IQ-ECho: Middleware Principles for Real-time Interaction

Across Heterogeneous Hardware/Software Platforms

Karsten SchwanGreg EisenhauerMatt WolfMustaq Ahamad(Nagi Rao - ORNLConstantinos Dovrolis)

College of ComputingGeorgia Tech

schwan/eisen/mwolf@cc.gatech.edu

http://www.cc.gatech.edu/systems/projects/IQECho/

GTGT

EmoryUniversity

EmoryUniversity

Cluster ComputerTerastream

Server

Cluster ComputerTerastream

Server

Transform Specialize

High End Users and Displays

TeragridAtlanta

Hub

High Performance Data Streaming

Scalable ServicesData

Cache

capture, transport, filter, select, sample, re-route

Large-scale CollaborativeApplications on Heterogeneous

Systems:Terastream Services and Teragrid

Data Transport

Visualization

Caching, Recovery, Logging, Security

Real-timeCollaborationand Inspection

WirelessUsers andDisplays

Data Source(e.g., spallationneutron source)

InstrumentedTestbed/FacilityLocal

Users

ORNLORNL

InstrumentedTestbeds/Facilities(e.g., for spallation neutron source)

High End Users and Displays

Real-time Collaboration: Molecular Dynamics

Requirements: Multiple

collaborators explore common data space

Personalized views, with ability to annotate and manipulate

Real-time sharing of data, even between different representations

Undamaged

Damaged

Undamaged

vs.

W / and L /

interatomic spacing

L

W

L

W

• Mechanical Engineering

• Physics

• Chemistry

•Aerospace Engineering

FCC

Twinning Plane

IQ-ECho: Middleware Principles for Network-aware Collaboration

Adaptive Peer-to-Peer Data Exchange:

IQ-ECho: High performance events:– Event-based peer-to-peer streaming data communications -binary

data exchanges (PBIO) for interactive apps (steering, real-time collaboration, …)

– Source-based filtering: IQ-services deployed to meet required application QoS, i.e., by disposition of application-specific code into remote sites and underlying platform

– Dynamic quality attributes: coordinated adaptation of platform (e.g., communication protocols) and of interactive applications

Network-awareness: adaptive communications (with Nagi Rao/ORNL, Constantinos Dovrolis/GT):– Runtime detection of congestion– Runtime response: adaptation: re-routing, concurrent paths,

coordinated protocol/application response (IQ-RUDP)

Real-time Collaboration with IQ-ECho

Filters

AdaptiveSource-based

Filtering

MultipleEvent Types

Dynamic QualityAttributes

Types of adaptation

Middleware- and/or Network-level:

Frequency

– Same amount of data but different rate

Resolution

– Same rate, different amount

Reliability

– Changing proportion of discardable packets

Multiple Connections

– Protecting critical connections from large-data traffic

Adaptive Communication

Adapt what?

– Congestion windows + data ratesIssues:– Transport cannot delegate all adaptation choices to

applications and still be fair to the network– Applications cannot delegate all adaptation to the

transport without limiting their choices or incurring difficulties (e.g., QoS translation)

Goal: – provide a mechanism to allow effective application

adaptations while remaining network-friendly

Coordinated Adaptation

Use `quality attributes’ to share information across middleware/protocol - IQ-Services

`Coordination methods’ Services/Protocol to address:– Conflicting adaptations– Combined effect of adaptation that may lead to overreaction– Limited application adaptation granularity– Others, ...

Problems important in networks where (delay * bandwidth) is large:– cost of adaptation– delay before correction of mistakes

Middleware/Protocol Interactions

IQ-Services in Middleware:– Application-relevant data manipulation:

• Data prioritizers, data filters, downsamplers

– Controlled by dynamic quality attributes

On-line Network Measurement: – e.g., Rao’s TCP-based methods

Using an Instrumented Protocol: IQ-RUDP extends Reliable UDP– TCP-friendly congestion control (LDA algorithm)– Exposes network performance metrics– Supports application-registered callbacks– Application-controlled adaptive reliability

Middleware Architecture

Evaluation of Coordinated Adaptation

How effective is coordination in two-layer adaptations?

– Metric is “smoothness” of delay over time– Evaluate three cases where coordination is necessary– Hold application traffic pattern constant, vary network

bandwidth• iperf used to generate background traffic

– Hold network bandwidth constant, vary application traffic

• Emulate content delivery server using MBONE trace– Drive adaptations using callbacks on error ratio

Example: Conflicting Adaptations

No Coordination– Transport unaware of adaptation– All packets sent regardless of priority– More unmarked packets delivered– Larger delay for marked packets

Coordination– Transport can drop non-priority packets– Better delay/jitter for high priority packets– Average delay improves due to spacing

Conflicting Adaptations

IQ-RUDP (on right) achieves lower avg delay

(emulation results)

Example: Metadata-based Filtering

Without IQ-ECho Level Adaptation

0

10

20

30

40

50

60

70

80

40 90 140 190 240

Time(s)

Fra

me

Ra

te(f

/s)

Message Deliver Rate(WithoutAdaptation)

Perturbation Introduced

IQ-RUDP (on right) achieves substantially higher frame rate (measured results)

With IQ-ECho Level Adaptation

0

10

20

30

40

50

60

70

80

90

4160 4180 4200 4220 4240 4260 4280 4300

Time(s)

Fra

me

Ra

te(f

/s)

Message Deliver Rate(WithAdaptation)

Perturbation Introduced

Conclusions and Status

Key technologies:• Adaptive, lightweight middleware services

– software release of IQ-ECho available soon (installation at ORNL in progress)

• Coordinated middleware/network (re)actions (through quality attributes)

– generalizes to other network efforts (e.g., Net100)

• Heterogeneous, distributed collaboration with high end data streams:

– Smartpointer (MD - SC2002)

Evaluation on wide area networks• Internet, GT/ORNL link (yet to come)

Focus on integration• MxN services, AG 2.x

Ongoing Efforts and Leverage

Deployment and Evaluation (Year 3):– Realistic applications and testbeds:

• deploy remote collaboration infrastructure (with ORNL) and experiment across ORNL/GT Gigabit Testbed (with N. Rao, ORNL)

• experiment with other data sets (e.g., spallation neutron source), other protocols, other network measurement methods (NSF/DOE)

– CCA/OGSI integration:• CCA integration: use MxN service as challenge example (joint with James

Kohl - ORNL)• OGSI integration challenge example: remote graphics services for AG->

OGSI, directory services

Leverage: CERCS and GT/ORNL efforts:– NSF Netreact project

• integrated network measurement - w. Dovrolis, Rao– NSF XML project - dynamic metadata– Teragrid and GT/ORNL and GT/NRL: high end network links

Future Work

Platform resources: effective deployment: – Servers: real-time data transformation with the

Terastream server (utilizing end points!)– Networks:

• application-specific processing on programmable routers• utilizing high end links, e.g.,Teragrid

Dynamic data interoperability:– heterogeneous data, using XML markups– automating XML/binary translations

Protected services:– controlling IQ-service execution

Publications

Qi He and Karsten Schwan, “IQ-RUDP: Coordinating Application Adaptation with Network Transport”, High Performance Distributed Computing (HPDC-11), July 2002.

Matt Wolf, Zhongtang Cai, Weiyun Huang, Karsten Schwan, ``SmartPointers: Personalized Scientific Data Portals in Your Hand'', Supercomputing 2002.

Fabian Bustamante, Patrick Widener, Karsten Schwan, ``Scalable Directory Services Using Proactivity'', Supercomputing 2002.

Patrick Widener, Greg Eisenhauer, Karsten Schwan, and Fabián E. Bustamante, "Open Metadata Formats: Efficient XML-Based Communication for High Performance Computing", Cluster Computing: The Journal of Networks, Software Tools, and Applications, 2003.

Greg Eisenhauer, Fabián Bustamante and Karsten Schwan, "Native Data Representation: An Efficient Wire Format for High-Performance Computing", IEEE Transactions on Parallel and Distributed Systems, 2003.