Optical Grids in USA Gigi Karmous-Edwards gigi@mcnc.org March 5th, 2006 International Workshop on Future of Optical Networking Outline • Motivation • GLIF • Research Challenges • OBS for Grids • EnLIGHTened Computing • Conclusions Motivation E-science and Grid Computing • E-science: global, large scale scientific collaborations enabled through distributed computational and communication infrastructure • Combines scientific instruments and sensors, distributed data archives, computing resources and visualization to solve complex scientific problems • In physics, molecular biology, environmental, Health, Entertainment, etc. • Grid computing: main enabler of E-science • Grid is concerned with "coordinated resource sharing and problem solving in dynamic, multi-institutional virtual organizations." (Foster) • E-science migrated to Grid for the reasons of affordability of high-bandwidth communication infrastructure, affordability of resources and inter-disciplinary nature of the research E-Science and Grid Opportunity • Governments world-wide promoting global E-science research programs • New Zealand, $43 million to establish the Advanced Network • United States: from Optiputer over Cyber-infrastructure to DDDAS , and DoE ($60M) • Canada: CA*net4 and i-Infrastructure • Australia: e-Research Initiative (Started Oct 2004) • Europe: 7th framework (€1B/year), GEANT2 (€93M), Netherlight, Geodise, UK E-science, etc. • Private sector setting a strong pace • Google building their own L1/L2 network • Multi $B investments from large businesses: IBM, HP, Intel,… • Smaller businesses: Grid-based enterprise applications, data centers solution, commercialization of Globus toolkit,… Advances in Optical Technologies • 1000 channels per fiber….. Experimentation with 160G per channel • Dark Fiber every where …. • Fiber is much cheaper…US Headlines: Google buys Fiber • All-optical switches are getting faster and smaller (ns switch reconfiguration) - useful for OBS • Control Plane protocols, SOA, continue to mature - but should revisit • Layer one Optical switches relatively cheaper than other technologies • Electronic Dispersion Compensation - advances in state of the art • Fiber, optical impairments control, and transceiver technology continue to advance while reducing prices New Demands on Networks Emerging High-End applications • High bandwidth connectivity of supercomputers (teraflops+) • Large file transfers, over long distances • Applications/end-users/sensors/instruments requesting optical networking resources host-to-host connections • Determinism (QoS), jitter and latency requirements • Coordination of network with computational and noncomputational resources (CPU, databases, sensors, instruments ) • Mechanisms for retrieving near-real-time information about network resources and network states • Mechanism for both advance and fast on-the-fly reservation and set-up • Low latency on-demand connection requests - real-time weather prediction • Policy and security enforcement in open scientific environments Why should this community work with Driving High-end Applications? When thinking of future Optical Architectures • Using advance applications to better understand the requirements on the network in order to architect future optical network architectures –Requires Interdisciplinary teams –Speaking the same language… • Help experiment with novel Optical technologies on experimental testbeds - NLR --> GENI (GLIF?) GLIF www.glif.is Global Lambda Integrated Facility World Map www.glif.is Visualization courtesy of Bob Patterson, NCSA. GLIF Automation? Client A Grid middleware Network Management NREN Control Plane WEB Services ? Network Management NREN Control Plane Client B Grid middleware Network Management NREN Control Plane GLIF Control Plane and Grid Middleware Integration wg Mission: To agree on the interfaces and protocols to automate and use the control planes of the contributed Lambda resources to help users on a global scale access optical resources on-demand or pre scheduled. Research Challenges Research Challenges • Coordination of resources per request for both on-the-fly and advanced reservations • Dynamic and Adaptive on-demand use of end-to-end networking resources Identification of functions and interactions between the control plane, management plane, and Grid middleware • Monitoring information of resources - identification of information, abstraction of information, and frequency of updates • Advanced software to support multiple classes of software including highlydynamic, and event-driven applications • – Dynamic lightpaths routing and resource allocation under constraints and • The more dynamic these reconfigurable lightpaths the closer they get to OBS Behavioral Control • Control/management planes interacting with middleware • Centralized vs. distributed functionality Applications Edge Routers Workflow Engines Resource Manager Co-Scheduler Resource Allocation Resource Monitoring One Definition of Control Plane “Infrastructure and distributed intelligence that controls the establishment and maintenance of connections in the network, including protocols and mechanisms to disseminate this information; and algorithms for engineering an optimal path between end points.” Draft-ggf-ghpn-opticalnets-1 Centralized vs. Distributed… Key areas for Today’s Control Plane are: 1) Provisioning Network 2) Recovery Behavioral Control Network Management (Hierarchical ) Network Management Migration NE NE NE Centralized (vertical) NE Protocols NE Protocols NE Distributed (Horizontal) Control Plane Functions OCS/OBS/OPS • Routing - Intra-domain and Inter-domain 1) automatic topology and resource discovery 2) path computation (How do we use the infrastructure) • Signaling - standard communications protocols between network elements for the establishment and maintenance of connections • Neighbor discovery - NE sharing of details of connectivity to all its neighbors (very powerful tool) • Local resource management - accounting of local available resources Centralized vs. Distributed Behavioral Control of Networks • Re-thinking control functionality in terms of (centralized or Distributed): – – – – – – – • • Information exchanged Algorithms for path computation and recovery Discovery and advertising of resources Scalability Timing Interdomain interactions Policy enabled Centralized Control: Management Plane – Management plane mechanisms rely on client/ server model, usually involving one or more management applications (structured hierarchically) communicating to each network element in its domain via a management protocol, (i.e., SNMP, Tl1, XML, etc). Distributed Control: Control Plane – “Infrastructure and distributed intelligence that controls the establishment and maintenance of connections in the network, including protocols and mechanisms to disseminate related information, and algorithms for engineering an optimal path between end points.” Control Plane Focus today 1. Network Configuration 2. Network Recovery GMPLS IP Control Protocols: Signaling: RSVP-TE, OBS Routing: OSPF-TE based on Link State protocols (LSA) Discovery: LMP OBS for Grids OBS vs. GMPLS signaling comparison, Advantage of OBS is fast reaction time • GMPLS RSVP-TE connection setup – PATH messages (from sender to receiver) – RESV messages (from receiver to sender) – Processing at each node for both message types to set up an end-to-end connection. • GMPLS RSVP-TE connection tear down – PATHTear message (from sender to receiver) – RESVTear message (from receiver to sender) – Processing at each node for both message types to set up an end-to-end connection. –Trade-offs • Two RTT to create and release end-to-end connection, However, it is a Reliable connection • OBS signaling can react much quicker - needed for nearrealtime reaction to events for High-end applications OBS more efficient use of the network than OCS OBS Control Plane Signaling Requirements Provides flexibility • Capability to send data w/out waiting for RTT - reduction of latency • Implicit resource release – Setup message contains duration of connection – Resources are released automatically w/out second control message • Explicit resource release – Separate release message is sent from application if duration of data transfer is unknown • Capability to behave in Connection Oriented manner – Wait for end destination ack prior to sending data • Variable payload sizes - faster switches and faster signaling processing – Very small payloads - packet – Long lived lightpaths • Hop-by-hop Path Fast Switch Reconfiguration • OBS signaling – Setup message processing - microseconds in hardware • Message parsing • Forwarding table look-up – Cross-connection - dependent on switching technology – MEMs - 12 - 15 ms – Polymer-based switches and SOA based - ns – Etc. – Advantages of Fast Switch Re-configuration times … – The smaller the payload … could be packet – The smaller the offset time – Quicker reaction to changing environments US Grid Testbeds • • • • • • • Optiputer TeraGrid Dragon Cheetah OmniNet UltraLight Enlightened EnLIGHTened Computing EnLIGHTened Computing Highly-dynamic Applications Driving Adaptive Grid Resources NSF seed funded project EnLIGHTened team significance key Institutions collaborating on the research efforts • • • • MCNC (Network research), -PI -Gigi Karmous-Edwards, Yufeng Xin, Steve Thorpe, Bonnie Hurst, Lina Battestilli, Mark Johnson, Joel Dunn LSU (Application and Grid research), PI -Ed Seidel, PI - Gabriele Allen, PI - Seung Jong (Jay) Park , Jon Maclaren, Andrei Hutanu, Lonnie Leger (Sub-Contract to MCNC) Renaissance Computing Institute, RENCI (Grid Middleware research): (a joint institute between UNC, Duke and NC State ), Dan Reed, Lavanya Ramakrishnan, NCSU (Network research), Savera Tanwir, Harry Perros, John Moore Key Partner Institutions (cost share) • • • • • • • Cisco, Javad Boroumand, Russ Gyurek, Wane Clark, Kevin McGratten AT&T Rick Schlichting, John Strand, Matti Hiltunen SURA Gary Crane Naval Research Lab (NRL) , Hank Dardy Calient Networks, Olivier Jerphagnon, Ron Mackey UCSB/Calient, John Bowers NLR EnLIGHTened Computing connectivity diagram with partners To Asia To Canada To Europe SEA POR BOI Chicago CAVE wave OGD EnLIGHTened wave (Cisco/NLR) DEN KAN SVL PIT CHI CLE WDC Cisco/UltraLight wave L.A. LONI wave San Diego Raleigh TUL DAL Baton Rouge HOU Members: - MCNC GCNS - Renaissance Comp. Inst. - LSU CCT Official Partners: - AT&T Research - SURA - NRL - Cisco Systems - Calient Networks - NLR NSF Project Partners - OptIPuter - UltraLight - WAN-in-LAB - DRAGON International Partners •LUCIFER •G-Lambda -GLIF Enlightened/LUCIFER Sister Projects - Similar Goals Testbeds ♣ 3 EU NRENs are partners + 3 national test-beds + 3 research networks in US and Canada + 5 expressed interest through LoIs ♣ These community representatives are willing to monitor project progress, collaborate and exploit its results GLIF www.glif.is Conclusions Further Reference IEEE Communications Magazine Feature Topic Optical Control Plane for Grid Networks: Opportunities, Challenges and the Vision http://www.comsoc.org/pubs/commag/cfpcommag306.htm Guest Editors: Gigi Karmous-Edwards and Admela Jukan March, 2006 Conclusions 1. 2. 3. 4. 5. 6. 7. 8. Control Plane research is vital to meeting future generation Grid computing - with a strong focus on “vertical integration” GLIF resources should be used for both network research and E-applications (morphnet concept) Reconfigurability is essential to bring down cost and meet application requirements. OBS or OCS Currently, we have a view of the behavior of potential future enterprise applications by focusing on the needs of high-end driving applications, but it is also important to understand the requirements of Industry. Next generation networks could be vastly different than today’s mode of operation - should not constrain research to today’s model The Research networks are the ones that will take these bold steps not the carriers… apply lessons learned to production quickly. International Collaboration is a very Key ingredient for the future of Scientific discovery - The Optical network plays the most critical role in achieving this! The OBS community can benefit by experimenting on national scale networks like NLR and working with E-science applications