Designing Resource-Aware Applications for the Cloud with ABS

Designing Resource-Aware Applicationsfor the Cloud with ABS

Einar Broch Johnsen

University of Oslo, [email protected]

1st Intl. Workshop on Formal Methods for and on the Cloud (iFMCloud)Reykjavik, Iceland, 04 June 2016

http://www.envisage-project.eu

Einar Broch Johnsen (UiO) Designing Resource-Aware Applications iFMCloud, 04.06.2016 0 / 22

[email protected]


http://cordis.europa.eu/fp7/home_en.html


We want to make e↵ective use of cloud computingto meet service requirements

Cloud API

ApplicationService

IVirtualization makes elasticamounts of resources available toapplication-level services

IMetered resources: Resources onthe Cloud are pay-on-demand

I Services need to share and scale

resources

IDigitalization: new services needto share resources with old services

Discovering bad resource management after deploymenton the Cloud can be a very costly (wasting both time and money!)



Services deployed on the cloud:Predicting behavior from models

ApplicationServer

DC4DC3

CalcServerDC2

CalcServer

DC1

CalcServer

AppWorkflow

CloudProvider

Invoke task

# of VMTotal Cost

Application Server =

Application Workflow +

Application Resource Management(load balancing, scalability)

CalcServer =

Independent tasks(can be parallelized)

AppRM

IResource-aware design:Build software that candynamically modifyits own deploymentto improve perfor-mance and/orreduce cost

IModel-based deployment decisions at design time using service models

IFormal semantics: Architects and developers can simulate and analyze atdesign time how an application runs on the cloud

Hahnle, Johnsen. Designing Resource-Aware Cloud Applications. IEEE Computer 48(6), 2015



What kind of questions can we answer using models?

Berndnaut Smilde: Nimbus II, 2012

Model-based analysis ofperformance vs. cost

1 How will the response time and cost of running my system changeif I double the number of servers?

2 Can I meet my performance requirements with my currentdeployment strategy? What about fluctuations in client tra�c?

3 Can I control the performance of my system better by using acustom resource manager?

Use the model topredict behavior



Conceptual Parts of a Deployed Cloud Service

Provisioning Layer

Legal Contract Layer

Formal Service Contract

Executable Model of Client Layer

Cloud API

Simulation

“early modeling”

Formal Methods

“early analysis”

Provisioning

“runtime monitoring”

Combine techniques based on abstract executable models

I Formal modeling using Abstract Behavioral Specifications (ABS)

I Formal methods: Verification, Performance Analysis, Cost Analysis,Advanced Type Systems, Code Generation, Test-Case Generation

I Monitoring: Framework to generate monitors for SLA-compliance



Example: Phone Services - Abstract Behavioral Model

Telephone Service

interface TelephoneService {Unit call(Int calltime);

}class TelephoneServer implements TelephoneService {Int callcount = 0;Unit call(Int calltime){while (calltime > 0) { [Cost: 1] calltime = calltime � 1;

await duration(1, 1); }callcount = callcount + 1;

}}

SMS Service

interface SMSService {Unit sendSMS();

}class SMSServer implements SMSService {Int smscount = 0;Unit sendSMS() {[Cost: 1] smscount = smscount + 1;}

}



Example: The New Year’s Eve Client Behavior

50 70

Alternate sms and call

Huge number ofsms per time interval

time

Alternate sms and call

Midnight Window

class NYEclient(Int frequency,TelephoneService ts,SMSService smss){Time created=now(); Bool call=false;Unit normalBehavior(){ ... }Unit midnightWindow(){ ... } // Switch at appropriate time...

}{// Main block:

DC smscomp = new DeploymentComponent(”smscomp”, Speed(50));DC telcomp = new DeploymentComponent(”telcomp”, Speed(50));[DC: smscomp] SMSService sms = new SMSServer();[DC: telcomp] TelephoneService tel = new TelephoneServer();Client c = new NYEbehavior(1,tel,sms); ... // Clients

}

How to deploythe services?



Example: Simulation Results

Einar Broch Johnsen, Olaf Owe, Rudolf Schlatte, Silvia Lizeth Tapia Tarifa:Dynamic Resource Reallocation between Deployment Components. ICFEM 2010: 646-661



Load Balancing in Deployment Scenarios

smscomp

telcomp

Client

Client

sms()

call(n)

sms()

call(n)

tel

sms

telb

smsb

request()

Resource awareness: resource reallocation, object mobility, job distribution

I dc.load(e): average load on dc during the last e time intervals

I dc.total(): currently allocated resources on dc

I dc.transfer(dc2, r): transfer r resources to dc2

Load Balancing Strategy for the Phone Services

Example: Reallocate 1/2⇥total resources upon request from partner



Example: Simulation Results

Johnsen, Owe, Schlatte, Tapia Tarifa:Dynamic Resource Reallocation between Deployment Components. Proc. ICFEM 2010



ABS: Abstract Behavioral Specification

ABS: Between design-oriented and implementation-oriented specification

I State-of-the-art modeling language: actors + OO

I Models follow the execution flow of OO programs,but abstract from implementation details using ADTs

I ABS allows time modeling and deployment modeling

I Java-like syntax: intuitive to the programmer

ABS is a formal, tool-supported modelling language

I Operational semantics allows advanced analysis techniques

I Simulation tool for rapid prototyping

I Automated worst-case resource and deadlock analysis

I Automated optimization of static deployment

I Semi-automated scalable verification of functional correctness (KeY)

I Code generation into Java and Haskell (preserves cost bounds)



Deployment Components

Deployment components are abstract execution locations

I Each deploymentcomponent has a givenresource capacity

I Objects execute in thecontext of a deploymentcomponent

Server ...

...objectEnv

[cost] Task1[cost] Task2object 1

[cost] Task1[cost] Task2

object n[cost] Task1[cost] Task2

I The resources are shared between the component’s objects

I Object execution uses resources in a deployment component(via Cost annotations)

I How resources are assigned and consumed,depends on the kind of resource

Johnsen, Schlatte, Tapia Tarifa. Integrating deployment architectures and resource consumptionin timed object-oriented models. J. Log. Algebr. Meth. Program. 84(1), 2015



A Resource-Aware Application in ABS

interface CalcServer {Unit process(Int cost);DC getDC();

}class Server implements CalcServer {Unit process(Int cost) { [Cost: cost] skip; }DC getDC() { return thisDC(); }

}interface ApplicationServer {Bool request(Int cost);

}

ApplicationServer

DC4DC3

CalcServerDC2

CalcServer

DC1

CalcServer

AppWorkflow

CloudProvider

Invoke task

# of VMTotal Cost




CalcServer =


AppRM

class ConstantBalancer(CloudProvider provider, Int serverSize) implements ApplicationServer {Server server; DC dc; Bool initialized = False;

Unit run() {Fut<DC> f = provider!createMachine(serverSize); await f?; dc = f.get;[DC: dc] server = new Server(); initialized = True;

}

Bool request (Int cost) {await initialized;Fut<Unit> r = server!process(cost); await r?; return (durationValue(deadline()) > 0);

}}



A Resource-Aware Application in ABS

interface CalcServer {Unit process(Int cost);DC getDC();

}class Server implements CalcServer {Unit process(Int cost) { [Cost: cost] skip; }DC getDC() { return thisDC(); }

}interface ApplicationServer {Bool request(Int cost);

}

ApplicationServer

DC4DC3

CalcServerDC2

CalcServer

DC1

CalcServer

AppWorkflow

CloudProvider

Invoke task

# of VMTotal Cost




CalcServer =


AppRM

class DynamicBalancer(CloudProvider provider) implements ApplicationServer {Map<Int, Set<Server>> sleepingMachines = EmptyMap;Int machineStartTime = ... // a constant representing the time it takes to start a machine;

Bool request (Int cost) {Int requiredResources = (cost / durationValue(deadline())) + 1 + machineStartTime;Server server = this.getMachine(requiredResources);Fut<Unit> r = server!process(cost); await r?;this.dropMachine(server); return durationValue(deadline()) > 0;

}

Server getMachine(Int size) { ... } // take machine of size if it exists, otherwise create one

Unit dropMachine(Server server) { ...}}



Rapid Prototyping: Simulation Results

I Define client behaviorto model a load spike

timeIncrease the # of requests

I Simulate the di↵erent scenarios with ABS simulator

User scenario

Load spikeStrategy QoS Total CostConstant balancer 53% 200As-needed balancer 100% 128

I QoS: measure the successful requests (i.e., requests completed withinthe deadline) divided by the total number of requests

I Total Cost: measures the accumulated sum of CPU resources madeavailable by the cloud provider.



Case Study: Montage (1)

Montage is a toolkit for assembling astronomicalimages into customized mosaics

mProject

mProjExec

mImgtbl mOverlaps

mDiffExec

mDiff

mFitExec

mFitplane

mBgModel

mBackground

mBgExec

mAdd

Re-project Image

Background modeling

Backgroundmatching

Final mosaic

Partly ordered workflow and highly parallelizable tasks.




DC4DC3

CalcServerDC2

CalcServer

DC1

CalcServer

CloudProvider

Invoke task

ApplicationServer =

AppWorkflow + AppRM

CalcServer =

Independent tasks

# of VMTotal Cost

ApplicationServer

AppWorkflow

AppRM

Model the Montage toolkit using the Cloud Provider API, run simulationsvarying the di↵erent deployment scenario and compare the results.




Cost vs. Time Tradeo↵: can we reproduce the resultsof other informal cloud simulation tools (GridSim)?

!"

!#"

!##"

!###"

!####"

!" $" %" &" !'" ($" '%" !$&"

!"

#$"

$!"

%$"

&!!"

&" #" '" (" &)" *#" )'" &#("Lo

garit

hmic

sca

le

60 cents for 1 processor

Loga

rithm

ic s

cale

approx 4 $ for 128 processors

approx 5.5 hrs for 1 processor

approx 18 min for 128 processors

432270

1152

CPU

Cos

t

89

28

Tim

e

The cost of doing science on the cloud: The Montage example.E. Deelman, G. Singh, M. Livny, G. B. Berriman, and J. Good. (SC’08), pages 1–12. IEEE/ACM, 2008.

Johnsen, Schlatte, Tapia Tarifa. Modeling Resource-Aware Virtualized Applications for the

Cloud in Real-Time ABS. Proc. ICFEM 2012



Case Study: Fredhopper Replication Server (1)

The Fredhopper Access Server (FAS) is a distributed, concurrent OOsystem providing search and merchandising services to e-Commercecompanies. The Replication Server is one part of FAS.

Acceptor

Cloud Provider

ClientJob

ClientJob

ClientJob

SyncClient

job(schedule)

SyncClient

SyncClient

... ...

LIVE STAGING

SyncServer create()

CLOUD

DC4

ConnectionThread

getConnection(schedule)



job(schedule)

job(schedule)

DC3

ConnectionThread

replication

DC2

ConnectionThread

replication

DC1

ConnectionThread

replication

I Very detailed model: consists of 5000 lines of ABS

Albert, de Boer, Hahnle, Johnsen, Schlatte, Tapia Tarifa, Wong.

Formal modeling and analysis of resource management for cloud architectures: an industrial

case study using Real-Time ABS. Service Oriented Computing and Applications 8(4), 2014



Case Study: Fredhopper Replication Server (2)

How does the accumulated cost in our modelcompare to the actual Java implementation?

0

17.5

35

52.5

70

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200

7500

15000

22500

30000

Runn

ing

time

[s]

Environments

Sim

ulat

ion

cost

Model simulation cost Implementation running time

Measured execution time of the implementation (left scale)

Accumulated cost of the simulation (right scale)

The deviation roughly seems to correspond to the start-up time of JVM



Case Study: Hadoop YARN Clusters (1)

Open-source software framework thatimplements a cluster managementtechnology for distributed processing.

Popular cloud frameworkfor big data processing:

I Resource allocation

I Code distribution

I Distributed data processing

Lin, Yu, Johnsen, Lee. ABS-YARN: A Formal Frameworkfor Modeling Hadoop YARN Clusters. Proc. FASE 2016



Case Study: Hadoop YARN Clusters (2)

How does the ABS YARN compare to the actual YARN implementation?

(a) The normalized starting time (b) The normalized finish time

(c) Cumulative completed jobs (d) Total number of completed jobs



The ABS Collaboratory

I ABS as a web service, with documentation and examplesunder development at http://www.abs-models.org

I Tools are open source: https://github.com/abstoolsI ABS API available for orchestration of Java codeI Eclipse plug-in for ABS

Get involved!


http://www.abs-models.org

https://github.com/abstools


Summary

Virtualization requires novel modeling abstractionsExecutable models, deployment components, reflection

ABS: Abstract Behavioral Specification

I Model deployed services with dynamic resource management

I High-level abstractions of low-level platform-specific concerns

I Analysis methods: performance, cost analysis, deadlock analysis, . . .

I More info and open source tools:www.abs-models.org

Make your deployment decisions at design time!ABS permits concise modeling and accurate prediction

Engineering Virtualized Services[www.envisage-project.eu]


www.abs-models.org

www.envisage-project.eu


Designing Resource-Aware Applications for the Cloud with ABS

Data & Analytics