YexiJiang Florida International UniversityAdvisor Dr Tao LiCollaborator Dr Charles Perng Dr RongChangPresentation OutlineBackgroundCloud Capacity PredictionPredict provisioning resource demandEstimat ID: 875502
Download Pdf The PPT/PDF document "Cloud Analytics for Capacity Planning an..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1 Cloud Analytics for Capacity Planning an
Cloud Analytics for Capacity Planning and Instant VM Provisioning YexiJiang Florida International UniversityAdvisor: Dr. Tao LiCollaborator: Dr. Charles Perng, Dr. RongChang Presentation Outline BackgroundCloud Capacity PredictionPredict provisioning resource demandEstimate de
2 provisioning requestsExperimental evalua
provisioning requestsExperimental evaluation resultsInstant Cloud ProvisioningPredict VM provisioning demandExperimental evaluation results��1YexiJiang http://users.cis.fiu.edu/~yjian004/ Background What is Cloud AnalyticsRapidly identify cloud resource o
3 r application trouble spots so you can s
r application trouble spots so you can solve the problem. What is the objective of cloud analytics? The cloud platform itself. What can cloud analytics do?Workload analysisSystem fault diagnostics��2YexiJiang http://users.cis.fiu.edu/~yjian004/ Smart Clo
4 ud Enterprise trace data 5 month, 35k+ r
ud Enterprise trace data 5 month, 35k+ requests, 120+ image types, 20+ features each record Important Features: Image Name, Owner, Start Time, End Time, ID ��3YexiJiang http://users.cis.fiu.edu/~yjian004/ Aggregating the Raw Data weeklydailyhourly Cannot
5 reflect real capacity Just right �
reflect real capacity Just right ��4YexiJiang http://users.cis.fiu.edu/~yjian004/ Aggregating the Raw Data MeasurementWeeklyDailyHourlyCoefficient of Variance (CV)0.56060.79151.2249Skewness0.32951.56445.4464Kurtosis1.625.884852.4103weeklydailyhourly Cann
6 ot reflect real capacity Just right Too
ot reflect real capacity Just right Too irregular ��5YexiJiang http://users.cis.fiu.edu/~yjian004/ Presentation Outline BackgroundCloud Capacity PredictionPredict provisioning resource demandEstimate deprovisioning requestsExperimental evaluation results
7 Instant Cloud ProvisioningPredict VM pro
Instant Cloud ProvisioningPredict VM provisioning demandExperimental evaluation results��6YexiJiang http://users.cis.fiu.edu/~yjian004/ Cost of Data Centers 31% of the cost is related to power.As hardware price continuously decreases, the proportion would
8 further increase. The US EPA estimates
further increase. The US EPA estimates the energy usage at data centers is experiencing successive doubling every five years. (7.4 billion in 2011) * From James Hamilton's Blog ��7YexiJiang http://users.cis.fiu.edu/~yjian004/ Motivation Reduce power cost
9 via capacity predictionCost of the Clou
via capacity predictionCost of the Cloud Provider Prepared Resource Real Requirement ��8YexiJiang http://users.cis.fiu.edu/~yjian004/ MotivationReduce power cost via capacity predictionCost of the Cloud Provider Prepared Resource Predicted Resource Real
10 Requirement ��9YexiJiang
Requirement ��9YexiJiang http://users.cis.fiu.edu/~yjian004/ Candidate Time Series Capacity time seriesNonstationary. Difficult to model directly Provisioning /deprovisioning time seriesObvious temporal pattern Better candidate ��10YexiJiang
11 http://users.cis.fiu.edu/~yjia
http://users.cis.fiu.edu/~yjian004/ Basic Idea Capacity = (# existing VMs) + (# provisioning) (# deprovisioning) Existing VM in cloud -+ Predicted Provisioning Predicted De - provisioning Predicted Capacity ��11YexiJiang http://users.cis.fiu.edu
12 /~yjian004/ Predicting Provisioning Dema
/~yjian004/ Predicting Provisioning DemandsEnsemble method for time series predictionIndividual prediction techniques used:Moving Average. Naïve predictor.Auto Regression. Linear predictor.Neural Network. Nonlinear predictor.Gene Expression Programming. Genetic algorithm.Suppo
13 rt Vector Machine. Linear predictor with
rt Vector Machine. Linear predictor with nonlinear kernel. Dynamic weighted linear combination Weight update (t)weight of predictor ppredicted value of individual predictor p(t)cost of predictor p at time t(t)error of individual predictor p��12YexiJiang h
14 ttp://users.cis.fiu.edu/~yjian004/ Cloud
ttp://users.cis.fiu.edu/~yjian004/ Cloud Prediction Cost Overprediction: cost of resource waste. function:Underprediction: cost of SLA penalty. function:Property: Nonnegative, Monotonic. ��13YexiJiang http://users.cis.fiu.edu/~yjian004/ ))(~),(())(~),((tv
15 tvTtvtvRC Prediction Result Ensemblehas
tvTtvtvRC Prediction Result Ensemblehas the best average performance.��14YexiJiang http://users.cis.fiu.edu/~yjian004/ Predicting DeprovisioningUse the life span CDF F(x)of VMs to estimate number of deprovisioning requests Estimation of distribution: st
16 epwise function. # of VMs with life span
epwise function. # of VMs with life span t (t1 t t2)&#x-200;&#x-200;15YexiJiang http://users.cis.fiu.edu/~yjian004/ provisioning evaluation Test data: last 60 day. Test methods:No preparation at all (None)Always prepare the maximum capacity (Maximum)Time series predi
17 ction (Time Series)Life span distributio
ction (Time Series)Life span distribution despite of image60 days of data (Dist 60)90 days of data (Dist 90) Global distribution estimation method outperforms the time series prediction method. ��16YexiJiang http://users.cis.fiu.edu/~yjian004/ Presentatio
18 n Outline BackgroundCloud Capacity Predi
n Outline BackgroundCloud Capacity PredictionPredict provisioning resource demandEstimate deprovisioning requestsExperimental evaluation resultsInstant Cloud ProvisioningPredict VM provisioning demandExperimental evaluation results��17YexiJiang http://use
19 rs.cis.fiu.edu/~yjian004/ Motivation Pro
rs.cis.fiu.edu/~yjian004/ Motivation Problem: Existing clouds are not “instant”, not suitable for midjob scaling and urgent tasks. VM preparation is fast, but patching, security assurance, manual process and other processes cost time. Known solutionsPrepare extreme l
20 arge number of different types of VMs. W
arge number of different types of VMs. Waste resource Ask customers to provide schedule. Impractical Our Idea: Make good use of the customer historical requests to infer the future demand. Reduce the average VM provisioning fulfillment time.��18YexiJiang
21 http://users.cis.fiu.edu/~yjian004/ Core
http://users.cis.fiu.edu/~yjian004/ Core Idea Model and predict d emands Predict Results Pre - provision at suitable time Wait for Requests Assign VMs to customers ��19YexiJiang http://users.cis.fiu.edu/~yjian004/ Focus on individual types No obvio
22 us temporal patterns for individual imag
us temporal patterns for individual image type. Ensemble is still required. ��20YexiJiang http://users.cis.fiu.edu/~yjian004/ Focus on popular VM types About 10% (12) of the 124 VM types consists more than 80% requestsInflection point divides the VM types
23 into popular group and rare group Requ
into popular group and rare group Requests for rare image types appear randomly. ��21YexiJiang http://users.cis.fiu.edu/~yjian004/ Workflow Overview ��22YexiJiang http://users.cis.fiu.edu/~yjian004/ Experimental Evaluation Ensembl
24 e method have the best performance in re
e method have the best performance in reducing waiting time and resource waste. ��23YexiJiang http://users.cis.fiu.edu/~yjian004/ ConclusionCapacity PredictionThe demand of cloud capacity can be estimated by predicting provisioning and deprovisioning requ
25 estsUse time series ensemble method for
estsUse time series ensemble method for provisioning predictionUse VM life span model for deprovisioning predictionInstant cloud provisioningPreprovision VMs before requests arrivePredict VM provision requests use time series ensemble methodThe average provisioning fulfillment
26 time can be reduced by 85%+ Future workI
time can be reduced by 85%+ Future workImprove prediction with user profileFinegrain adjustment with control theory��24YexiJiang http://users.cis.fiu.edu/~yjian004/ Thank you!��25YexiJiang http://users.cis.fiu.edu/~yjian004/ Thank y
27 ou Related Paper:Intelligent Cloud Capac
ou Related Paper:Intelligent Cloud Capacity Management. (NOMS 2012)ASAP: A SelfAdaptive Prediction System for Instant Cloud Resource Demand Provisioning. (ICDM 2011)Patent:Cloud Provisioning Accelerator, Serial # 13306506, Pending��26YexiJiang http://us