Transcript Resource Optimization in Hybrid Core Networks with 100G Links

Resource Optimization in Hybrid
Core Networks with 100G Links
Malathi Veeraraghavan
University of Virginia
[Collaboration with Admela Jukan]
Date: Sep. 28, 2009
1
Outline
•
•
•
•
•
•
•
•
Problem statement
Hybrid network architecture
Traffic monitoring/characterization
Traffic engineering
Network engineering
Traffic analysis & simulations
DOE testbed
Workplan
2
Big picture question
• As
– link rates reach 100Gbps
– switching capacities reach tens of Tb/s
• Do we need a new backbone network
architecture?
3
Hybrid node architecture
Unfolded view
Ethernet
control-plane
port
Hybrid node
Node controller
Administrative
interface (CLI, TL1)
Signaling
(provisioning)
protocol
SNMP
MIB+agents
Routing protocol
Data plane
Input
interfaces
Layer-3 (IP)
router (1Tbps)
Carrier
Ethernet
switch (10Tbps)
WDM switch
100Tbps
Output
interfaces
Hybrid network architecture
“put the traffic where the bandwidth is”
REF
Hybrid Traffic engineering (TE) system
Modify routing metrics and/or
write routing table entries
Hybrid Network Engineering (NE) system
Traffic monitoring/
characterization system
K circuits: IP-routed partition
N-K: Dynamic-circuit partition
Hybrid
Node
“put the bandwidth where the traffic is”
Obtain data
Request dynamic circuit
setup/release
DOE-implemented control-plane
software systems
Hybrid
Node
Hybrid
Node
Shared single core pool
of N fibers
Hybrid
Node
Hybrid
Node
REF: Report of US/EU Workshop on Key Issues and Grand Challenges in Optical Networking.
Available: http://networks.cs.ucdavis.edu/mukherje/US-EU-wksp-June05-Final-Report.pdf
5
Traffic monitoring/characterization system
Hybrid Traffic engineering (TE) system
Hybrid Network Engineering (NE) system
Traffic monitoring/
characterization system
DOE-implemented control-plane
software systems
Traffic monitoring/characterization system
K circuits: IP-routed partition
• Reads parameters necessary only for TE/NE applications
N-K: Dynamic-circuit partition
Hybrid
• Characterizes traffic matrix
Node
Hybrid
• Not itself a general-purpose
monitoring system such as PerfSONAR
but could interface with such systems to obtain data
Node
Hybrid
Node
Shared single core pool
of N fibers
Hybrid
Node
Hybrid
Node
Report of US/EU Workshop on Key Issues and Grand Challenges in Optical Networking.
Available: http://networks.cs.ucdavis.edu/mukherje/US-EU-wksp-June05-Final-Report.pdf
6
Hybrid Traffic Engineering system
Hybrid Traffic engineering (TE) system
Hybrid Network Engineering (NE) system
Traffic monitoring/
characterization system
DOE-implemented control-plane
software systems
Hybrid Traffic Engineering (TE) system
K circuits: IP-routed partition
• Obtains data from Traffic monitoring/characterization system
N-K: Dynamic-circuit partition
Hybrid
• Computes optimal routes for load balancing
Hybrid
• Issues CLI commands toNode
hybrid nodes to modify routing metrics
and/or write routing table entries
Node
Hybrid
Node
Shared single core pool
of N fibers
Hybrid
Node
Hybrid
Node
Report of US/EU Workshop on Key Issues and Grand Challenges in Optical Networking.
Available: http://networks.cs.ucdavis.edu/mukherje/US-EU-wksp-June05-Final-Report.pdf
7
Hybrid network engineering system
Hybrid Traffic engineering (TE) system
Hybrid Network Engineering (NE) system
Traffic monitoring/
characterization system
DOE-implemented control-plane
software systems
Hybrid Network Engineering (NE) system
K circuits: IP-routed partition
• Obtains
monitoring/characterization system and
N-K: Dynamic-circuit
partitiondata from Traffic
Hybrid
DOE-implemented control-plane
Node software
Hybrid
• Determines if thresholds are crossed to trigger setup/release of
Node
dynamic circuits
Hybrid • If triggered, sends request for dynamic circuit setup/release to
DOE-implemented
software
Sharedcontrol-plane
single core
pool
Node
• Commands Hybrid TE system to make routing table updates
of N fibers
Hybrid
Node
Hybrid
Node
Report of US/EU Workshop on Key Issues and Grand Challenges in Optical Networking.
Available: http://networks.cs.ucdavis.edu/mukherje/US-EU-wksp-June05-Final-Report.pdf
8
What is today’s backbone
network architecture
• Two separate networks (separate nodes; separate links)
– IP-routed service (black links)
– SDN for dynamic-circuit service (blue links)
SDN switch
IP router
Summer 09 ESnet map: http://www.es.net/pub/maps/current.pdf
9
Nodes and wide-area links
• IP routed network
– 16 routers (MX960 or M320)
– 20 inter-city (WAN) links – 10GbE
• SDN network
– 17 SDN switches (MX 960)
• two in New York, two in Chicago
– 29 inter-city (WAN) links – 10 GbE
• Metro rings
– SUNN, CHIC, NEWY
10
Future hybrid network:
Part of the ESnet topology
PNWG
100 GbE
DENV
10GbE
PNNL
LANL
SNLA
ALBU
SUNN
•
For simplicity, let us drop the
WDM switch from our hybrid
node and just have one
connection-oriented switch
(e.g., carrier Ethernet, MPLS,
WDM, SONET)
ELPA
Hybrid node
Connection-Oriented (CO) switch
IP Router
11
Hybrid architecture
DENV
1 GbE
SDN switch
NREL
IP router
10 GbE
NREL
10 GbE
CO switch
IP router
100 GbE
10 GbE
DENV
What rate?
IP router
SDN switch
ALBU
1 GbE
SNLA
10 GbE
LANL
Current: Separate
CO switch
IP router
10 GbE
SNLA
100 GbE
ALBU
dynamic circuit LANL
Future: Hybrid
12
Comparison
Current architecture
Hybrid (future) architecture
Separate inter-city (backbone)
links – Two or more 10GbE
One inter-city (backbone) link
- 100 GbE
Most enterprise access links
terminate only on IP router –
10GbE or 1GbE
(some on both IP router and SDN
switch)
Enterprise access links terminate
only on CO switch – 10 GbE or 100
GbE
Backbone links are 10GbE (same as Can we make IP router-to-router
some enterprise access links)
circuits of lower rate?
Link loads are low now; nevertheless 10Gb/s
links are required to avoid backbone links
from becoming bottlenecks
13
Questions
• Backbone router to backbone router
circuits:
– bandwidth-limited?
• not mandatory to limit circuit bandwidth if CO switch
is MPLS or carrier Ethernet
– bottleneck link implication on long-flow TCP throughput
– if so, what rate?
• Access circuits (enterprise-to-backbone):
– what rate?
14
IP datagram flow in
future hybrid network
PNWG
100 GbE
DENV
10GbE
PNNL
LANL
SNLA
ALBU
SUNN
ELPA
Hybrid node
Connection-Oriented (CO) switch
IP Router
15
Implication
• Packet forwarding:
– CO switches handle packets at intermediate nodes
– Backbone IP routers involved only at source and
destination PoPs
– “Cut-through” paths
• By reducing number of packets handled by IP
router:
– its’ switching capacity can be smaller
– access and backbone circuit rates between
access/backbone routers can be lower
– low rates will not affect short flows, but will impact long
flows
– can long flows can be redirected to CO network?
16
Hybrid architecture
feasibility study
• Traffic analysis
• Simulations
– to enable planning of node size (port
speeds, switching capacities)
17
Traffic analysis
• Analyze Netflow data
– characterize “short” and “long” flows
• flow inter-arrival time distributions
• flow duration distributions
– need to characterize lengths of silences
between long flows generated by an
enterprise
– determine what applications generate
“long” flows
18
Simulations
• Use Netflow data-validated input traffic model
for ESnet
• Create a simulation model of an alternate “hybrid”
ESnet, one in which inter-router links are of lower
rates than in deployed network but used only for
short flows
• Long flows are directed to CO network with
dynamic circuit setup between enterprise routers
• Obtain performance metrics
– short flows
– long flows
19
SNMP traffic analysis
• Obtain SNMP link traces for
backbone and access links
• Are some links that are not on many
shortest paths loaded at lighter
levels than others?
• Time-of-day dependencies?
• Correlations?
• Stationarity?
20
DOE-provided testbed
• Demonstrate operation of hybrid networks
on DOE-provided testbed with 3 to 4
nodes.
– Hybrid nodes (provided to us):
• IP router (layer 3) + SDN switch (layer 2.5, 2 or 1)
• DOE-implemented control-/management-plane
software
– OSCARS, PerfSONAR
– Implemented by us:
• Traffic characterization software
• Traffic/networking engineering management systems
21
Work plan
• Year 1: Architecture and Analysis
• Year 2: Algorithm design and
software implementation
• Year 3: Prototyping on DOE-provided
testbed
22
Requests to ESnet
•
•
•
•
Netflow data
Raw SNMP data files
RIBs (BGP and IGP) and IGP metrics
Can a full-mesh of MPLS tunnels - without bandwidth allocations –
be created on existing ESnet?
–
•
Simplifies the process of generating traffic matrices (LSP SNMP data)
Are there experimental PCs at PoPs for researcher use?
–
Virtual machines?
– PCs connected to both IP router and SDN switch?
– Run BWdetail (data-plane experiments)
• Is http://netinfo1.es.net to be maintained? If so, the Public
Access link gives error messages. Will Level2 access be
useful to us?
23