Ray Bair

High-Performance Networksfor High-Impact Science: 

High-Performance Networks for High-Impact Science Ray Bair April 10, 2003

VISION: 

VISION Science applications and specialized experimental facilities are n-way interconnected. terascale computing specialized instrumentation petascale storage high end visualization remote collaborators Office of Science networking environment can move science, especially large scale science, to a new regime. seamless collaboration among scientists seamless collaboration among scientists and experimental and computational resources scientific progress fostered through the interplay of theory, simulation and experiment

Workshop Charge: 

Workshop Charge Focus on 'What is possible in the realm of science?' Science unfettered by communications – scenarios How do the high impact applications requirements impact… Network provisioning? Network research? Middleware research? Are there alternative business models that make sense in the context of the scenarios?

Workshop Findings: 

Workshop Findings Role of Advanced Infrastructure in Realizing DOE Science Visions Enabling Middleware Research Enabling Network Research Network Provisioning Model Governance Model Path Forward Workshop Record: doecollaboratory.pnl.gov/meetings/hpnpw/

Diverse Application Characteristics: 

Diverse Application Characteristics Climate Modeling Few large data repositories, many computing sites Spallation Neutron Source Many users on tight schedules, high data rates/volumes Macromolecular Crystallography Remote collaboratory operation of many beamlines High-Energy Physics Extremely data-intensive, global-scale collaborations Magnetic Fusion Energy Sciences Complex analysis on tight schedule; real time collaboration Chemical Sciences Link high-throughput exp’ts to simulations; cross scales Bioinformatics Very large research community; many large databases

Many Shared Requirements: 

Many Shared Requirements

Advanced Infrastructure Enables DOE Science: 

Advanced Infrastructure Enables DOE Science Much of science is already a distributed endeavor, or rapidly becoming so There is considerable commonality in services needed  we can define an infrastructure for distributed science Services must allow science to scale in many ways Science paradigm shifts depend critically upon an integrated advanced infrastructure well beyond today’s Revolutionary shifts can only arise from a well-integrated, widely deployed, and highly capable distributed computing and data infrastructure All together, not just any one element

Infrastructure Middleware Services: 

Infrastructure Middleware Services Global Services Support all users, e.g., directories, certificates Community Services Underpin virtual organizations, e.g., collaboration primitives, membership Resource Services Enable a site capability to participate across the network, e.g., Grid managed resources

High Priority Middleware Research Areas: 

High Priority Middleware Research Areas Secure control over who does what Information integration and access Coscheduling and quality of service Effective network caching and computing Community services to support collaborative work Monitoring and problem diagnosis

Evolution of NetworkServices Requirements: 

Evolution of Network Services Requirements

Evolution of NetworkServices Requirements: 

Evolution of Network Services Requirements

High Priority Network Research: 

High Priority Network Research Ubiquitous monitoring and measurement infrastructure Middleware often requires understanding of the underlying network to make decisions Monitoring data must be published in a scalable format, understandable by the middleware High-performance transport protocols TCP has well-known performance limitations Research into both improving TCP and into new protocols is necessary (1)

High Priority Network Research: 

High Priority Network Research Multicast Large, distributed collaborative projects are increasingly common in DOE, e.g., Access Grid use IP-multicast is a fragile technology today Mechanisms are needed to make IP multicast more robust Guaranteed performance and delivery Need to determine what network service model will satisfy DOE research needs, and work across a variety of sites and networks Trade-off between predictability and reliability New approaches to network management needed (2)

High Priority Network Research: 

High Priority Network Research Intrusion detection Main unsolved problem is predictive analysis Based on what happened in the recent past, one can get a warning an attack is about to occur Distributed systems vs. firewalls Vetting traffic through firewalls is increasingly hard because of increased traffic, much encrypted Mechanism is needed to integrate Grid security with firewalls, so the firewall can allow authorized streams (3)

SC programs need an integrated Network Provisioning Model: 

SC programs need an integrated Network Provisioning Model Production Level Networking In support of base DOE science requirements Resources for High Utilization Science In support of challenging science applications Providing both capability networking and advanced services Resources for Network Research Easily separable for running controlled experiments Over time, services, capabilities and app’s migrate

Develop an Integrated Network Provisioning Strategy: 

Develop an Integrated Network Provisioning Strategy Integrate planning, coordination, funding and implementation across all three elements Cooperate in moving new technologies into common usage – reduce barriers Evolve a more distributed service model, as more production services feature end-to-end support Maintain an agile provisioning stance; be able to adapt to changes Develop shared visions/metrics of success

Essential Features of the Path Forward: 

Essential Features of the Path Forward Integrated Road Map Shared vision across network resources and programs Network andamp; Middleware Randamp;D Programs Geared to address issues of scale and complexity Motivated by urgent needs of DOE science Deployment program Incorporate Randamp;D advances Network Provisioning Model Flexible and dynamic Business and Governance Models Optimize services in a dynamic arena

Report of the High-Performance Network Planning Workshop: 

Report of the High-Performance Network Planning Workshop DOE Organizing Comm. Mary Anne Scott, Chair Dave Bader Steve Eckstrand Marvin Frazier Dale Koelling Vicky White Workshop Organization Support Ray Bair Charlie Catlett Bill Johnston Contributors to the Report Ray Bair, Editor Deborah Agarwal G. McDermott Arthur S. Bland Sandy Merola Julian Bunn Thomas Ndousse-Fetter Charles Catlett Harvey Newman C.W. Cork Larry Rahn David Dixon David Schissel T.N. Earnest Mary Anne Scott Ian Foster Gary Strand Dennis Gannon Rick Stevens M.J. Greenwald J.R. Taylor Jason Hodges Brian Tierney William Johnston James B. White III William Kramer Michael Wilde James Leighton Linda Winkler