Dynamic_Cloud_Application_Redistribution_Performance_Optimization

ResearchResearch

University of StuttgartUniversitätsstr. 3870569 StuttgartGermany

Phone +49-711-685 88337 Fax +49-711-685 88472

Santiago Gómez Sáez, Vasilios Andrikopoulos, Frank Leymann, and Steve StrauchInstitute of Architecture of Application Systems

{gomez-saez, andrikopoulos, leymann, strauch}@iaas.uni-stuttgart.de

Towards Dynamic Application Distribution Support for Performance

Optimization in the Cloud

IEEE CLOUD 2014

Research

© Santiago Gómez Sáez 2

Agenda

Motivation Experiments

Methodology & Setup Application Persistence Analysis & Evaluation

Performance-aware Application (Re-)Distribution Process

Conclusion & Future Work

33© Santiago Gómez Sáez

Research

Motivation – Efficient Application Distribution

WebShop: WAR

Apache_Tomcat:Servlet_Container

Ubuntu10.04:Virt_Linux_OS

IBM_Server:Physical_Server

Product_DB:SQL_DB

MySQL: SQL_RDBMS_Server

AWS_EC2_m1.xlarge:

AWS_EC2



AWS_RDS_mediumDB: AWS_RDS

MySQL: SQL_DBaaS

AWS_EC2_m1.medium:

AWS_EC2


AWS_Elastic_BeansTalk: Application_Container

Partial vs. Complete Migration Multiple deployment options Multi-dimensional & Evolving problem Application workload behavior fluctuations Resources Demands Evolution

Performance-aware Information

WL Specification

alt_hosted_onhosted_on

interacts-with

Andrikopoulos et al.: Optimal Distribution of Applications in the Cloud. In: Proceedings of CAiSE’14


Research

Motivation - Perspectives & Approaches

Performance-aware Specification

Workload Model Derivation & Characterization

Application Workload Evolution Monitoring

Top-down Bottom-up


Research

App. Persistence Experiments – Methodology & Setup (1)

Evaluate the application persistence layer performance Under different deployment scenarios For different workload characteristics Towards maximizing its performance

Emulated a three-layered application Consolidate the top-down and bottom-up performance analysis

approaches over time Schema & Empirical analysis Derive the workload behavior model Characterize the different operations which constitute the

workload Generate workloads with different characteristics Performance Analysis


Research

WebShop: WAR

…


IBM_Server:Physical_Server

Product_DB:SQL_DB


AWS_EC2_m1.xlarge:

AWS_EC2


AWS_RDS_xlargeDB: AWS_RDS

MySQL: SQL_DBaaS

alt_hosted_onhosted_on

WL Specification


FlexiScale4vCPU:FlexiScale_VM

on-premise

Interacts-with

IaaS DBaaS



Research


Using the TPC-H Benchmark workload & data Initial workload -> 23 SQL (SELECT) queries 1 GB data

Apache JMeter 2.9 as the load driver Measurements & Rounds

Throughput (Req./s) 10 Rounds/day during one month (Q4 2013)

Characterization of the Workload items Generated multiple workloads with different characteristics (CL, CM, and

CH) Probability of occurrences of queries QCL, QCM, and QCH

Analyze the variation with respect to the initial workload behavior model

Analyzing the performance improvement/degradation for the persistence layer deployment alternatives


Research

App. Persistence Experiments – Sample Query

Table joints Subqueries Embedded operations Conditional selection Ordering & Limit


Research

App. Persistence Analysis – Workload Analysis


Research

App. Persistence Analysis – Workload Characterization


Research

App. Persistence Analysis – Performance Improvement


Research

Synthetic Workload Generation


Research

Generated Workload Analysis – Cumulative Distribution Fit

(i) On-premise (ii) DBaaS (iii) IaaS


Research

Generated Workload – Alternative Topologies Evaluation


Research

Participants Application Developer

Application design and realization Application topology specification, e.g. using TOSCA

Application Cloud Distribution Design Support System Application (Re-)distribution towards proactively react to fluctuating workloads

Topology+ : enriched application topology model with performance awareness (e.g. expected throughput/operation or

component, resource consumption) application workload characteristics (e.g. probability matrix of operations)

Topology*: viable topology describing application distribution alternatives and specifying Cloud services dynamic resource adaptation configurations

Approach – Performance Aware Application (Re-)Distribution Process


Research

Model Application

Topology

Specify Performance-

awareness

Discover Application Distribution

Evaluate Application &

Distribute

Register & Monitor

Performance

Re-distribution

• Requirements• Capabilities• Constraints

• Expected Performance• Workload Behavior Specification• Topology+

• Workload Model Derivation• Cloud Offerings Matching• Similarity & utility-based Analysis• Topologies*

• Topology* instance• Synthetic Workload Generation• Distribution Performance Evaluation

• Demanded vs. Provided Performance• Workload behavior Patterns• Statistical Classification

Collaborative Loop• Proactiveness• Workload Fluctuation• Application Performance Evolution• Optimize performance vs. cost

tradeoff

ResourcesAdaptation

Application Developer Distribution Support System


Research

Conclusions

Collaboration Loop Experiments focusing on the Application persistence

Deriving the workload model Generating workloads with different characteristics Multiple deployment scenarios

Process-based approach to distribute the application towards optimizing its performance Functional & Non-functional aspects Identifying the need to enrich the topology Focusing on the application workload behavior evolution


Research

Future Work

Reuse existing tools & implement new ones of the proposed tool chain

Performance-aware specification language Evaluate the performance of the overall process Experiments in the application upper layers

Santiago Gómez SáezE-mail: [email protected] of Architecture of Applications Systems (IAAS)University of Stuttgart (Germany)

mailto:[email protected]

19

Research

Thanks for your attention!!

Dynamic_Cloud_Application_Redistribution_Performance_Optimization

Technology