Elastic software infrastructure to support the industrial internet

transcript

Douglas C. Schmidtd.schmidt@vanderbilt.edu

Elastic Software Infrastructure to

Support the Industrial Internet

Institute for Software Integrated Systems

Vanderbilt University

Nashville, TN

RTI Webinar Series, October 23rd, 2013

• Context & terminology

• Prior R&D progress

• Current R&D trends & challenges

• A promising solution

• Concluding remarks

Outline of Presentation

• The Industrial Internet is a term coined by GE that refers to the integration of complex physical machinery with networked sensors & software

Overview of the Industrial Internet

en.wikipedia.org/wiki/Industrial_Internet

• It combines fields such as machine learning, big data, the Internet of things, & machine-to-machine communication to

• Connect machines embedded with sensors to other machines (& end users)

• Enable access & control of mechanical devices

• Extract data from these devices, make sense of it, & deliver the right information to the right people at the right time (& in real-time)

• Derive some form of value in terms of improved utility, & cost savings

At the heart of the Industrial Internet are cyber-physical systems & clouds

• A cyber-physical system (CPS) features a tight coordination between the system’s computational & physical elements

Overview of Cyber-Physical Systems

en.wikipedia.org/wiki/Cyber-physical_system

Overview of Cyber-Physical Systems• A cyber-physical system

(CPS) features a tight coordination between the system’s computational & physical elements

• CPSs increasingly use networked processing elements to control physical, chemical, or biological processes or devices

www.ge.com/stories/industrial-internet has other apt examples

• A cyber-physical system (CPS) features a tight coordination between the system’s computational & physical elements

• CPSs increasingly use networked processing elements to control physical, chemical, or bi ological processes or devices

• In CPSs the “right answer” delivered too late becomes the “wrong answer”• i.e., dependability has

a temporal dimension

(& increasingly a security dimension)

This talk focuses on distributed CPSs rather than standalone CPSs

Overview of Cyber-Physical Systems

• Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources • e.g., networks, servers,

storage, applications, & services

Resource pooling

Rapid elasticity

Broad network access

On-demand self-service

Measuredservice

Overview of Cloud Computing

• Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources

• These resources can be rapidly provisioned & released with minimal management effort or service provider interaction

Resource pooling

Rapid elasticity

Broad network access

On-demand self-service

Measuredservice

csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf

Overview of Cloud Computing• Cloud computing is a model for enabling

ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources

• Cloud offerings enable “economies of scale” via multi-tenancy & elasticity• e.g., run atop shared

(often virtualized) data access, storage, hardware, software, middleware, etc.

Overview of Cloud Computing• Cloud computing is a model for enabling

ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources

• Cloud offerings enable “economies of scale” via multi-tenancy & elasticity

• Cloud services don’t require users to know of the configuration & physical location of the computing & communication infrastructure delivering services• Similar to traditional utilities, such as power grids, water, sewer,

as well as datacom/telecom service providers

Some implementations of cloud computing may be at odds with CPS needs..

• Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources

• Cloud offerings enable “economies of scale” via multi-tenancy & elasticity

• Cloud services don’t require users to know of the configuration & physical location of the computing & communication infrastructure delivering services• Similar to traditional utilities, such as power grids, water, sewer,

as well as datacom/telecom service providers

The designs of legacy CPSs tend to be:• Stovepiped• Proprietary • Brittle & non-adaptive• Expensive to develop & evolve• Vulnerable

From this design paradigm…

AirFrame

Nav WTS

SPLnner IFF

Cyclic Exec

Problem: Small changes can break nearly anything & everything

Prior R&D Progress for Cyber-Physical Systems

…and this operational paradigm…

Resources

Utility “Curve”

“Broken” “Works”

“Hard” Requirements

Problem: Lack of any resource can break nearly everything

Real-time QoS requirements for legacy CPSs:• Ensure predictable end-to-end

QoS, e.g.,• Bound latency, jitter, &

footprint• Bound priority inversions

• Allocate & manage resources statically & avoid sharing

This is not at all what we think of as a computing cloud!

Real-time QoS requirements for legacy CPSs:• Ensure predictable end-to-end

QoS, e.g.,• Bound latency, jitter, &

footprint• Bound priority inversions

• Allocate & manage resources statically & avoid sharing

Information Backbone

…to this design paradigm…

Event Channel

ReplicationService

PLanner IFF FLIR

AirFrame

AP Nav WTS The designs of today’s leading-edge CPSs tend to be more:• Layered & componentized• Standards- & COTS-based • Robust to failures & adaptive to

operating conditions• Cost effective to evolve & retarget

Result: changing requirements & environments can be handled more flexibly

Resources

Desired Utility Curve

“Working Range”

“Softer” Requirements

Result: better support for operations with scarce/contended resources

• Ensure acceptable end-to-end QoS, e.g.,• Minimize latency, jitter, & footprint• Minimize priority inversions

• Resources are allocated/managed dynamically & can be shared

Some CPS operating platforms have much in common with computing clouds

See www.dre.vanderbilt.edu/~schmidt/JSS-DRM.pdf for more info

Key solution space challenges• Enormous accidental & inherent

complexities• Continuous evolution & change• Highly heterogeneous platform,

language, & tool environments

Key problem space challenges

• Dynamic behavior• Transient overloads• Time-critical tasks• Context-specific requirements• Resource conflicts• Interdependence of (sub)systems• Integration with legacy

(sub)systems

New Challenge: Ultra-Large-Scale Cyber-Physical Systems

(sub)systems

Mapping problem space requirements to solution space artifacts is very hard!

(sub)systems

See www.dre.vanderbilt.edu/~schmidt/PDF/FOME-HCDS-paper.pdf for more

Ultra-Large-Scale CPSs are well beyond scope of today’s computing clouds

(sub)systems

“Gentlemen, we have run

out of money. It is time to start

thinking.”

en.wikiquote.org/wiki/Talk:Winston_Churchill

Convenient Trend: Elastic Hardware Platforms• “Elastic hardware” based on

multi-core & distributed-core architectures now available at reasonable prices

en.wikipedia.org/wiki/Elasticity_(cloud_computing) has more info

• Elastic hardware has potential to substantially accelerate performance by parallelizing application work loads & auto-scaling data processing at runtime – Goal is to add/utilize more

hardware without changing application business logic or configurations

• Elastic hardware has potential to substantially accelerate performance by parallelizing application work loads & auto-scaling data processing at runtime

• Current focus of elastic hardware is largely on web hosting applications in public cloud environments

• Elastic hardware has potential to substantially accelerate performance by parallelizing application work loads & auto-scaling data processing at runtime

• Current focus of elastic hardware is largely on web hosting applications in public cloud environments

Elastic hardware is necessary—but not sufficient—for elastic CPS applications

Impediments to Applying Elastic Hardware for CPSs

• Inadequate programming models– Complicated & obtrusive APIs– Can’t use hardware predictably

& scalably

ISR Processing SCADA Systems Air Traffic Mgmt Aerospace

• Inadequate programming models

• Inadequate knowledge of real-time, concurrency, & networking – e.g., high probability of race

conditions, deadlocks, priority inversion, & missed deadlines

• Inadequate knowledge of real-time, concurrency, & networking

• Inadequate mechanisms to transition seamlessly from multi- to distributed-core environments

• Inadequate knowledge of real-time, concurrency, & networking

• Inadequate quality-of-service (QoS) support at scale – e.g., lack of system-wide control

over key QoS impacting resource usage & end-to-end data deliver semantics

• Inadequate programming models– Complicated & obtrusive APIs– Can’t use hardware predictably

& scalably• Inadequate knowledge of real-

time, concurrency, & networking – e.g., high probability of race

conditions, deadlocks, priority inversion, & missed deadlines

• Inadequate quality-of-service (QoS) support at scale – e.g., lack of system-wide control

over key QoS impacting resource usage & end-to-end data deliver semanticsSome impediments affect many types of systems, some mostly

affect CPSs

Key Research Challenges for Elastic CPSs

Meeting these challenges requires rethinking some cloud computing tenets

1. Precise auto-scaling of resources with a system-wide end-to-end focus

2. Flexible optimization algorithms to balance real-time constraints with cost & other goals

3. Improved fault-tolerance fail-over that supports real-time requirements

4. Data provisioning & load balancing algorithms that consider physical properties of computations & storage

1. Precise auto-scaling of resources with a system-wide end-to-end focus– State-of-the-art in auto-

scaling algo rithms manage services in isolation• CPSs require auto-

scaling algo rithms to operate on end-to-end task chains

CPU utilization

1. Precise auto-scaling of resources with a system-wide end-to-end focus– State-of-the-art in auto-

scaling algo rithms manage services in isolation

– Physical stability & safety properties may require exceedingly complex analyses• e.g., reachability of

hybrid cyber-physical states

CPU utilization

2. Flexible optimization algorithms to balance real-time constraints with cost & other goals– CPS deployments must be

schedulable on all resources acquired from cloud providers to ensure real-time response times, while optimizing desired objective functions• e.g., minimizing costs

Response Time Scalability

Multi-dimensional Resource Management

2. Flexible optimization algorithms to balance real-time constraints with cost & other goals– CPS deployments must be

schedulable on all resources acquired from cloud providers to ensure real-time response times, while optimizing desired objective functions

– Principled means are needed to co-schedule and/or per form admission control & eviction of mixed-criticality task sets deployed on cloud resources

Response Time Scalability

Multi-dimensional Resource Management

3. Improved fault-tolerance fail-over that supports real-time requirements– Some cloud platforms

tolerate faults for provisioned re sources• This is insufficient for

CPSs where real-time fault-tolerance of end-to-end task chains must be met simultaneously

Key Research Challenges for Elastic CPSsISR Processing SCADA Systems Air Traffic Mgmt Aerospace

3. Improved fault-tolerance fail-over that supports real-time requirements– Some cloud platforms

tolerate faults for provisioned re sources

– Reasoning about the consequences of faults is an important open re search area due to the complex & stochastic nature of many CPSs

Key Research Challenges for Elastic CPSsISR Processing SCADA Systems Air Traffic Mgmt Aerospace

4. Data provisioning & load balancing algorithms that consider physical properties of computations & storage– CPSs generate load on

a computing clouddue to physical stimuli

social network linkages

geographic associations

power consumption

cache affinity

– To build more scalable & high-performance CPSs, algorithms & techniques are needed to • Exploit physical

characteristics of data & computation

power consumption

cache affinity

• Improve the distribution of work in a computing cloud

power consumption

cache affinity

• Improve the distribution of work in a computing cloud

power consumption

cache affinity

We need a holistic solution that provides an elastic CPS software infrastructure

Requirements for Elastic CPS Software Infrastructure

• Flexibility – Loosely coupled components that can be analyzed, replaced, reused, distributed, & parallelized dependably

Middleware

Dynamic Discovery

Load Balancing

Data DistributionLow Latency

Dependability

• Adaptability – Provide APIs that adapt to existing code, rather than always having to adapt code to an API Middle

Dynamic Discovery

Load Balancing

Dependability

• Adaptability – Provide APIs that adapt to existing code, rather than always having to adapt code to an API

• Uniformity – Seamless (ideally standards-based) support for multi-core & distributed-core

Middleware

Dynamic Discovery

Load Balancing

Dependability

• Scalability – Static & dynamic load balancing ensures best & dependable utilization of available elastic hardware resources

Middleware

Dynamic Discovery

Load Balancing

Dependability

• Scalability – Static & dynamic load balancing ensures best & dependable utilization of available elastic hardware resources

Middleware

Dynamic Discovery

Load Balancing

Dependability

Middleware is a key element of elastic CPS software infrastructure

Key Layers of CPS Software Infrastructure

Provide mechanisms to manage end-system resources, e.g., CPU scheduling, inter-process communication, memory management, & file systems

Domain-SpecificServices

CommonMiddleware

ServicesDistributionMiddleware

Host InfrastructureMiddleware

Operating Systems &

Protocols

CommonMiddleware

Operating Systems &

Protocols

Encapsulates & enhances native OS mechanisms to create reusable network programming components

Key Layers of CPS Software InfrastructureISR Processing SCADA Systems Air Traffic Mgmt Aerospace

CommonMiddleware

Operating Systems &

Protocols

Defines higher-level programming models whose reusable APIs & components automate & extend native OS capabilities across distribution boundaries

CommonMiddleware

Operating Systems &

Protocols

Augment distribution middleware by defining higher-level domain-independent services that focus on programming “business logic”

CommonMiddleware

Operating Systems &

Protocols

Tailored to requirements of particular domains, such as SCADA, avionics, aerospace, vehtronics, C4ISR, air traffic management, integrated healthcare, etc.

Promising Elastic CPS Middleware: DDS• The OMG Data

Distribution Service (DDS) promotes a pattern language that yields loosely coupled, polyglot, evolvable, scalable, efficient & dependable CPSs

en.wikipedia.org/wiki/Data_Distribution_Service has a good DDS overview

Promising Elastic CPS Middleware: DDS• The OMG Data Distribution

Service (DDS) promotes a pattern language that yields loosely coupled, polyglot, evolvable, scalable, efficient & dependable CPSs– DDS supports relational

& OO information modeling• Data-Centric Publish-

Subscribe (DCPS) & Data Local Reconstruction Layer (DLRL)

Global Data Space

Distribution Service (DDS) promotes a pattern language that yields loosely coupled, polyglot, evolvable, scalable, efficient & dependable CPSs– DDS supports flat,

relational, & OO information modeling

– DDS global data space allows apps to read/write data anonymously & asynchronously, decoupled in space & time

Data Reader

Data Writer

Data Reader

Subscriber Publisher Subscriber

Domain Participant

Topic Topic

Global Data Space

– DDS pub/sub model allows apps to produce/consume information into/from the global data space

Topic Topic

Data Reader

Data Writer

Data Reader

Domain Participant

– DDS pub/sub model allows apps to produce/consume information into/from the global data space

DDS mainly provides distribution middleware & common middleware services

Global Data Space

Topic Topic

Data Reader

Data Writer

Data Reader

Domain Participant

• DDS controls resource usage, end-to-end data delivery, & data availability via a rich set of QoS policies, e.g.:– Batching– Priority– Deadline– Data Durability– Redundancy– Data History

Promising Elastic CPS Middleware: DDSISR Processing SCADA Systems Air Traffic Mgmt Aerospace

Subscriber

Data Reader

Data Writer

S6 S5 S4 S3 S2 S1S7 S7X

HISTORY

RELIABILITYCOHERENCY

RESOURCE LIMITS

LATENCY

TopicR

Publisher

Promising Elastic CPS Middleware: DDS

www.dre.vanderbilt.edu/~schmidt/PDF/CrossTalk-2008-final.pdf

Subscriber

Promising Elastic CPS Middleware: DDS• DDS controls resource

usage, end-to-end data delivery, & data availability via a rich set of QoS policies, e.g.:– Batching– Priority– Deadline– Data Durability– Redundancy– Data History

DDS’s request/offered (RxO) model matches QoS policies between pub & sub

Subscriber

Data Reader

Domain Participant

RequestedQoS

OfferedQoS

Subscriber

Data Reader

Domain Participant

OfferedQoS

Promising Elastic CPS Middleware: DDS• DDS controls resource

usage, end-to-end data delivery, & data availability via a rich set of QoS policies:– Batching– Priority– Deadline– Data Durability– Redundancy– Data History

• Bridges are available across technologies to expose relevant data to heterogeneous network protocols, without imposing changes into existing legacy systems

Promising Elastic CPS Middleware: DDS• DDS is an OMG standard

that itself is based on many associated open standards

• Key DDS implementations are now available in open-source form• Many opportunities

for researchers to influence DDS standard & implementations

See www.dre.vanderbilt.edu/~schmidt/PDF/DDS-WAN.pdf for recent paper

• DDS is used in many CPS research projects & production systems

• portals.omg.org/dds provides more info on DDS activities & projects

Concluding Remarks• Despite advances in elastic hardware, deploying

CPSs in cloud environments is hard without adequate support from elastic software infrastructure – It’s unlikely that public clouds will work for

mission-critical Industrial Internet applications

Key characteristics of computing clouds for CPS are multi-tenancy & elasticity

CPSs in cloud environments is hard without adequate support from elastic software infrastructure

• Standards-based DDS middleware provides key open-source building-blocks to create a dependable elastic CPS software infrastructure

There are many hard research challenges remaining

CPSs in cloud environments is hard without adequate support from elastic software infrastructure

• There are many hard research challenges remaining

www.industrialinternet.com/blog/three-qs-professor-douglas-schmidt/

Concluding Remarks

www.coursera.org/course/posa

“Big breakthroughs often happen when

what is suddenly possible meets what is

desperately necessary”

– Thomas Friedman

• Despite advances in elastic hardware, deploying CPSs in cloud environments is hard without adequate support from elastic software infrastructure

• There are many hard research challenges remaining

• See www.isis.vanderbilt.edu/workshops/cc4cps for info on an NSF workshop on Computing Clouds for Cyber-Physical Systems (CC4CPS) • Attended by ~50 researchers

funded by the NSF•Topics of workshop included

• Role of computing clouds in data collection, integration, analysis, & mining for CPS

• Roles of computing clouds in CPS control

• Stability, safety, security, privacy, & reliability considerations in integrating cloud computing with CPS

• Programming models & paradigmsfor computing clouds that support CPS

The NSF CC4CPS workshop report will be available later this year

Additional Information

Ultra-large-scale (ULS) systems are socio-technical ecosystems comprised of software-reliant systems, people, policies, cultures, & economics that have unprecedented scale:

• # of software & hardware elements

• # of connections & interdependencies

• # of computational elements

• # of purposes & perception of purposes

• # of routine processes & “emergent behaviors”

• # of (overlapping) policy domains & enforceable mechanisms

• # of people involved in some way• Amount of data stored, accessed, &

manipulated

• … etc …

www.sei.cmu.edu/uls

See blog.sei.cmu.edu for more discussions of software R&D activities

Sponsored by Office of the Secretary of Defense (OSD) with assistance from the National Science Foundation (NSF), & Office of Naval Research (ONR), www.nap.edu/openbook.php?record_id=12979&page=R1

The report focuses on ensuring the DoD has the technical capacity & workforce to design, produce, assure, & evolve innovative software-reliant systems in a predictable manner, while effectively managing risk, cost, schedule, & complexity

NRC Report Critical Code: Software Producibility for Defense (2010)

See blog.sei.cmu.edu for more discussions of software R&D activities

• The Institute for Software Integrated Systems (ISIS) was established at Vanderbilt in 1998

• Research at ISIS focuses on systems with deeply integrated software that are networked, embedded,& cyber-physical

• Key research areas at ISIS: • Model-Integrated

Computing

• Middleware for distributed real-time & embedded (DRE) systems

• Model-based engineering of cyber-physical systems• Wireless sensor networks• Systems security & privacy

www.dre.vanderbilt.edu/~schmidt/ISIS-research.pdf has more info on ISIS

Elastic software infrastructure to support the industrial internet

Technology