The Road towards Optical Networking. SURFnet. University of Amsterdam SARA NIKHEF UCL

Transcription

1 The Road towards Optical Networking Cees de Laat λaat EU SURFnet University of Amsterdam SARA NIKHEF UCL

2 What is this buzz about optical (2 of 15) networking Networks are already optical for ages Users won t see the light Almost all current projects are about SONET circuits and Ethernet (old wine in new bags?) Are we going back to the telecom world, do NRN s want to become telco s Does it scale Is it all about speed or is it integrated services

3 VLBI (3 of 15)

4 Know the user (4 of 15) # of users A B C ADSL GigE LAN BW requirements A -> Lightweight users, browsing, mailing, home use B -> Business applications, multicast, streaming, VPN s, mostly LAN C -> Special scientific applications, computing, data grids, virtual-presence

5 What the user (4a of 15) Total BW A B C ADSL GigE LAN BW requirements A -> Lightweight users, browsing, mailing, home use B -> Business applications, multicast, streaming C -> Special scientific applications, computing, data grids, virtual-presence

6 So, what s up doc (5 of 15) Suppose: Optical components get cheaper and cheaper Dark (well, dark?) fibers abundant Number of available λ/user -> Speeds of 10, 100, 1000 Gbit/s make electrical domain packet handling physically difficult Gbit/s = 30 ns = 15 meter fiber QoS makes no sense at these speeds Cost per packet forwarding lower at L1 / L2 Then: Long term view ---> full optical λ provisioning for grid applications How low can you go

7 Optical networking, 3 scenarios (6 of 15) Lambdas for internal ISP bandwidth provisioning An ISP uses a lambda switching network to make better use of its (suppliers) dark fibers and to provision to the POP's. In this case the optical network is just within one domain and as such is a relatively simple case. Lambda switching as peering point technology In this use case a layer 1 Internet exchange is build. ISP's peer by instantiating lambdas to each other. Is a N*(N-1) and multi domain management problem. Lambda switching as grid application bandwidth provisioning This is by far the most difficult since it needs UNI and NNI protocols to provision the optical paths through different domains.

8 Current technology + (re)definition Current (to me) available technology consists of SONET/SDH switches which transport Ethernet DWDM+switching coming up Starlight and CANET use for the time being VLAN s on Ethernet switches to connect [(not)exactly] two ports So redefine a λ as: (7 of 15) a λ is a pipe where you can inspect packets as they enter and when they exit, but principally not when in transit. In transit one only deals with the parameters of the pipe: number, color, bandwidth

9 (8 of 15) Application Middleware Application Middleware High bandwidth app Transport Transport UvA SURFnet5 3 rd party carriers ams chi Switch Router GbE Router Router GbE Router Router 2.5Gb lambda Lambda Switch Lambda Switch lambda for high bandwidth applications Bypass of production network Middleware may request (optical) pipe RATIONALE: Lower the cost of transport per packet GbE UBC Vancouver Router Switch

10 Architectures - L1-3 (9 of 15) R R R

11 Other architectures - Distributed (10 of 15) virtual IEX es vlan a vlan b vlan d vlan a vlan b vlan c Problem: vlan tag distribution ==> gmpls

12 Distributed L2 (11 of 16) Layer 2 VPN Univ A Univ X Univ B Univ Y SN5 CHICAGO lambda SN5 A DAM

13 (12 of 16) CA*net 4 Architecture Edmonton CANARIE GigaPOP ORAN DWDM Carrier DWDM Saskatoon Victoria Calgary Vancouver Regina Winnipeg Thunder Bay Quebec Montreal Ottawa St. John s Charlottetown Fredericton Halifax Seattle Chicago CA*net 4 node) Possible future CA*net 4 node Toronto Windsor Boston New York

14 Nested Lightpaths University 15 Carrier A Regional Network D 13 AS 1 2 AS AS Regional Network C 8 6 AS 6 University LightPath switch Carrier B STAR LIGHT router

15 (12a of 16) Mini-IX with OON Grooming Agents Optional Signal and Control Plane Agents Grooming Agents OSPF OSPF OBGP OBGP OBGP OBGP Customer A Customer A OBGP OBGP Customer B OSPF OSPF Customer D Switch Matrix Customer C X Customer A partition X Customer C X X X X X X Note: Switch and Signal Planes are partitioned to respective condominium owner. All objects are controlled remotely via SOAP X X Customer C partition X X Switch Control Agents

16 Recursive Onion control model of administrative domains Selector Switch Domain X Distributor Switch (13 of 16) Domain Y AAA AAA

17 Daisy Chain control model of administrative domains Selector Switch Distributor Switch (13a of 16) AAA AAA Domain X Domain Y

18 (13b of 16) XML based AAA request for Lightpath element -<AAARequest version="0.1" type="service"> - <Authentication protocol="simple"> <Identity>Arie</Identity> <AuthenticationData type="password">abc123</authenticationdata> </Authentication> - <ServiceData> - <SwitchData> <Source> </Source> <Destination> </Destination> <Bandwidth>1000</Bandwidth> <StartTime>12:45</StartTime> <Duration>2:31:45</Duration> </SwitchData> </ServiceData> </AAARequest>

19 if ( ( Simple policy example executed by Rule Based Engine (13c of 17) ASM::RM.CheckConnection( Request::ServiceData.SwitchData.Source, Request::ServiceData.SwitchData.Destination ) && ( Request::ServiceData.SwitchData.Bandwidth == 1000 ) ) ) then ( ASM::RM.RequestConnection( Request::ServiceData.SwitchData.Source, Request::ServiceData.SwitchData.Destination, Request::ServiceData.SwitchData.Bandwidth, Request::ServiceData.SwitchData.StartTime, Request::ServiceData.SwitchData.Duration ) ; Reply::Answer.Message = "Request successful" ) else ( Reply::Error.Message = "Request failed" )

20 Clients X Windows Web Browser PDA GGF-GPA: Combined Grid and Web Services Architecture http, https. etc. Application Portals Discipline / Application Specific Portals (e.g. SDSC TeleScience) Problem Solving Environments (AVS, SciRun, Cactus) Environment Management (LaunchPad, HotPage) composition frameworks (e.g. XCAT) Python, Java, etc., JSPs Apache SOAP,.NET, etc. XML / SOAP over Grid Security Infrastructure Web Services Job Submission / Control File Transfer Data Management Monitoring Events Credential Management Workflow Management other services: visualization interface builders collaboration tools numerical grid generators etc. CoG Kits implementing Web Services in servelets, servers, etc. Apache Tomcat&WebSphere &Cold Fusion=JVM + servlet instantiation + routing Grid Protocols and Grid Security Infrastructure Grid Services: Collective and Resource Access CORBA Condor-G Data Replica and Metadata Catalog Grid X.509 Certification Authority Grid Information Service Grid Monitoring Architecture Grid Web Service Description (WSDL) & Discovery (UDDI) Grid ssh GRAM SRB/ Metadata Catalogue GridFTP MPI Secure, Reliable Group Comm. Grid Protocols and Grid Security Infrastructure (13d of 17) Resources Compute (many) Storage (many) Communication Instruments (various)

21 i The International Virtual Laboratory Grid 2oo2 (intermezzo I) September 2002 Amsterdam Science and Technology Centre (WTCW) The Netherlands A showcase of applications that are early adopters of very-high-bandwidth national and international networks What can you do with a 10Gbps network? What applications have insatiable bandwidth appetites? Scientists and technologists to optimally utilize 10Gbps experimental networks, with special emphasis on e-science, Grid and Virtual Laboratory applications Registration is open ( igrid is not just a conference/demonstration event, it is also a testbed!! Contact maxine@startap.net or delaat@science.uva.nl

22 The International Virtual Laboratoryi Grid 2oo2 (intermezzo II)

23 First experiences with SURFnet (14 of 18) pure for research Lambda 2.5 Gbit SONET λ delivered dec 2001 Took about 3 months, should be 300 ms First generation equipment delivered nov 2001 Back to back tests => OC12 limit -> 560 Mbit/s 1 unit shipped to Chicago (literally, took 3 weeks) End to end now 80 Mbit/s So, what is going on? Second generation equipment delivered in april unit shipped to Chicago (yes, it took 3 weeks)

24 First generation equipment! (15 of 18) slope = 100 ms

25 (16 of 18)

26 Layer - 2 requirements from 3/4 (16a of 18) WS fast L2 fast->slow high RTT L2 slow->fast fast WS TCP is bursty due to sliding window protocol and slow start algorithm. So pick from menu: Flow control Traffic Shaping RED (Random Early Discard) Self clocking in TCP Deep memory Window = BandWidth * RTT & BW == slow fast - slow Memory-at-bottleneck = * slow * RTT fast

27 kbyte UDP packets (16b of 18)

28 fast - slow Memory = * slow * RTT fast (16c of 18) Memory MByte fast slow For RTT = 100 ms

29 Layer - 2 requirements from 3/4 (16d of 18) WS fast L2 fast->slow high RTT L2 slow->fast fast WS Window = BandWidth * RTT & BW == slow fast - slow Memory-at-bottleneck = * slow * RTT fast Given M and f, solve for slow ===> f * M 0 = s 2 - f * s + RTT f M s 1,s 2 = ( 1 +/- sqrt( 1-4 ) ) 2 f * RTT

30 Forbidden area, solutions for s when f = 1 Gb/s, M = 0.5 Mbyte AND NOT USING FLOWCONTROL s = 158 ms = RTT Amsterdam - Vancouver or Berkeley rtt

31 Self-clocking of TCP (17 of 18) WS fast L2 fast->slow high RTT L2 slow->fast fast WS 14 µsec 20 µsec 20 µsec 20 µsec 20 µsec

32

33

34 Revisiting the truck of tapes Consider one fiber (18 of 18) Current technology allows for 320 λ in one of the frequency bands Each λ has a bandwidth of 40 Gbit/s Transport: 320 * 40*10 9 / 8 = 1600 GByte/sec Take a 10 metric ton truck One tape contains 50 Gbyte, weights 100 gr Truck contains ( / 0.1 ) * 50 Gbyte = 5 PByte Truck / fiber = 5 PByte / 1600 GByte/sec = 3125 s one hour For distances further away than a truck drives in one hour (50 km) minus loading and handling tapes the fiber wins!!!

35 (15 of 14) The END Thanks to TERENA: David Williams SURFnet: Kees Neggers UIC&iCAIR: Tom DeFanti, Joel Mambretti CANARIE: Bill St. Arnaud This work is supported by: SURFnet EU-IST project DATATAG 26/11/ /9/2002