1 GMPLS optical networks Malathi Veeraraghavan Professor Charles L. Brown Dept. of Electrical & Computer Engineering University of Virginia [email protected]ETRI, Korea Feb. 2009 GMPLS: Generalized MultiProtocol Label Switched networks (MPLS, SONET, WDM, SDM, VLAN)
62
Embed
1 GMPLS optical networks Malathi Veeraraghavan Professor Charles L. Brown Dept. of Electrical & Computer Engineering University of Virginia [email protected].
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
GMPLS optical networks
Malathi VeeraraghavanProfessor
Charles L. Brown Dept. of Electrical & Computer EngineeringUniversity of [email protected]
• State of the art on different applications & networks– Commercial– Research-and-Education (REN) networks
13
Observations
• "Many e-science experiments ... are optimized to provide maximum throughput to a few facilities, as opposed to moderate throughput to millions of users, which is the raison d'etre for commercial networks."
• Networks should be scalable:– Metcalfe's statement: Value of a network
increases exponentially with the number of users
14
Key difference between DRAGON and CHEETAH
• DRAGON focus:– For eScience
• Small number of users• High throughput to a few facilities
– Transfer technology to Internet2 • Implement and deploy software for book-ahead reservations
and circuit provisionining by teaming with ESNet and DANTE
• CHEETAH focus:– General-purpose commercial network goal to bring GMPLS
services to millions of users– But not with just moderate throughput, but also high-rate– Analyze GMPLS network bandwidth sharing modes (BA +
IR)– Implementation: IR
Background
• Types of switches• Types of bandwidth-sharing modes
– IP networks vs connection-oriented (GMPLS) networks
• Tradeoffs in GMPLS network modes– Immediate-request mode (e.g., Plain
Old Telephone Service)– Book-ahead (advance-reservation)
15
16
Types of switches
Multiplexing technique on
data-plane linksAdmissioncontrol in control plane?
Circuit switch (CS)- position based (port, time, lambda)
Packet switch (PS)- header based
Connectionless (CL) - no admission control
Not an option
e.g., Ethernet
Connection-oriented (CO)- admission control
e.g., telephoneSONET WDM, SDM
Virtual-circuit e.g., MPLS, ATM, PBBTE
GMPLS network switches
17
Difference between bandwidth (BW)-sharing modes
• In connectionless networks (e.g., IP)– Pre-1988 IP network:
• Just send data without reservations or any mechanism to adjust rates congestion collapses in the Internet in the 80s!
– Van Jacobson's 1988 contribution:• Added congestion control to TCP• Sending TCP adjusts rate
– TCP congestion-control pros and cons:• Pros: Proportional fairness and high utilization• Cons: No rate guarantees & No temporal fairness (job
seniority)
• In connection-oriented networks (e.g., GMPLS)– Key: Admission control
18
Bandwidth sharing modes in GMPLS networks
• Can execute admission control in two ways:– Bufferless (immediate-request)– With buffers (book-ahead is effectively the same as having buffers to hold
calls to start in the future)
• Immediate-request: M/G/m/m model– m: number of channels on a link (servers)– if all channels are occupied, reject call
• Book-ahead: M/G/m/p model– p: max number in system: advance-reservation window K = p/m timeslots– waiting time and call blocking – K cannot be : need to block calls if per-server traffic intensity can be > 1– Or engineer the system so per-server traffic intensity ≤ 1
• Difference:– Not as the names suggest: IR calls need bandwidth immediately
• Misconception: BA with book-ahead time of “now” IR NOT TRUE
– Instead, call duration needs to be specified to support BA mode– For IR mode, applications do not need to specify duration
19
IR mode: M/G/m/mErlangB formula
mP
u
k
mP
bb
m
k
k
m
b
)1(
!/
!/
0
: offered traffic load in Erlangs: call arrival rate
1/: mean call holding time/m: per-server traffic intensitym: number of circuitsPb: call blocking probabilityub: utilization
For a 1% call blocking probability, i.e., Pb = 0.01
m ua
24.8%58.2%84.6%
110100
417117
If m is small, high utilization can only beachieved along with high call blocking probability
20
• Example – To achieve a 90% utilization
with a call blocking probability less than 10%• BA-First schemes are needed
when m < 59
– To achieve a 90% utilization with a call blocking probabilityless than 20%• BA-First schemes are needed
when m < 32
Comparison of Immediate-Request (IR) and Book-Ahead (BA) schemes
U: utilizationK: number of time periods in advance-reservation window
m=10, U = 80%: PB = 23.6% m=100, U = 80%: PB = 0.4%
IR m=10, K=10, U = 80%: PB = 0.4%
BA
Link capacity C = 10Gbpsm = 10 if per-call allocation = 1Gbps
Needed if per-callcircuit rate is a largefraction of link capacity(e.g., 1Gbps circuits on a 10Gbps link, m = 10)
22
Relate BW sharing modes to network types
Bandwidth-sharing mechanisms
Book-Ahead (BA)(high rate per call)
Immediate-Request (IR)(moderate rate per call)
eScience networks(small number of users)
Very large (TB, PB) file transfers need high-BW and long holding time + remote viz. need to reserve other resources such as displays.Centralized control-plane solution sufficient, since call durations are high (OSCARS+DRAGON)
What applications?Centralized control-plane(DRAGON)
general-purpose networks(large number of users)
To assign 1Gb/s on 10Gb/s per file transfer, m=10, need BA mode. Need distributed control-plane solution: small durations implies high call arrival rate at same util (load)
Moderately large (100MB, GB) file transfers assigned moderate-BW (100-300Mbps)(CHEETAH)
23
References on bandwidth sharing modes
• IR mode for file transfers with moderate-BW allocation (100Mbps on 10Gbps link)– X. Fang and M. Veeraraghavan, “On using a hybrid architecture
for file transfers,” acceptedto IEEE Transactions on Parallel and Distributed Systems, 2009.
– X. Fang and M. Veeraraghavan, On using circuit-switched networks for file transfers,” in IEEE Globecom, New Orleans, LA, Nov. 2008.
– X. Zhu, X. Zheng, and M. Veeraraghavan, "Experiences in implementing an experimental wide-area GMPLS network," IEEE Journal on Selected Areas in Communications (JSAC), Apr. 2007.
– M. Veeraraghavan, X. Fang, and X. Zheng, “On the suitability of applications for GMPLS networks,” in IEEE Globecom, San Francisco, CA, Nov. 2006.
• Large-scale deployment of BA mode: (mean waiting time, blocking rate)– X. Zhu and M. Veeraraghavan, "Analysis and Design of Book-
ahead Bandwidth-Sharing Mechanisms," IEEE Transactions on Communications, Dec. 08.
– X. Zhu, M. E. McGinley, T. Li, and M. Veeraraghavan, "An Analytical Model for a Book-ahead Bandwidth Scheduler," in IEEE Globecom Washington, DC, Nov. 2007. Heterogeneous rate allocation
Is an opportunity being missed if distributed IR bandwidth sharing mode is not explored?
• Yes. Four reasons:1. Increase end-to-end rate relative to IP service; possible in the
presence of admission control (programmable patch panels to share ports)
2. Enable the creation of large-scale circuit/VC networks with moderate-rate circuits that can support a brand new class of applications
• economic value for the networking industry3. A "reservations-oriented" mode of networking to complement
4. Alternative pricing models for bandwidth• Leased lines and IP service are at two extremes• Usage based pricing• Dedicated (moderately high) bandwidth for short durations instead of low
bandwidth for all time24
To increase end-to-end rate• Problem:
– WDM allows 40Gbps/channel with 80 channels/port– But, end-to-end rate is still on the order of tens of Mbps– Why? Access link rates: both for enterprises and residences
• Inter-domain link cost:– Internet2 charges $250K/year for a 1Gbps Ethernet connection– Why so high? High router port cost and no sharing
• Router port cost:– One-port 10Gbps or ten-port 1Gbps interface card costs $150-200K
• 2007 data for local access links in US:– 1.5M T1, 183K T3, 44K OC3, 21K OC12, 2K OC48 and 2.5K OC192
• Add leased lines to terminate on a space-division switch - for moderate rate, connect to sub-Gbps ports– With admission control for ports, connect high-speed link for short
duration for single flows based on request from file-transfer apps.
State of the art on different applications & networks– Commercial– Research-and-Education (REN) networks
Spectrum of services
34
Leased line Verizon BoD eScience 10G POTS IP
New services
35
Commercial uses
• Semi-permanent MPLS virtual circuits– Traffic engineering– Voice over IP
• QoS concerns: telephony has a 150ms one-way delay requirement (with echo cancellers)
– Business or service provider interconnect • interconnecting geographically distributed
campuses of an enterprise• interconnecting wide-area routers of an ISP
service provider
36
Traffic engineering (TE)
• Since BGP and OSPF routing protocols mainly spread reachability information, routing tables are such that some links become heavily congested while others are lightly loaded
• MPLS virtual circuits are used to alleviate this problem– e.g., NY to SF traffic could be directed to take an
MPLS virtual circuit on a lightly loaded route avoiding all paths on which more local traffic may compete
• This is an application of MPLS VCs without bandwidth allocation
37
Business or service provider interconnect (leased lines)
– Virtual Label Swapping Router – VLSR• Implements OSPF-TE, RSVP-TE• Run on control PCs external to switches (since not all switches
implement these GMPLS control-plane protocols)• Communicates with switches via SNMP, TL1, CLI to configure circuits.
– Client System Agent – CSA• End system software for signaling into network (UNI or peer mode)
– Application Specific Topology Builder – ASTB• User Interface and processing which build topologies on behalf of users• Topologies are a user specific configuration of multiple LSPs
47http://dragon.east.isi.edu
Open Source DCN Software Suite
• OSCARS (IDC)– Open source project maintained by ESNet and Internet2– Uses WDSL, XML, SQL database to store reservations– Reservations accepted with 1 minute granularity
• DRAGON (DC)– NSF-funded Open source project maintained by USC ISI
EASTand MAX
• Version 0.4 of DCNSS current deployed release– https://wiki.internet2.edu/confluence/display/DCNSS
• DCN workshops offered for training:– http://www.internet2.edu/workshops/dcn/index.html
• The following organizations have implemented/deployed systems which are compatible with this IDCP
– Internet2 Dynamic Circuit Network (DCN)– ESNet Science Data Network (SDN)– GÉANT2 AutoBahn System– Nortel (via a wrapper on top of their commercial DRAC System)– Surfnet (via use of above Nortel solution)– LHCNet (use of I2 DCN Software Suite)– Nysernet (use of I2 DCN Software Suite)– LEARN (use of I2 DCN Software Suite)– LONI (use of I2 DCN Software Suite)– Northrop Grumman (use of I2 DCN Software Suite)– University of Amsterdam (use of I2 DCN Software Suite)– DRAGON Network
• The following "higher level service applications" have adapted their existing systems to communicate via the user request side of the IDCP:
– LambdaStation (FermiLab) – CMS project on Large Hadron Collider– TeraPaths (Brookhaven) - ATLAS project on Large Hadron Collider– Phoebus
• Advance reservation request and circuit provisioning at scheduled time:• End user signals IDC with a reservation request• Authenticate requester and check authorization• Request reservation (create time, bandwidth, VLAN tag)• Signaling: creation of circuit (automatic or in response to message to IDC)
Route selection,admission controlcentralized per domain at IDC
Intra-domain operations
• Using DRAGON in Internet2 DCN– NARB does intra-domain path computation after
collecting routing information by listening to OSPF-TE exchanges between VLSRs
– These intradomain paths are provided to IDC for use during resource scheduling (upto 3 path options are considered)
– 5 VLSRs serve 22 CD-CIs: “subnets of CD-CIs”– In Signaling phase, VLSR sends TL1 command to
edge CD-CI, which initiates proprietary hop-by-hop signaling to configure circuit through subnet
53
54
GOLE: GLIF open lightpath exchange
55
DOE networks
• ESnet and Science Data Network (SDN)– OSCARS: an advance-reservation system– Science Data Network: MPLS network
• UltraScience Network– Research network for DoE labs– GbE and SONET (Ciena CD-CI)– Centralized scheduler for advance-reservation calls– 5-PoP network: ORNL, Atlanta, Chicago, Seattle,
Sunnyvale– Connections to Fermi Lab, PNNL, SLAC, CalTech
• Lambdastation: CMS project– Between Fermi Lab and Univ. of Nebraska
Spectrum of services
56
Leased line Verizon BoD eScience 10G POTS IP
Plain Old Telephone Service (64kbps)Immediate-Request (IR) mode
( holding times can be shorter)Distributed path computation/admission controlHigh call handling volume
New services
CHEETAH
57
NSF-funded CHEETAH network GbEthernet and SONET
TN PoP
Controlcard
GbE/10GbEcard
GA PoP
Controlcard
GbE/10GbEcard
SN16000
Controlcard
OC192card
OC-192
GbE/10GbEcard
End hosts
NC PoP
SN16000 SN16000
GaTech
End hosts
End hosts
ORNL
OC192cards
NCSUOC192card
OC-192
UVaCUNYGbE
GbEGbEs
GbE
GbE
GbEGbE
GbE
Sycamore SN16000SONET switch with GbE/10GbE interfaces
58
Networking software
• Sycamore switch comes with built-in GMPLS control-plane protocols:– RSVP-TE and OSPF-TE
• We developed CHEETAH software for Linux end hosts:– circuit-requestor
•allows users and applications to issue RSVP-TE call setup and release messages asking for dedicated circuits to remote end hosts
– CircuitTCP (CTCP) code
http://www.ece.virginia.edu/cheetah/
59
CHEETAH network usage
Application
DNS client
RSVP-TE module
TCP/IP
CTCP/IPNIC 1
NIC 2
End Host CHEETAH software
IP-routed
network
SONET circuit-switched network
CircuitGateway
CircuitGateway
Application
DNS client
RSVP-TE module
TCP/IP
CTCP/IPNIC 1
NIC 2
End HostCHEETAH software
• Bandwidth-sharing mode:• Immediate-request mode (blocked calls fall back to IP
path)• Heterogeneous rate allocation under high loads:
• higher BW for large files than for small files• Applications:
• Common file transfers (web, P2P, CDN, storage)• attempts circuits for large files (if blocked, use IP-routed
path)• use IP-routed path for small files
60
End-to-end call setup delay measurements
• Delays incurred in setting up a circuit between host zelda1 (in Atlanta, GA) and host wuneng (in Raleigh, NC) across the CHEETAH network
• Observations:– Setup delays for SONET circuits (OC1, OC3) are small (166ms) – Setup delays for Ethernet-over-SONET (EoS) hybrid circuits are much higher