Dynamic Cloud Resource Reservation via Cloud Brokerage

Transcription

1 Dynamic Cloud Resource Reservation via Cloud Brokerage Wei Wang*, Di Niu +, Baochun Li*, Ben Liang* * Department of Electrical and Computer Engineering, University of Toronto + Department of Electrical and Computer Engineering, University of Alberta July 10, 2013

2 Growing Cloud Computing Costs Drastic increase in enterprise spending on Infrastructure-as-a- Service (IaaS) clouds 41.7% annual growth rate by 2016 [CloudTimes 12] IaaS cloud will be the fastest-growing segment among all cloud services 2

3 Tradeoffs in Cloud Pricing Options On-demand instances No commitment Pay-as-you-go Linux/UNIX Usage Windows Usage Standard On-Demand Instances Small (Default) $0.080 per Hour $0.115 per Hour Medium $0.160 per Hour $0.230 per Hour Large $0.320 per Hour $0.460 per Hour Extra Large $0.640 per Hour $0.920 per Hour Reserved instances Reservation fee + discounted price Suitable for long-term usage commitment 3

4 On-demand v.s. Reservation Pros Cons On-demand 1. Flexible 2. Fits sporadic workload Expensive for longterm usage Reservation Cost efficient for longterm usage 1. Long-term usage commitment 2. Expensive for sporadic workload 4

5 User s Problem Hard to choose among different pricing options Lacks sufficient expertise Cost savings due to the reservation option are not always possible Depends on the user s own demand pattern Must be long-term and heavy usage 5

6 Can we go beyond the limitation of demand pattern of a single user and lower the cost?

7 A Cloud Brokerage Service A cloud broker reserves a large pool of instances Users purchase instances from the broker in an on-demand fashion User IaaS Cloud Providers Reserved/On-demand Instances Broker Broker cost "On-demand" Instances User cost User User Figure 1: The proposed cloud broker. Solid arrows show the 7

8 Why cloud broker?

9 Better Exploiting Reservation Options Statistical multiplexing increases the utilization of reserved instances Aggregating all users demands smoothes out the bursts A flat demand curve is more friendly to reserved instances The true cost of reserved instance is reduced due to the increased instance utilization 9

10 Reducing Wasted Cost Due to Partial Usage Alleviate the pricing inefficiency of on-demand instances Partial usage is counted as a full billing cycle The broker can time-multiplex partial usage Billing cycle (an instance-hour) Without broker User 1 User 1 User 2 Instance 1 Instance 2 With broker User 1 User 2 User 1 10

11 Enjoying Volume Discounts Most IaaS clouds offer significant volume discounts Amazon provides 20% or even higher volume discounts in EC2 The sheer volume of the aggregated demand makes cloud broker easily qualify for such discounts 11

12 A Win-Win Solution Users receive a lower price when trading with the broker No upfront payment for reservation No money wasted on idled reservation instances Broker makes profit by leveraging the wholesale (reservation) model A significant price gap between on-demand and reserved instances Aggregate demand is more amenable to the reservation option 12

13 How many instances should a broker reserve?

14 On-demand and Reserved Pricing On-demand instances Fixed hourly rate p Reserved instances Upfront reservation fee: Reservation period: Instances reserved at time t: r t # of reserved instances that are effective at time t, the number of res n t = P t i=t +1 r i. may not be sufficient 14

15 Dynamic Resource Reservation Cloud users submit demand predictions to the broker ycle. Broker Henc reserves instances based on the aggregate demand d 1,...,d T Total cost = Reservation cost + On-demand cost where, P T t=1 r t + P T t=1 (d t n t ) + p, the number of res n t = P t i=t +1 r i. may not be sufficient # of reserved instances that are effective at t 15

16 The Cost Minimization Problem Make dynamic reservation ycle. decisions Henc accommodate demands d 1,...,d T r 1,...,r T to min {r 1,...,r T } cost = P T t=1 r t + P T t=1 (d t n t ) + p This is an integer program! 16

17 Optimal Solution: Dynamic Programming

18 The Curse of Dimensionality High dimensional dynamic programming High dimensional state: x i : # of instances that are reserved no later than t and remain effective at t+i er t Exponential time and space complexity The curse of dimensionality s t := (t, x 1,...,x 1 ), that are reserved 18

19 Approximate Solution

20 A 2-Competitive Heuristic Segment the demand into intervals each spanning one reservation period Interval Demand Time Make optimal instance reservation decisions per interval 20

21 Demand Optimal Instance Reservation within an Interval Stratify demand into levels For each level, decide if a reserved instance should be used Example On-demand rate: $1 per hour Reservation: $2.5 for 6 hours Should reserve when instance usage >= 3 hours L Time (hour) L5 L4 L3 L2 21

22 Cost Performance Per-interval reservation is 2-competitive Incurs at most twice the optimal cost in the worst case Interval Demand Time All reservations are made at the beginning of the interval 22

23 An Improved Greedy Algorithm Do not segment demand into intervals Stratify demands into levels Demand Make reservations top-down L6 L5 L4 L3 L2 L1 Time At each level, apply dynamic programming V l (t) =min{v l (t )+,V l (t 1) + c l (t)} Strictly better than Per-Interval Reservation, and is also 2- competitive 23

24 When demand predictions are unavailable

25 Online Algorithm Make instance reservation decisions without future information Algorithm 3 Online Reservation Made at Time t 1. Let g i =(d i n i ) + for all i = t +1,...,t. 2. Run Algorithm 1 with g t +1,...,g t as the input demands. Let x be its output. 3. Reserve r t = x instances at time t. 4. Update n i = n i + r t for all i = t +1,...,t+ 1. The best that we can do [Wang et al. ICAC 13] 2-competitiveness for the deterministic online algorithm d Std High y = 5x 25 M

26 Trace-Driven Simulations

27 Dataset and Preprocessing Google cluster-usage traces 900+ users usage traces on a 12K-node Google datacenter We convert users computing demand data to IaaS instance demand Users are classified into 3 groups based on demand fluctuation level Standard deviation vs. mean in hourly demand 250 Demand Std High y = 5x Medium y = x Low Demand Mean Figure 8: Demand statistics and the division of users into 3 groups according to demand fluctuation level. 27

28 Demand Curve Highly fluctuated Medium fluctuation Relatively stable # Instances # Instances (k) # Instances User Time (hour) User Time (hour) User Time (hour) Figure 7: The demand curves of three typical users. 28

29 Aggregation Smoothes Out Demand Bursts y = 0.363x Demand Mean : medium fluctuation Demand Std y = 1.774x Demand Mean (a) Group 1: high fluctuation Demand Std y = 0.363x Demand Mean (b) Group 2: medium fluctuation Figure 300 9: Aggregation suppresses the 300demand fluctuation of individual users. Ea the demand fluctuation level (the ratio between the demand standard deviation an Demand Std Wasted instance hours (k) Before aggregation After aggregation y = 0.058x Demand Mean 400 (c) Group 3: low fluctuation % Demand Mean 5.6% 30.5% 200 and fluctuation of individual users. 16.5% Each circle represents a user. The line indicates ween the demand standard deviation 0 and mean) in the aggregate demand curve. High Medium Low All Demand Std y = 0.061x (d) All the users. each reservatio discount of 50 on-demand ins general pricing Group Divi tics of users, w deviation for e been mentione 29 vations critical

30 The Reduction of Partial Usage Figure 9: Aggregation suppresses the demand fluctuation of individu the demand fluctuation level (the ratio between the demand standard Wasted instance hours (k) 0 Before aggregation After aggregation 16.5% 30.5% 5.6% High Medium Low All Demand Fluctuation 23.4% Figure 10: Aggregation reduces the wasted instance-hours due to partial usage. disk, memory, etc. Instance Scheduling. We take such a dataset as input, and ask the question: How many computing instances would 30 ea dis on ge tic de be va flu sta thr rat de

31 Cost Savings Due to the Broker Saving Percentage (%) Heuristic Greedy Online High Medium Low All Demand Fluctuation No volume discount Cost w/ Broker (k $) y = x Cost w/o Broker (k $) (a) Group 2: medium fluctuation Cost w/ Broker (k $) y = x Cost w/o Broker (k $) (b) All the users Figure 15: Cost without the broker vs. with the broker for individual users, using Greedy strategy. Each circle is a user. 31

32 Conclusions We propose a smart cloud brokerage service Reserves a pool of instances to serve the aggregated demand Leverages the price gap between the wholesale and retail model to reap the profit while offering lower price to cloud users Cloud users purchase instances from the broker as if instances were offered on demand Design and analyze three instance reservation algorithms for the broker and evaluate them via trace-driven simulations More detailed analysis of online algorithms are given in our follow-up work [Wang et al. ICAC 13] 32

33 Thanks!