Service Morphing: Integrated System and ApplicationLevel Service Adaptation in Autonomic Systems

1
Service Morphing Integrated System- and
Application-Level Service Adaptation in Autonomic
Systems

• Christian Poellabauer, Karsten Schwan, Sandip
  Agarwala, Ada Gavrilovska, Greg Eisenhauer,
  Santosh Pande, Calton Pu, Matthew Wolf
• schwan_at_cc.gatech.edu
• Center for Experimental Research in Computing
  Systems
• Georgia Institute of Technology

2
Overview

• Key problems
• High-end, dynamic applications on distributed
  systems
• focus on interactive applications
• Flexibility in how, where, and when an
  applications required processing and
  communication actions are performed, and
• Continuous quality management meet end-user
  needs despite run-time variations in service
  locations, platform capabilities, and user
  requirements.

3
Application Drivers Services
High End Users and Displays
ORNL
Cluster Computer Terastream Server

Atlanta BioRing
downsample
transform
Instrumented Testbeds/Facilities
End Users and Displays

• Services
• dynamically deployed,
• cooperating components
• dynamic, hetero. platforms
• end-to-end QoS guarantees
• application-dependent

Specialize
differentiate
GT
Color Reduction
Window Filter
Reduce Rate 10x
Window Filter
Reduce Rate 2x
Data reduction/fusion Information
creation/manipulation
4
Airplane Data Traffic
Delta Air Example Operational Information
Systems
Operational Flight Displays
Airport LAN
High Performance Computing Real-time Decision
Tools
FAA Flight Data
Real-Time Information Transport
capture, display, transport, filter, transform
Simulation
Optimization
Gate Readers
Equipment Inspection

Cluster Server
Wide-area Transport
Passenger paging and response
Airport LAN
Crew and Equipment Status
Visualization
Real-time Situation Assessment
Baggage Displays
Scalable Robust Services
Storage
Baggage Status
Recovery and Replay
Security Systems
5
Applications - Needs

• Grids/Enterprise/Embedded Systems
• Heterogeneous and dynamic.
• Robust service delivery users assume services to
  be available despite varying resource levels and
  end user needs.
• This talk focus on performance.
• Also dynamic trust models/protection and high
  availability.

6
Approach Agile Services and Q-Fabric

• Agile Service Platform thin InfoFabric service
  layer permits applications to control information
  exchange
• ECho publish/subscribe middleware allows clients
  to subscribe to information channels of interest
  to them and to apply services to such information
  represented by dynamically generated handlers.
• Agility Using dynamic binary code generation,
  service code is deployed/specialized to match
  current user needs to available platform
  resources.
• Q-Fabric Configurable Resource Management
  kernel-level resource management deals with
  run-time changes in resource availability.
• Evaluation with High-end Applications display of
  real-time captured sensor data, remote viz,
  high-end actions on devices with limited
  resources (i.e., PDAs).

7
Agile, Managed Services Idea

• Services self-modify, which can entail combined
  use of middleware, compiler, and system
  technologies
• Self-modification may involve coordination across
  multiple levels of abstraction (e.g., middleware,
  network)
• Self-modification is assisted by underlying
  platforms (incl. OS extensions).

8
Results and Status

• InfoFabric middleware SPARC, i86, MIPS, Itanium,
  ARM, XScale.
• Agility dynamic deployment dynamic binary code
  generation, run-time methods for dynamic linking,
  repository mode- and parameter-based adaptation
  rich meta-information.
• Q-Fabric used for dynamic end-to-end control of
  interactive video applications and for
  distributed Internet applications (e.g.,
  scientific collaboration).

9
InfoFabric Basics

• Information channels in InfoFabric are based on
  direct source-to-sink links between subscribers.
• Channels use self-describing information items
  (PBIO).
• A service is a middleware-accessible set of
  computations applied to information items.
• InfoFabrics operation can be changed dynamically
  via run-time adaptation, taking advantage of its
  rich set of meta-information about typed
  information flows, items, services, and code
  modules
• Information with quality attributes quality
  refers to degrees of similarity and consistency
  of information acquired, processed, and
  delivered. Quality captures information
  completeness, timeliness, etc.

10
Middleware-level Adaptation Mirroring for OIS

• Adaptive replication and redundancy
• distributed algorithms decide how information is
  routed and processed (e.g., dynamic mirroring)
  based on high-level performance requirements
• Operational Information Systems
• continuous support for a companys daily
  operations
• Airline flights, crews, passengers
• OIS combines
• continuous data capture
• continuous state updates - via EDE event
  derivation engine
• short response times

11
Agile Services and Mirroring Response Time

• Adaptive event mirroring for improved performance
  and responsiveness events are mirrored to
  additional server engines to achieve low average
  delays for event derivation and publication.
• Granularity at which monitoring is performed and
  thereby level of data consistency across nodes
  are modified dynamically.
• Experiment average event update delays increase
  with increased rate of incoming client requests
  (blue), which can be avoided by moving requests
  to mirror nodes.

12
Q-Fabric (1)

• Kernel-level Quality Management
• Based on same set of mechanisms that implement
  InfoFabric.
• Lightweight, low granularity operation with
  unrestricted access to system resources.
• Inherently distributed.
• Enforced participation.

13
Q-Fabric (2)
14
Results (1) Remote Visualization

• Service behavior adjusted in response to
  Q-Fabrics resource monitoring in a distributed
  scientific visualization adapting information
  flows between nodes.

Client-aware data streaming change in latency
with varying network load
Client-aware data streaming change in event rate
with varying network load
15
Results (2) Video Conferencing

• Q-Fabric manages the resources underlying a video
  conferencing application (vic) CPU (Linux
  real-time round-robin scheduler), network (DWCS -
  Dynamic Window-Constrained Scheduler). The
  application can be adapted by varying the image
  quality in a range of 1 (low) to 95 (high).

Jitter at a client with distributed resource
management. Management at Client.
Jitter at a client with perturbation at sender
and the client, and without adaptation.
16
Results (3) Service Adaptation plus Resource
Management
Jitter with distributed resource management AND
application adaptation.
17
Results (4) Managing Power Consumption
On ARM-based handhelds without/with
Q-Fabric-based Management
18
Selected Publications

• Applications SC 2002
• Middleware SPDT98, HPDC2000, SC2001, HPDC2001,
  IPDPS 2001, Middleware 2003
• Q-Fabric NOSSDAV 2001, ACM Multimedia 2001, ARCS
  2002, ACM Multimedia 2002, HPDC 2003

19
Future Work

• Distributed algorithms for information routing
  and processing across distributed embedded
  platforms (overlay networks).
• Service disassembly, specialization, and
  reassembly to match behavior to the current
  capabilities of the underlying system.
• Employing dynamic compilation
• In place and Remote compilation to deploy
  code at certain overlay network nodes.
• Power-aware code generation, and code
  specialization to refine the parameter- and
  version-based adaptations used by InfoFabric.
• Interface with standard platforms .NET, CORBA,
  OGSA interfaces.