Seagull: Intelligent Cloud Bursting For Enterprise Applications Tian Guo , Upendra Sharma, Timothy Wood‡, Sambit Sahu†,Prashant Shenoy University of Massachusetts Amherst, The George Washington University ‡ , IBM Research † U NIVERSITY OF M ASSACHUSETTS , A MHERST • Department of Computer Science
Cloud Computing § Cloud Computing: Pay-as-you-go service • Rent Resources • Infrastructure as a Service • - Virtualization technology, rent VMs - Popular for Apps with dynamic workload 10 § Benefits 8 Sever Needed Flexible pricing model • 6 Agile to workload changes 4 • 2 P r i v a te S i te ( 5 s e r v e r s ) 0 Mon Tues Wed Th u r Sa t Sun Fri Tian Guo(tian@cs.umass.edu) 2
Cloud Bursting § Enterprises own private data centers Try to use the existing infrastructure (hybrid) • § Cloud Bursting Enables Enterprise to use local data center • rents public resource upon workload changes • seamless and transparent resource sharing between local and • public cloud § Challenges 10 Cloud Burst Servers Needed 8 When to trigger cloud bursting? • 6 Which Apps to cloud Burst? • 4 How to balance cost and time trade-off? 2 • Private Site (5 servers) 0 § Seagull Mon Tues Wed Thur Fri Sat Sun Cloud Bursting Algorithm • Precopying Algorithm • Tian Guo(tian@cs.umass.edu) 3
Seagull Cloud Bursting Algorithm 1 1 1 1 2 2 1 2 2 1 1 1 1 1 1 Naive Seagull Seagull 2 Seagull 1 Naive Naive 1 2 1 1 1 vcpu 1 1 vcpu 1 vcpu 1 1 1 1 vcpu 1 vcpu 1 vcpu 2 2 2 2 2 vcpu 2 2 2 vcpu 2 2 2 vcpu 2 vcpu 2 vcpu 2 vcpu § Which applications to cloud burst? Naive approach: move overloaded applications • • Incurs high cost and overhead Seagull approach: Pick the cheapest applications • • Multi-resource bin packing problem • Greedy approach • Metric: App_Costs/ VM_cores to run in public cloud Tian Guo(tian@cs.umass.edu) 4
How to Lower Migration Time ? § Cloud bursting on demand e.g 5 GB disk state, takes a long time ( ~22 mins) • § Opportunistic Precopying Copys app vm state to the public cloud in the background • Benefit: Dramatically shortens the migration time • Some experiments: • ~120 secs Tian Guo(tian@cs.umass.edu) 5
Seagull Precopying Algorithm 1 1 1 1 1 1 1 1 1 1 1 1 Naive Naive 1 Seagull 1 1 1 Seagull 1 vcpu 1 1 vcpu 1 1 vcpu 1 2 1 vcpu 2 vcpu 2 2 vcpu 2 2 vcpu 2 2 vcpu § How to balance cost and time trade-off? Naive Precopying: Precopying overloaded applications • - Not Necessary lower migration time Intelligent Precopying • - Intuition: Choose the apps that are most likely to be migrated Tian Guo(tian@cs.umass.edu) 6
Cloud Bursting Algorithm Evaluation § Experiment Setup 3 hosts and 5 Apps • Varying workload of A for 4 hours • § Seagull is cost Efficient Lowers cost by 25% over 4 hours • 25% saving Tian Guo(tian@cs.umass.edu) 7
Precopying Algorithm Evaluation § Experiment Setup Emulation with 200 quad-core hosts • 40 applications, 30% were overloaded • Precopying freqency: 1 hr & total time: 24 hrs • § Seagull balances time and cost well Spends 22% more money • Transmits 95% less data • 95% saving 22% more Tian Guo(tian@cs.umass.edu) 8
Summary § Cloud Bursting Hybrid solution for dynamic workload • Good for Enterprises with private data centers • § Seagull: Intelligent and automated Cloud Bursting Determines which Apps to Cloud Burst • - Lowers Cost by 25% Determines which Apps to Precopy • - Saves 95% Data Transmission Tian Guo(tian@cs.umass.edu) 9
Recommend
More recommend