Downloaded 13 times
Uploaded byFei Dong
Decision support for Amazon Spot Instance
This document discusses using Amazon EC2 spot instances for cost-effective cloud computing. It presents challenges in optimizing costs while meeting performance needs. Various instance types and pricing models are examined, including predicting spot prices using statistical methods. An algorithm and bidding strategy are proposed to minimize costs based on deadline, budget and other constraints. An experimental evaluation compares performance and costs of spot instances versus on-demand instances.
Decision support for Amazon Spot Instance
- 1. Decision Support for Amazon EC2 Spot Instances Fei Dong 2011-‐11-‐28 11/27/11 1
- 2. A Glimpse of Amazon EC2 • Reserved Instance, On-‐demand Instance, and SI • Different scenarios: Cluster × Workload EC2 Node CPU Memory I/O Per-‐hour Type (#EC2 units) Performance Cost m1.small 1 1.7 GB moderate $0.085 m1.large 4 7.5 GB high $0.34 m1.xlarge 8 15 GB high $0.68 c1.medium 5 1.7 GB moderate $0.17 c1.xlarge 20 7 GB high $0.68 cc1.4xlarge 33.5 23 GB very high $1.60 11/27/11 2
- 3. MulU-‐objecUve Cluster Provisioning 1,200 Running Time 1,000 800 (min) 600 Actual 400 200 0 m1.small m1.large m1.xlarge c1.medium c1.xlarge 10.00 8.00 Cost ($) 6.00 4.00 Actual 2.00 0.00 m1.small m1.large m1.xlarge c1.medium c1.xlarge EC2 Instance Type for Target Cluster 11/27/11 3
- 4. Spot Instance • Spot instances enable you to bid for unused Amazon EC2 capacity. Instances are charged the Spot Price which is set by Amazon EC2 and fluctuates periodically depending on the supply of and demand for Spot Instance capacity. 11/27/11 4
- 5. Challenges & AssumpUons • Challenges: – Minimize monetary costs for a user while meeUng Service constrains. – Know nothing about Amazon pricing strategy and other bid strategy. • AssumpUons: – Bid price is fixed. – Instance Type is fixed (no mix strategy) – Not consider the overhead to recover spot instances. 11/27/11 5
- 6. Pricing PredicUon Model • Linear Regression • Normal DistribuUon • ExponenUal DistribuUon n ∑ p (1− p )i −1 (i ) H i =1 11/27/11 6
- 7. Predict Price Algorithm 1. Collect the prices over a period of Ume, in order to esUmate mean and variance. 2. Use the exponenUal approximaUon fidng, calculate x given the CDF(X<x) = Prob. 3. Compare other models and pick a maximum value as a bid price. 4. If the bid price is smaller than the spot price, thus increase the bid by 33% for the next interval. 11/27/11 7
- 8. Price PredicUon eu linux.m1.small spot price on 11/18/2011 0.35 0.3 0.25 0.2 Price ($) Actual 0.15 Predict Adjust Predict 0.1 0.05 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Time (Hours) Var(Predict Adj) = 0.000769 11/27/11 Var(Predict) = 0.001786 8
- 9. Bid Strategy Utility = F(d, b, e, t, n) Deadline d, budget b, Estimated Time e, copt = arg max F(d, b, e, t, n) c∈S Cluster Type t, Number n Min Time Mode: Can the job be execute as soon as possible under specified budget and deadline constrains? Min Money Mode: What is the bid price and instance type that minimize the total monetary cost? ExhausUve Search 11/27/11 9
- 10. Experimental EvaluaUon • Choose 5 Spot Instance Types – M1.small, m1.large, m1.xlarge, c1.medium, c1.xlarge • Run 5 Instances compared with on demand instances. 11/27/11 10
- 11. Experimental EvaluaUon (Ctd.) 1600 1400 Running Time (Min) 1200 1000 800 on-‐demand 600 SI budget intensive 400 SI Ume intensive 200 0 m1.small m1.large m1.xlarge c1.medium c1.xlarge EC2 Instance Type 14 12 10 Cost ($) 8 on-‐demand 6 SI budget intensive 4 SI Ume intensive 2 0 m1.small m1.large m1.xlarge c1.medium c1.xlarge EC2 Instance Type 11/27/11 11
- 12. Case Analysis M1.small Linux on 11/18/2011 M1.small spot instance, bid strategy 11/27/11 12
- 13. Conclusions & Future Work • Conclusions – More cost-efficient than fixed-size instance choice – Spot Instances not always provide inexpensive resources for transient workloads • Future works – Consider to mix other instance types (e.g. spot instances & reserved instances) – Disaster Recovery, checking point. 11/27/11 13
- 14. Reference • hjp://aws.amazon.com/ec2/instance-‐types • H. Herodotou, F. Dong, and S. Babu. No One (Cluster) Size Fits All: AutomaUc Cluster Sizing for Data-‐intensive AnalyUcs. (Slides)In Proc. of the ACM Symposium on Cloud CompuUng 2011 (SOCC '11), October 2011. • D. Ardagna, B. Panicucci and M.Passacantando. A Game TheoreUc FormulaUon of the Service Provisioning Problem in Cloud Systems. WWW2011 Proceedings, 2011 • N. Jain, I. Menache, and O. Shamir. On-‐demand or Spot? Learning-‐based Resource AllocaUon for Delay-‐Tolerant Batch CompuUng. 11/27/11 14
- 15.