Reconfigurable Real-Time Middleware for Distributed Cyber-Physical Systems with Aperiodic Events

Reconfigurable Real-Time Middleware for

Distributed Cyber-Physical Systemswith Aperiodic Events

Yuanfang Zhang, Christopher Gill, Chenyang LuDepartment of Computer Science & Engineering

2

Motivation

• Cyber-Physical Systems (CPS) require integrated design of computing & physical systems.

• Challenge: Diversity of CPS applications– Avionics, automobile, manufacturing, medical, power grid…

– Different CPS applications need different middleware configurations.

– Existing real-time middleware provides fixed sets of services.• Real-Time CORBA, Real-Time Java, CORBA Component Model.

• Goal: reconfigurable middleware for diverse CPS applications.Tailor middleware services to specific needs of a CPS application.

Facilitate integrated design of CPS.

3

Outline

• Middleware architecture• Alternative service strategies

– encapsulated in configurable middleware components

• Map CPS application characteristics to service strategies– supported by configuration tools

• Implementation and empirical evaluation

4

Application Model• End-to-End Task Ti = chain of subtasks (Ti,1, Ti,2, …Ti,k)

– Aperiodic or periodic– Subject to end-to-end deadline– Example: Subtask triggered by event from predecessor

• Job: an instance of a task

Application Processor 1

Application Processor 2

Application Processor 3

T1,2 T1,3T1,1

ECEC EC

Gateway

Gateway

Real-time Event Service on CORBA

5

Middleware Architecture

Task manager• Admission Control (AC)

• Load Balancing (LB)

Application processors• Idle Resetting (IR)

• Task Effector (TE)

OriginalComponent

DuplicateComponent

Original TaskAllocation

TaskReallocation

Task Manager

EC/ORB

AC

LB

ApplicationProcessor 1

EC/ORB

T1,1

TE IR

ApplicationProcessor 2

EC/ORB

T1,2

TE IR

ApplicationProcessor 3

EC/ORB

T1,1

TE IR

ApplicationProcessor 4

EC/ORB

TE IR

ApplicationProcessor 5

EC/ORB

T1,3

TE IR

T1,2

6

Admission Control Strategies

• Admission test based on aperiodic utilization bound [Abdelzaher04]

– Guarantee end-to-end deadlines of admitted tasks/jobs.

• AC per Task– Perform the admission test for an entire task when it arrives. Reserve capacity for all jobs of an admitted task no job skipping. More pessimistic admission test.

– Example: Digital control.

• AC per Job– Perform the admission test for each job of a task. No reservation for a task may skip some jobs of a task. Less pessimistic admission test.

– Example: Non-critical image acquisition.

7

Load Balancing Strategies

• Redirect events to replicas located on least loaded processors• Light weight: No state synchronization among replicas.• LB per Task

– The path of a task is determined upon arrival same path for all jobs. Achieve state persistency between jobs. Less performance benefit.

– Example: Integral control, video.

• LB per Job– Different jobs may be redirected to different paths. No state persistency between jobs. More performance benefit

– Example: Proportional control, image acquisition.

8

CPS Applications Services

• Can tolerate job skipping? Per Task AC (example: digital control) Per Job AC (example: image acquisition)

• Component Replication? No load balancing Load balancing

• Require state persistency between jobs? Per Job LB (example: Proportional control) Per Task LB (example: Integral control)

9

Configuration Space

• 15 valid configurations difficult to configure manually!

• Some combinations are invalid: Per Task AC/Per Job IR

Lo

ad B

alan

cin

g

Idle Resetting

No

ne

Per

Tas

k

Per

Jo

b

None Per Task Per Job

Per Job

Per Task

Admission Control

10

Configuration Tools

• Input: Application characteristics.– Does your application allow job skipping? [yes (Y), no (N)]– Does your application have replicated components? [yes (Y), no (N)]– Does your application require state persistence? [yes (Y), no (N)]

• Configuration Engine– Generate XML-based deployment plan– Avoid invalid combinations of strategies

• Deployment Engine (DAnCE) [Deng07] executes deployment plan.

11

1. N2. Y3. Y

4. PT

Workload

ConfigurationEngine

XML-baseddeployment

plan

Parse theplan

ComponentRepository

Deploy components on each node

Select

Createcomponent

server

CreateContainer

Deployment::NodeImplementationInfo

Deployment::DeploymentPlanDAnCE

PlanLauncher

DAnCEExecutionManager

Front End

DAnCENode

ManagerNode

ApplicationManager

set_configurationDAnCE

NodeApplication

Container

Configuration<instance id="Central-AC"> ....... <configProperty> <name>LB_Strategy</name> <value> <type> <kind>tk_string</kind> </type> <value> <string>PT</string> </value> </value> </configProperty>

12

Component Middleware

• Based on CIAO 0.6 [Wang04], open-source implementation of Light Weight CORBA Component Model (CCM) specification.

• Implemented real-time services as configurable components.– Supports real-time, aperiodic and periodic, end-to-end tasks.

13

Experimental Platforms

harry.csePentium4 2.53GHz1G RAM512KB cacheKURT-Linux 2.4.22

hermoine.csePentium4 2.80GHz1G RAM512KB cacheKURT-Linux 2.4.22

norbert.csePentium4 2.53GHz1G RAM512KB cacheKURT-Linux 2.4.22

ron.csePentium4 2.80GHz1G RAM512KB cacheKURT-Linux 2.4.22

neville.docPentium4 3.40GHz2G RAM2048KB cacheKURT-Linux 2.4.22

angelina.docPentium4 3.40GHz2G RAM2048KB cacheKURT-Linux 2.4.22

100Mbps Ethernet switch

14

Imbalanced Workloads

0

0.2

0.4

0.6

0.8

1

T_N_N

T_N_T

T_N_J

T_T_N

T_T_T

T_T_J

J_N_N

J_N_T

J_N_J

J_T_

N

J_T_

T

J_T_

J

J_J_

N

J_J_

T

J_J_

J

Av

era

ge

ac

ce

pte

d u

tiliz

ati

on

ra

tio

AC_IR_LB

N: None

T: Per Task

J: Per Job

Easy to generate different configurations.Middleware configurations have significant impact on real-time performance

15

Conclusions

• Configurable real-time middleware– Configuration tool maps application characteristics to

middleware configurations– Components middleware implements configurable services

Facilitate integrated design of diverse CPS applications