized traffic savings

64 Trace-based analysis (Synthetic) Normalized traffic savings Good

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)

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)

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)

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)

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)

70 Trace-based Analysis Spotify traces (all requests for 1M random tracks; 1 week) Prediction policies: (i) oracle, (ii) 24h, (iii) interval average

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

72 Trace-based Analysis Spotify traces (all requests for 1M random tracks; 1 week) Prediction policies: (i) oracle, (ii) 24h, (iii) interval average Normalize against offline global knowledge policy that stores most popular files NDC fails; 3-SLA works reasonably well Dynamic allocation with k-sla outperform LRU by far

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

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

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

76 Dynamic Content Allocation Problem 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 76

77 Dynamic Content Allocation for Cloudassisted Service of Periodic Workloads György Dan (KTH) and Niklas Carlsson (LiU) Thank you! Niklas Carlsson