Quality-Assured Cloud Bandwidth Auto-Scaling for Video-on-Demand Applications

Transcription

1 Quality-Assured Cloud Bandwidth Auto-Scaling for Video-on-Demand Applications Di Niu, Hong Xu, Baochun Li University of Toronto Shuqiao Zhao UUSee, Inc., Beijing, China 1

2 Applications in the Cloud WWW Netflix moved to Amazon Web Services in

3 Traditional Enterprise Operation Server Capacity Demand Days Over-provision 3

4 Auto Scaling in the Cloud # VMs CPU Demand Days Reduced cost: from infrastructure investment to metered billing (pay-as-you-go) 4

5 Available Bandwidth Bandwidth Demand Days No bandwidth guarantee! Unappealing to delay-sensitive applications Video-on-Demand, Gaming 5

6 Bandwidth Reservation is becoming feasible for traffic from a VM to the Internet H. Ballani, P. Costa, T. Karagiannis, and A. Rowstron, Towards Predictable Datacenter Networks ACM SIGCOMM 11 C. Guo et al. SecondNet: a Data Center Network Virtualization Architecture with Bandwidth Guarantees ACM CoNEXT 10 6

7 Can We AutoScale Bandwidth? A Naive Idea: if bandwidth utilization > 90%, reserve more bandwidth. Passive! No guarantee! Apps don t know demands! Individual reservations are costly! 7

8 Utilize the Data and Computing Power in the Cloud 8

9 From IaaS to Platform as a Service (PaaS) VoD Apps Payment Performance Guarantee Learning Cloud Provider Prediction Data Center Bandwidth Reservation Request Direction Data Center Predictive Quantitative Multiplex 9

10 Roadmap Tenant Payment Performance Guarantee 1 Cloud Provider Demand Prediction Data Center 2 Request Direction Resource Reservation Data Center 10

11 Data Mining UUSee: a VoD provider based in China 200+ GB traces, 21 days reports of online users every 10 minutes Bandwidth demand in each video channel 11

12 1000 A Typical Video Channel Bandwidth (Mbps) Time periods (1 period = 10 minutes) 12

13 Prediction based on Learning Predict expected demand Estimate demand variation 13

14 Seasonal ARIMA Models 500 Original data 10-minute-ahead prediction Bandwidth (Mbps) Time periods (1 period = 10 minutes) 14

15 GARCH modeling of volatility Prediction error Estimated conditional error standard deviation Mbps Time periods (1 period = 10 minutes) 15

16 Prediction Results Expected demand of each tenant μi Demand standard deviation σi Demand covariance matrix Σ = [σij] 16

17 Roadmap Tenant Payment Performance Guarantee 1 Cloud Provider Demand Prediction μi Σ Data Center 2 Request Direction Data Center Bandwidth Reservation 17

18 Individual Reservation Tenant i : random demand Di Tenant i wants to reserve Ri bandwidth s.t. Pr(D i >R i ) < If Di is Gaussian, then Small constant Service Level Agreement R i = µ i + θ()σ i A constant 18

19 Mixing anti-correlated demands saves bandwidth reservation 19

20 .%/%01%*'!#2#3,-4 A 1.%/%01%*'!#2#3,-4 A 2 )#$*+,*-"!"#$$%&'(!"#$$%&'9 < :,;% (<'=,$>-%/ < :,;% (<'=,$>-%/ Asum < :,;% (<'=,$>-%/ 20

21 Statistical Bin Packing (Integer Programming) M. Wang, X. Meng, and L. Zhang, Consolidating Virtual Machines with Dynamic Bandwidth Demand in Data Centers INFOCOM 2011 Mini-Conference A. Epstein, D. Breitgand Improving Consolidation of Virtual Machines with Risk-aware Bandwidth Oversubscription in Compute Clouds INFOCOM 2012 Mini-Conference 21

22 Reservation A 1 Data Center Request Direction (Fractional in [0,1]) w 11 w 12 Tenant 1 Random Demand D 1 A 2 Data Center w 21 w 22 Tenant 2 D 2 Problem Formulation: min s A s Total bandwidth reservation Capacity of datacenter s s.t. A s C s, Pr(L s >A s ), s w si =1. Load of datacenter s: L s = i w sid i All Di served 22

23 Bandwidth reservation minimization min s A s s.t. A s C s, Second order cone programming Pr(L s >A s ), s w si =1. For Gaussian demands Pr(L s >A s ) A constant E[L s ]+θ() var[l s ] A s E[L s ]= i w si µ i Expected demand i var[l s ]= i,j σ ij w si w sj covariance between i and j 23

24 Theorem 1: Optimal Solution (Closed Form) Tenant 1 Tenant 2 Tenant 3 Cloud Provider MIX Data Center Data Center Problem: each video is replicated to every data center! 24

25 Suboptimal Solutions Per-DC Optimal min s A s s.t. A s C s, Pr(L s >A s ), s w si =1. Solve for each s one by one mina s s.t. A s C s, Pr(L s >A s ) Per-DC Limited Channels Add channel number constraint per data center: Integer Programming; Heuristic solutions 25

26 Trace-Driven Simulations 26

27 1000 Replication Degree Reserved Bandwidth (Gbps) 10 Replication Degree % saving Gbps 1 Optimal Per-DC Opt. Per-DC Lim. Random Individual 7 35 data centers; Each capacity 300 Mbps; 176 channels (aggregated); Peak demand: 7.55 Gbps; Mean demand: 5.62 Gbps; ε = 2% 27

28 7.00 # Shortage Channels Over-Provision Ratio (%) 130% Shortage Channels % 115% 107.5% 0 Optimal Per-DC Opt. Per-DC Lim. Random 100% 35 data centers; Each capacity 300 Mbps; 176 channels (aggregated); Peak demand: 7.55 Gbps; Mean demand: 5.62 Gbps; ε = 2% 28

29 Conclusions Use smart data and prediction for cloud workload management Probabilistic QoS guarantee and statistical multiplexing Similar to financial risk management Pricing this guaranteed service 29

30 Thank you! Google Di Niu 30