Copyright © 2009, Telcordia Technologies, Inc. All Rights Reserved. 1 / 16 Model-based SOA Performance Profiling using CloudLauncher Other project members

Copyright © 2009, Telcordia Technologies, Inc. All Rights Reserved. 1 / 16

Model-based SOA Performance Profiling using CloudLauncher

Other project members

Francesco CarusoJosephine Micallef

Sumant Tambe,[email protected] University, Nashville TNSummer Intern @ Telcordia Technologies , June-Aug 2009


Outline

Background: Service-Oriented Architecture (SOA) and Quality of Service (QoS)

QoS Design Problem in SOA Overview of State of the Art Project Objectives Solution: SOA Dynamic Designer for QoS

A model-driven deployment and performance profiling tool for SOA QoS design: CloudLauncher

Amazon Elastic Compute Cloud (EC2) Testbed

Concluding remarks


Service-Oriented Architecture (SOA) & QoS

Imp

lem

en

tatio

n la

yer

Se

rvic

es

inte

rfa

ce la

yer

Bu

sin

ess

p

roce

ss la

yer

..NET J2EE Legacy

composite services

atomic services

E2E QoS (SLA)

Service QoS

Environment QoS

SOA System


Motivation Mission-critical systems with high-assurance

requirements – in defense, communications, healthcare, are increasingly adopting SOA Often non-intuitive tradeoffs are necessary between multiple QoS

(e.g., security, reliability, availability, performance)

Lack of engineering methodology for SOA system QoS

QoS design today is an ad-hoc, manual and costly one-off job for every SOA system

Increases risk and cost for successfully deploying these solutions No procedures or tools available to intelligently explore the solution

space to find the best tradeoffs


Overview of State of the Art SOA standard specifications for security and reliable messaging

policies are maturing Implementations exist for enterprise middleware platforms However, no support in higher-level QoS design methodology or tool

Several research projects addressing SOA QoS Primary goal is for “service matching”– selecting the service that meets

the required QoS ”– not system design Provide point solutions to very specific problems – not “end-to-end”

QoS Modeling for Distributed Real-time Embedded (DRE) systems E.g., Component QoS Modeling Language (CQML), Vanderbilt Univ. Fault-tolerance, Network QoS etc.


Project Objectives Provide system engineers with a methodology and

supporting toolset to enable them to design SOA-based systems that achieve end-to-end QoS

Use model-driven techniques to generate configurations adapted to the target deployment environment that meet the QoS requirements Reduce development burden and configuration

errors Facilitates deployment onto heterogeneous

environments

Enable adaptation to changes in system’s QoS requirements or deployment environment by re-computing new target configuration Incremental computation to minimize rework Minimize configuration delta to reduce impact


Solution: SOA Dynamic Designer for QoS

Model System End-to-End QoS

Requirements

Suggest Composition Patterns for

QoS

Build Service QoS

Profile

DeployPromising Candidates

SOA Dynamic

Designer for QoS

Design-Time Intelligently present design

options based on all constraints

Allows users to do “what-if” analysis on alternative designs

Automatically configure “tunables” on services and deployment environment

1 2

34


Order Manager

Credit Check

Model-based Design Specification (1/3) System model specifies composition of SOA services

Structure Behavioral Abstractions

Parallel Serial

Inventory Check

Customer Data

Inventory Data

External Credit Check

Serial

Parallel

E2E QoS: Support up to N orders/hour while respondingwith order confirmation in no more than X sec.

Order Confirmation Response Time

T = max (Ta, Tb)

Ta Tb

Credit CheckResponse

TimeTa = Tx + Ty

Tx

Ty

T


Model-based Design Specification (2/3) Specify QoS attributes (secure channel, reliable channel etc) Apply composition patterns

Encryption strength At least once, exactly once delivery, etc

Secure Channel

Reliable Channel

Order Manager

Credit Check Inventory Check

Customer Data

Inventory Data


Encryption Mediator

Persistent Message Queue


Collocated and Clustered

Model-based Design Specification (3/3) Specify deployment pattern alternatives

Single/multiple core(s), small/medium/large memory resources Clustering Vs. standalone Collocated Vs. distributed

Order Manager

Credit Check Inventory Check

Customer Data

Inventory Data


Inventory Check

Inventory Data

Inventory Check

Inventory Data

Inventory Check

Inventory Data

Less Resourceful Machine

Resourceful Machine


Build Service QoS Profile

CloudLauncher Model-based tool for empirical load testing Use Amazon Elastic Compute Cloud (EC2) as testbed Automatically instantiate

different deployment patterns different deployment host resource settings different QoS composition patterns (in progress)

Use Grinder for generating dynamic load (in progress)

Measure data points (response time, load, utilization) Interpolate to estimate service QoS profile (in progress)


CloudLauncher Supported Deployment Patterns

Amazon Cloud (US-East-1* Zone)

Amazon EC2 Instance

App Server: Order

Manager

DB: Customer

DataJDBC

Single Machine Instance

App Server: Order

Manager

DB: Customer

DataJDBC

Distributed on 2 Machines

DB: Customer

Data

App Server: Order Status



App Server: Order

Manager

JDBCLoad Balancer

Clustered App Server

App Server: Order Manager

DB: Customer Data

JDBC


DB: Customer Data

JDBC


DB: Customer Data

JDBC

Load Balancer

Replicated DB


Architecture of Grinder

SOA Dynamic

Designer for QoS

Amazon Cloud (US-East-1* Zone)

Amazon EC2 Instance

The GrinderController

Test Clients

Distributes testplan to agents,

And collects teststatistics DB:

Customer Data




App Server: Order

Manager

JDBC

Load Balancer

Clustered App ServerThe GrinderAgent G

enerate


Example Service Profile for Different Deployment Patterns

Response Time

Load

S Host

L Host

XL Host

Response Time

Load

1H-Cluster

4H-Cluster

Response Time (Load) = a + b * Load

2H-Cluster

Machine Type Clustered Configuration

L1 L2 L1 L2

T1

T2

T1

T2


Concluding Remarks

SOA Dynamic Designer for QoS capabilities Model end-to-end QoS requirements

Intelligently suggest QoS patterns

Generate configuration for target deployment environment

Build QoS profile of the system by empirical testing on cloud platforms


Thank you!


Extra slides


Generic Modeling Environment (GME) Developed at Institute for

Software Integrated Systems (ISIS), Vanderbilt University

Under development since 1995 (GME 9 in beta testing)

Enables layered, multi-view system modeling, model transformation, model analysis and validation, and model execution

API support in C++, Java, Python

Reference: Composing Domain-Specific Design Environments, Akos Ledeczi et al., IEEE Computer, Nov 2001

Parts browser with custom icons

Tree viewSystem Model


Cloud Vs. Grid Computing

19

Cloud Grid

Transactional apps (e.g. web sites) High performance, one application

Real-time resource allocation, Fork-a-server now!

Best-effort scheduling of queued allocation requests

Requests of up to 10-50 nodes Requests of up to 500-1000 nodes

Same administrative domain (so far) Sharing of resources belonging to different admin domains

Pay-as-you-go (frequent scale out and scale down)

infrequent

Virtualization is crucial Virtualization is just entering

Usability: Easy (PaaS, SaaS) Heavy

Documents

Copyright © 2009, Telcordia Technologies, Inc. All Rights Reserved. 1 / 16 Model-based SOA Performance Profiling using CloudLauncher Other project members