Transcript Ingen lysbildetittel

1
TM8106-Optical Networking
Optical IP Switching: A Flow-Based Approach
to Distributed Cross-Layer Provisioning
Authors
Marco Ruffini, Donal O’Mahony, and Linda Doyle
Other References: M. Ruffini, D. O’Mahony, and L. Doyle, “Optical IP switching for dynamic traffic engineering in
next-generation optical networks,” in Proc. of the Optical Network Design and Modeling Conf., 2007, pp. 309–318.
By Urooj Fatima
31-10-2012
Optical IP Switching: A Flow-Based Approach to Distributed Cross-Layer Provisioning
2
Overview
•
•
•
•
Why we need hybrid architectures?
Contribution of the paper
Previous work
Optical IP Switching
– Architectures
– Mechanisms
• Simulation results
– Technical (architecture simulations)
– Economical (cost evaluation)
• Testbed implementation
– Effects on TCP and UDP
– Summary
3
Why hybrid architectures?
Problem
• Revolution in optical transmission
improvements – WDM and EDFAs
due
to
technological
– Increase in bandwidth availability
– Decrease in cost of data transfer
– Boost in development and deployment of Internet
• IP routing technology struggles to deliver the necessary
bandwidth at competitive costs - network bottleneck
Solution
• Hybrid electro-optical architectures – bridging the gap between
optical transport and electronic routing
– Solution to reduce costs at the IP layer
– Deliver new revenue generating services and applications
4
Contribution
• Most of the hybrid architectures have focused on end-to-end
lightpath provisioning
– Centralized management plane
• This paper proposes Optical IP Switching (OIS) – a hybrid
electro-optical network architecture that combines IP routing and
wavelength switching using distributed decision-making
process.
• Technical and economic analysis is reported based on
simulations
• Test-bed results – Effects on TCP and UDP
5
Hybrid Electro-Optical Solutions-I
• Optical Circuit Switching (OCS)
– Dynamic setup of optical lightpaths to transport data transparently
– Group and switch all the packets sharing a common route into
dedicated all-optical channels, bypassing some of the intermediate
IP hops
• Optical Packet Switching (OPS)
– Packet by packet routing
– Electronic processing of packet header and determination of next
hop
– Activation of optical switch to route payload in optical domain
– Not cost effective for large scale deployments
6
Hybrid Electro-Optical Solutions-II
Focus of the paper
• Optical IP Switching (OIS)
– Relates to OCS
– Creation of shortcut (cut-through connection) directly at optical level
i.e. Without optical-to-electrical conversion process
– Distributed approach - Optical paths are automatically engineered
by analyzing local traffic at each node.
– OIS nodes classify IP packets based on routing destination prefixes
instead of considering distinct IP flows.
– Aggregates the classified packets into dynamically created optical
paths
7
Previous work-I
• Standardizing optical control plane
– Developing one integrated
layers
control suite operating at different
• International bodies involved:
– ITU-T worked on ASON architecture
• basic functional requirements of ON
• Principles for UNI and NNI
– IEFT worked on GMPLS architecuture
– OIF (Optical Internet-working Forum) focuses on
implementation of the control interfaces (UNI and NNI)
actual
• These standards facilitate optical paths provisioning by
automating tasks
– Topology discovery and connection provisioning
8
Previous work-II
Examples of Optical Control Plane Implementation
• Hybrid architecture from NTT (Nipon Telegraph and Telephone)
– GMPLS is used by electronic routers to create new optical paths
during congestion
• Optical Flow switching (OFS) – in process by MIT
– Creation of highly dynamic end-to-end lightpaths (requested by
users) for traffic flow from source-to-destination LANs.
– Transparent to MANs and WANs
• Grid network architectures
Research is focusing on increasing dynamic capability of
wavelength-switched networks, although the physical layer is not
fully flexible
9
The need for automatic lightpath provisioning
• Global IP forecast – growth of cunsumer traffic will totally
dominate over business traffic
10
The need for automatic lightpath
provisioning-II
• To accomodate variable traffic demands from different users
(large customers from banks, small customers from residential
areas) automatic lightpath provisioning is required for optical
networking reconfiguration.
• Current traffic engineering operations
– mainly human driven
– suitable for reconfiguration over large time scales (weeks – months)
– Not suitable to address traffic dynamics and unpredictability
• OIS – continuous adaptation of the wavelength topology by
distributed decision making
– Suitable, scalable and allows traffic variations for short time scales
(seconds to minutes)
11
Optical IP Switching
• The idea is inherited from IP switching
– Bypassing the IP layer by creating switched connection i.e. all
packets flowing to the same destination proceeds directly without
being analyzed one after the other.
• Switching data directly in the optical domain without OEO
conversion
• Advantage
– Cost saving – optical switch ports costs tens of times less than IP
ports
• Disadvantage
– Lacks the packet granularity offered by electronic routers and
switches
– Cannot be used to offer deterministic QoS for individual flows
12
OIS Architecture
13
OIS Architecture-II
• Close intergration of IP routing and Optical switching
has advantages:
– Introduction of a mechanism to engineer both layers together
– Guarantees full backward compatibility with IP protocol – an
important characterisitic for the practical implementation of
Internet architectures
14
OIS Mechanisms-I
First Step - Traffic Analysis
• OIS node is in observation state – performs constant traffic
analysis
• Information that needs to be collected includes:
– Interface from which the packet arrived
– Output interface – which is selected by the router through longest
prefix matching algorithm
– Payload size
– Arrival time
15
OIS Mechanisms-II
• Original electronic IP switching approach
– Create an optical path as soon as IP flow of suitable size (dubbed
‘elephant flow’ in literature) is observed
– Does not scale for optical IP switching – because average elephant
flow rate differ by 3 to 4 orders of magnitude
How to increase the wavelength utilization?
16
OIS Mechanisms-III
Second step – Aggregation
• Aggregation of multiple flows into the same light path
• Typical flow routing techniques are not feasible
– Individual TCP flows are identified by 5-tuple (transport protocol,
src port, dst port, src IP address, dst IP address)
– This granularity requires routers to keep track of millions of flows at
the same time
• Solution – Aggregation Matrix
17
OIS Mechanisms-IV
Prefix-based aggregation mechanism
•
•
Basic Idea : Classification of the forwarded IP traffic using the network prefixes
stored in the IP routing table
Packet classification depends on arrival (In) interfaces (number of columns) and
departure (Out) interfaces (number of rows)
– The generic matrix cell (i,j) identifies traffic incoming from interface i, relayed through
interface j
•
Within each of the previous classes operates a finer classification by destination
prefix: in each cell of the matrix a list of destination prefixes is built up, reachable
through the corresponding interface
18
OIS Mechanisms-V
Prefix-based aggregation mechanism-II
• For each packet the router checks its destination address and
determines the output interface, using the longest prefix
matching algorithm.
• The size of the packet payload adds up to the total amount of
data carried by its matching prefix within the cell (i,j).
19
OIS Mechanisms-VI
Prefix-based aggregation mechanism-III
Advantages
•
•
•
•
The size of the routing at the upstream node is not unduly increased –
each prefix summarizes a large amount of IP flows
Upstream node rarely needs to add new entries – most of the prefix
entries in RTs of peering nodes are similar
Prefix summarization diminishes the signaling overhead – only prefixes
are signaled upstream
Traffic analysis phase is simplified – information is processed at the
granularity of the prefixes
Disadvantage
•
Information about each flow is disregarded
– Not possible to guarantee QoS to individual flows
20
OIS Mechanisms-VII
Third step - Optical Path Creation (Provisioning state)
• At decision time, the router analyzes the statistics collected in the observation
state
• It sums up the amount of data brought by the different prefixes within each cell
• Only cells whose cumulative data is over a pre-established “path threshold” (100
Mbps in this case) are considered for Optical IP Switching
• The router signals the upstream and downstream neighbors (using interfaces i
and j) checking their capability to support a new optical cut-through path and
proposing a suitable wavelength
• After both neighbors have acknowledged the request, the router passes
upstream the list of prefixes to be switched through the new optical path.
21
OIS Mechanisms-VIII
Optical Path Creation-II
•
•
•
The upstream neighbor updates its IP routing table and starts injecting
packets into the cut-through path
The field ‘Dynamic link’ is added to enable coexistance of the legacy IP
and OIS protocols without interfering with each other
Dynamic cut-through paths are hidden to the link discovery protocol of the
IP layer to avoid stability problems caused by frequent link reconfiguration
22
OIS Mechanisms-IX
Path Extension
•
•
•
•
•
One of the cases where the
outgoing interface of the selected
cell is the source of an already
existing path, the node creates an
upstream extension to an already
existing path.
This algo selects a subset of the
prefixes switched by the original
paths
Only this subset will be carried by
the new extended path –
diminishing the amount of data
transported by optical channel
decreasing channel efficiency
Longer cut-through paths increase the number of transparent hops, enhancing the
cost-saving potentials of optical switching
This algorithm plays important role in the trade-off between length of the optical
path and amount of data carried by the path.
23
OIS Mechanisms-X
Path Cancellation
• The cut-through paths carrying data rates below pre-established
‘path cancellation threshold’ are deleted.
– Path cancellation threshold < path creation th., path extension th.
• Existing cut-through paths are deleted to free resources such as
interfaces, optical ports and underexploited wavelength channels
• Only cut-through path source and destination can cancel.
• After path cancellation, the traffic that was being switched returns
to the default links
24
Architecture Simulations-I
Nothing
about Path
cancellation
threshold
• Threshold values are statically assigned
– Could be dynamically adapted to the available resources(future)
• Observation time is the time interval during which each node
examines the traces and is set equal to the duration of traffic
traces
25
Architecture Simulations-II
• Reference topology used
is GÉANT
– pan-European data
communication network
– Provides real traffic
traces
– BGP RTs – allowed
reconstruction (C-BGT
simulator)
• Reconstructed topology
shown includes:
– 23 nodes
– Average distance 797
km
• Transit traffic is 36% of
the total
26
Architecture Simulations-III
• Simulator is written in PERL
• Each node operates independently
• Initial setup:
–
–
–
–
Nodes start from an initial blank state (contrast to real networks)
No optical cut-through path setup
Receives full traffic (from the traces recorded from GÉANT)
Default links need to be provisioned to handle entire traffic
• Network Steady State:
–
–
–
–
Nodes start operating cut-through paths
Traffic redirected from default link to dedicated optical paths
Paths created and cancelled (no details about cancellation)
Initial default IP capacitiy remains unused
• Due to simulation and does not reflect real network behaviour
27
Architecture Simulations-IV
Comparison between OIS and a Centralized Architecture
• Example of Centralized Architecture – Transparent Overlay
network model
– Network administrator centrally provisions optical paths by
analysing traffic demand matrices
• Assumption on end-to-end architecture:
– Traffic demand generated by all nodes is available in a central DB
– Dedicated end-to-end path is provisioned for date rate > 100 Mbps
• For comparison with OIS
• Optimization is not operated – out of scope
28
Architecture Simulations-V
Comparison between OIS and a Centralized Architecture
(transparent switching within single domain)
Switched data
• As traffic increases, both architectures reaches 36% (transit traffic)
• OIS presents higher switching capability for lower levels of traffic
–
Optical paths are built in distributed fashion allowing better traffic aggregation
Channel Occupancy
• In favour of end-to-end provisioning architecture because of lower number of
paths created
29
Architecture Simulations-VI
Comparison between OIS and a Centralized Architecture
(transparent switching across multiple domains)
•
Extending cut-through paths transparently to the external networks
–
•
•
•
•
GÉANT becomes highly transparent core network, traffic mostly switched at optical level
and routing is mainly operated by external domains
Overall transit traffic among core GÉANT nodes is over 98%
Both protocols reached switching ratios of over 93%
Increase in network performance
Network operators cannot control and filter traffic entering and leaving the domain
30
Cost Evaluation of Presented Architectures-I
•
Focused on savings in capital expenditures
–
•
Cost Model considers:
–
–
–
•
Lower cost associated with optical switching compared with electronic IP routing with a
cost difference per port
reduction in routing equipment allowed by optical by-pass
Increase in optical devices to implement OIS
Increase of transport costs in terms of optical regenerators, longer-reach transmitters
and links, higher capacity WDM systems
All devices are bidirectional
31
Network Equipment for cost analysis
•
Cost of IP routing is
considered propotional to
the amount of data routed
ateach node rather than
to the number of ports
32
Cost Evaluation of Presented Architectures-II
Network traces collected in 2005
Observations
• Negligible difference between OIS
and Overlay
• For Low level traffic
–
–
•
Single domain
IP over WDM is advantageous
Optical paths cannot exploit optical
bandwidth, hence not cost efficient
For High level traffic
–
Transparent switched
advantageous
models
are
•
100 times traffic increase shows 20%
cost advantage
• For multidomain transparent vs
opaque savings reaches above 80%
Not Considered
•
•
OIS coexistance with other architecutres in
access and metro areas
Resiliency-associated costs
Interdomain
33
Testbed Evaluation-I
•
Analysis of dynamic transparent switching operation effect on TCP and UDP
transport protocols
Testbed setup
•
•
•
Hardware - Off-the-shelf
OIS software - click language
Core and edge nodes
– INTEL machines running Linux
– 3 GHz Pentium 4 processors
• Optical Switch
- 16 X 16 port MEMsbased
- Switching time 25 ms
34
Testbed Evaluation-II
• Dynamic creation, extension and cancellation of optical paths
might cause:
– Packet loss
• caused by the switching time of the optical devices
– Jitter
– Out-of-order arrival
• packets traveling optical paths experience negligible transit time, and
can overtake IP routed path
• Effect on TCP
– Packet jitter is negligible as long as it is not above TCP time out
– Packet loss in OIS leads to erroneous network congestion
35
Testbed Evaluation-III
Out-of-order arrival situation
• ∑ (transit times of routers bypassed) > Tgap
(1)
– where Tgap is the time gap between two consecutive packets
Tgap =
𝐵
𝑅
(2)
– B = packet size (bytes)
– R = sending data rate
Combining (1) and (2)
R>
𝐵
𝑡𝑟𝑎𝑛𝑠𝑖𝑡 𝑡𝑖𝑚𝑒𝑠 𝑜𝑓 𝑟𝑜𝑢𝑡𝑒𝑟𝑠 𝑏𝑦𝑝𝑎𝑠𝑠𝑒𝑑
• The higher the rate above threshold, the higher the number of
packets that will arrive out of order.
36
Testbed Evaluation-IV
Tests on TCP protocol
•
Congestion is simulated by increasing the transit time at the routerprovoking significant out of order arrival during switching time
Solutions
• Introducing guard time at the upstream router
delaying the sending of data on new optical
path
• Creating optical bypass on ack path
Congestion on transmission
path
Congestion on transmission
and ack paths
37
Testbed Evaluation-V
Avoidance of packet loss during path extension
•
•
•
Basic idea – avoid packets
crossing optical switch while
switching is in operation
Extension times measured
– order of 50 ms
Buffer required at the
source node
– The cost is negligible
compared to the overall
router cost
•
Good for TCP but jitter
introduced might create
disruption for UDP
38
Summary
•
•
The comparison between
overlay and OIS models
is summarized
Advantages of
distributed decision
making
– Increased scalability
– Quicker reaction times
– High support for
interdomain networking
•
Disadvantage is
probable failure to
converge to optimal
solutions
39