Center for Experimental Research in Computer Systems

Spring 2007 IAB Meeting

Karsten Schwan, Calton Pu, Douglas Blough, Sudhakar Yalamanchili

IUCRCERCS NSF Industry University Co-operative Research Center

Mission

Scientific

.Grid

Scientific

.Grid

EnterpriseEnterprise

EmbeddedEmbedded

Lead the innovation of new information and computing technologies, to construct the interactive information systems of the future, and to create the intellectual capital that

can advance these technologies and fuel future advances.

Information anytime, anywhere Timeliness!Quality!Security!Robustness!

Remote access toInformation System

Strategic Thrusts - Highlights

– Scientific/Technical Computing -- Dynamic Data Management and• GT: IHPCL Laboratory (e.g., new cluster machines, including substantial Intel donations for

education and for multicore computing initiative)• DOE:ORNL, Sandia: High Performance I/O initiative; involvement with startups• Cisco (MPI and IB QoS); Dell; HP/Intel (Gelato, Itanium donations); IBM (VMM power

management, Cell SUR grant), RNet communication processor design• News: Multicore Focus (HP, IBM, Intel); Ongoing I/O Initiative, Virtualization in HPC (ONRL,

Sandia, UNM)

Strategic Thrusts - Highlights– Enterprise Computing -- Autonomic/Adaptive and Service-Oriented

Systems:• IBM, Intel, TCS (autonomic and critical enterprise systems, dynamic content

distribution/event-based systems, SOA, virtualization – hypervisor scaleout, I/O virtualization, trusted passages, metering, power management, failure diagnosis and fault containment)

• HP (deployment and management, system monitoring, risk-based control, stream data mining)

• Worldspan (runtime behavior detection and QoI, virtualization)• LogicBlox (Dynamic code generation for efficient data access)• Delta, Raytheon (policy/performancerobustness, runtime behavior modeling)• Cisco, Intel (network data services, heterogeneous multicore, IB network

virtualization)• News/Outreach: IBM SUR (joint with Ohio State), OSU industry partners, OSU

IAB meeting, exploring new links: Benchmark, Earthlink, McKesson, NSF CRIAirportLAN

AirportLAN

AirportLAN

AirportLAN

Cluster ComputingReal-Time

Information Processing

Cluster ComputingReal-Time

Information Processing

Optimization

Baggage StatusBaggage Displays

Crew andEquipmentStatus

FAA FlightData

Passenger paging andresponse

GateReaders

Wide-area

Transport

Real-time Decision Tools

Scalable Robust Services

StorageDatabases

capture, display, transport, filter, transform

Real-Time Information Transport

Visualization

Real-timeSituation

Assessment

High Performance Computing

OperationalFlight Displays

Strategic Thrusts - Highlights– Embedded Systems/Architecture

• Boeing (testing, software correctness)• Intel (in-vehicle computing and lightweight methods for system virtualization, system-level power

management, computer architecture)• Motorola (middleware for pervasive and mobile applications)• Federal: pervasive applications (transportation, robotics), upcoming Cyberphysical Systems program• IBM, Intel (network processors and heterogeneous multicore)• Sony (gaming applications)• News: Korea program in Embedded Systems, Samsung educational program, Robotics Center liason,

Sensor (OSU) and MobiEMU testbeds

image quality,end-to-end delay,jitter, loss rate

throughput,response times

CERCS Personnel

• Faculty– Mustaque Ahamad, Mostafa Ammar, Doug Blough, Constantinos

Dovrolis, Greg Eisenhauer, Richard Fujimoto, Ada Gavrilovska, Alexander Gray, Mary Jean Harrold, Hsien-Hsin Lee, Wenke Lee, Ling Liu, Gabriel Loh, Pete Manolios, Alex Orso, Henry Owen, Santosh Pande, Milos Prvulovic, Calton Pu, Kishore Ramachandran, Jay Ramanathan (Ohio State), Rajiv Ramnath (Ohio State), George Riley, David Schimmel, Karsten Schwan, Olin Shivers, Matthew Wolf, Hongyan Zha, Sudhakar Yalamanchili, Ellen Zegura

• Research Staff– Steve Ferenci, David Hilley– Supported by DARPA, DOE, NSF, (CoC), (ECE)

• Associated Faculty/Researchers– David Bader, Tucker Balch (Robotics), Patrick Bridges (UNM),

Robert Butera, Steve DeWeerth, Irfan Essa, Phil Hutto, Byron Jeff (Clayton State), Scott Klasky (ORNL), Kang Li, Sung Kyu Lim, Arthur Maccabe (UNM), Vincent Mooney, Jeff Nichols (ORNL), Krishna Palem, Kalyan Perumalla (ORNL), Jeff Vetter (ORNL), Patrick Widener (UNM)

Industrial Relations

• IUCR CERCS Center – Contributors (GT): Boeing, Cisco, Delta, DOE, HP, IBM, Intel,

LogicBlox, TCS, Worldspan– Industry Workshops and Industrial Advisory Board

• Joint initiatives - e.g., TIE grant with UFL, expansion to Ohio State (joint curriculum/facility efforts), planned expansion to UNM

• Internship Program– Amazon, ATT, CISCO, Delta, (DoCoMo), DOE, Google, HP, IBM,

Intel, Microsoft, Motorola, NetApp, Radisys, TCS, VMWare, Worldspan

• Evolving relationships:– ATT, DoCoMo, Microsoft, Motorola, NetApp, Netronome, Raytheon,

RNet, VMWare, Xilinx

Overview - Current Industry Engagements

• One slide per ongoing project• Federal projects (e.g., joint work with DOE ORNL)

elided, so, few HPC efforts described• Same order: HPC, Enterprise, Embedded

HPC: Cisco - Infiniband-based Research

Gavrilovska, Schwan, Wolf

• Mechanisms for delivering end-to-end QoS levels in challenging settings:– Multi-core nature of future HPC nodes– I/O limitations in high-performance infrastructures– End-to-end virtualized environments

• Two main efforts under current investigations:

– Data virtualization: ‘Datatap’ mechanism on top of low-level IB verb interface to (1) extract data from IB infrastructure, (2) middleware mechanisms to support dynamic extensions for service-oriented applications, (3) dynamic, resource-aware routing and data distribution to meet application QoS requirements.

– Platform virtualization: (1) integrate x86-based virtualization solutions into Infiniband settings, (2) develop mechanisms for end-to-end QoS for VM-to-VM interactions by improved and dynamic resource management and scheduling mechanisms.

Additional Effort :RNET – high end NIC for science applications

Enterprise: Elba Project – HP Labs

Calton Pu

• Apply code generation techniques to automate large system deployment, measurement, evaluation, and management

• Collaborative work (1 faculty-Pu, 1 industry-Sahai, 6 PhD stud., 4 MS stud., 1 undergrad.)

• 8 published papers in 2 years, several more in the pipeline

AutomatedSystem Mgmt

(1) Design

(2) Code Generation

(4) Evaluation& Analysis

(5)Reconfiguration

(3)Deployment

Current work: (step 4 above)Evaluation of 3-tier benchmark (RUBiS) using generated scripts(millions of lines of deployment,measurement, and analysis scripts)

Enterprise:Robust Delivery of Quality Data - Worldspan

Client 1

Message queueClient 2

Client 3

Clearinghouse(CW)

AirlinesGDS

Karsten Schwan, Mohamed Mansour, Jay Lofstead

Problem: Complex GDS with potentially unanticipated behaviorsExample: Variable search times due to caching effectsSolution:Runtime behavior detection, model construction, and mitigationSpecific approach:Mitigation via request reordering

Enterprise:Runtime Behavior Diagnosis – Delta Air Lines

Sandip Agarwala, Mohamed Mansour, Karsten Schwan

• Investigation of multiple enterprise architectures– Revenue Pipeline, delta.com, DNS

• Path detection in complex systems• Autonomic workflows• Monitoring and management in SOA systems (proposed

work)

Enterprise: Collaboration with IBM Research

Ling Liu

• Distributed Systems and Software– Dynamic Content Dissemination: Architectures and

Optimizations– Collaborators: Arun Iyengar, Fred Douglis, Isabelle Rouvellou

• Event Streams and Security– Sensor Stream Processing and Optimization (e.g., load

shedding, load balancing, motion adaptive indexing)– Event Stream Mining– Collaborators: Philip Yu, Bugra Gedik, Rong Chang

• Service Oriented Computing– Secure publish-subscribe systems, Secure Event Dissemination– Collaborators: Arun Iyengar, Liang Jie Zhang

13

Enterprise/HPC:High Productivity Computing

• Stream computing programming model– Kernels expressed in a declarative programming language– Custom hardware for accelerating data intensive kernels– Explicit interaction model: non-coherent shared memory

• Focus on applications such as retail forecasting and data analytics

LocalMemoryCache

ACC

DMA

FIFOLocalMemoryCache

LocalMemoryCache

ACC

DMA

FIFOACC

DMA

FIFO

Network (e.g., Hypertransport)

CPU CPU CPU

Run-timeKernel

Kernel

Inputs Outputs InputsOutputs

Application

Memory Sudha Yalamanchili

with LogicBlox Inc.

Enterprise/HPC: Databus: Runtime Rule Generation - LogicBlox

Greg Eisenhauer

• Fine grain retail data analysis (what-if calculations)

• Rule-based declarative language• Compile down to interpreted

“FactBus”,• Rule objects and variable objects.

Apply rules top to bottom, fallback on failure.

• Use DCG for “Just-In Time” compilation

• Initial results, speedups of 3.• Additional improvements anticipated.

Write

Test

Iterator

Lookup

BinOp

BinOp

FactBus Variables

Rules

DC

G s

ubse

t

apply

apply

apply

apply

apply

apply

DBS

trin

g

int

int

bo

ol

do

ub

le

do

ub

le

Enterprise/HPC:Scalable Hypervisors - Intel

KarstenSchwan

Enterprise:Power Management in Virtualized Systems

Platform HW

VMM

Dom0

OS

Application

VM 3

OS

Application

VM 2

VPM Channel

VP

M M

ech

anis

ms

PMPolicy

Platform HW

VMM

Dom0

OS

Application

VM 1

VPM Channel

VP

M M

ech

anis

ms

PMPolicy

Heterogeneity-awareAllocation Policy

Leverage heterogeneity in:• Performance capabilities• Power efficiency of resources• Power management support

Leverage heterogeneity in:• Performance capabilities• Power efficiency of resources• Power management support

Coordinate virtualized system management:• Enable VM management independence • Decouple virtual and physical resources for management• Introduce “soft” scaling for flexible management

Coordinate virtualized system management:• Enable VM management independence • Decouple virtual and physical resources for management• Introduce “soft” scaling for flexible management

Ripal NathujiKarsten Schwan

IBM + Intel

Enterprise:Trusted Passages on Virtualized Platforms – Intel/NSF

Hypervisor

ServiceVM

GuestVM1

Guest VM2

BE

NIC

Trust ControllerApp.

FEFE

Hypervisor

ServiceVM

GuestVM1

Guest VM2

BE

NIC

Trust ControllerApp.

FEFE

network

network

Host1 Host2

Overlay node1 Overlay node2

network

network networ

k

network

Trusted passage

Trusted passage

Mustaq AhamadGreg Eisenhauer

Wenke LeeKarsten Schwan

Run trusted services across untrusted platforms:• Trust models and trust controller mechanisms for evolving node trust• Virtual Machines Monitoring and Introspection to support trust controllers• Data Interception and Redirection as Remedial Measures

Research Question

How can we perform fault analysis at

the system-model level and make this

information accessible to developers?

Fault Propagation for Safety (Boeing)Problem

Critical avionics systems

• now use integrated modular avionics

• making fault analysis for the entire system difficult

Embedded/Enterprise:Aristotle Research Group (Mary Jean Harrold)

Testing Evolving Software (TCS) (with Alex Orso)

Problem

Changes

• require rapid modification and testing for quick release

• causing released software to have many defects

Research Question

How can we test well (to gain

confidence in changes before

release of changed software)

MaTRIX

Computes conditions test

cases must satisfy to test

changes well

FauPA

Propagates injected faults

forward to determine impact;

Traces faulty components

backward to find root cause

Embedded:Correct Software Assemblies - Boeing