Title: New Networking Architectures and Technologies
1New Networking Architectures and Technologies
- Martin Zirngibl
- Director, Bell Labs Research
- mz_at_lucent.com
- Dimitri Stiliadis, Peter Winzer, Harsha Nagesh,
Vishy Poosala
2Background
- Fiber amplifier and WDM have allowed orders of
magnitude increase in point-to-point bandwidth - Data traffic dominates, but voice traffic still
creates most revenues - Problems
- Switching has become bottleneck
- Legacy networks not optimized for data traffic
- Outline of Talk
- Are there network architectures that can be
efficient for both, voice traffic and data
traffic? - How can we scale SONET/SDH networks?
- How can we scale packet switches?
3Network Architectures for Bursty Traffic
- Point-to-Point
- Logical point-to-point
- Use SONET/SDH layer for muxing
- Packet Switched
- Muxing is done in IP
- Burst Switched
- Dynamic provisioning of the network resources
- Load-Balanced
- New proposal to combine advantages of
point-to-point and packet switched.
4Statistical Multiplexing
Bandwidth demand
- Statistical Muxing aggregates Bandwidth gt Better
Efficiency - Will be used when traffic is bursty and bandwidth
expensive
C
C
C
Packet Buffer And mux
C
C
Total demand
- Where will statistical muxing be used, core or
edge? - Not Clear Because
- Edge bandwidth Bursty but Transport Cheap
- Core Traffic Smooth but Bandwidth Expensive
5Todays Network Architecture
Logical mesh Between edge and access routers
High Order Circuit Switching (STS-1)
High Order Circuit Switching (STS-1)
Data over SONET
Enterprise Data Access DSO/DS1/DS3
HUB
POP Ethernet Network
SONET/WDM Optical Core
Low Order Circuit Switching (VT1.5)
Enterprise Voice DSO/DS1
TDM
Core Routers
Edge Routers ATM/Frame Relay
Circuit switched core SONET/SDH optical
Circuit-switched access network
Packet aggregation on customer premise
Packet aggregation At edge of core network
6Statement of the Problem
- N nodes want to exchange packet traffic
- Each node wants to deliver and receive packets
at an average rate C - Packets from any node may go to any other
node(s) at a rate(s) between 0 and C
C
C
C
C
1
2
C
Demand X to Y
3
4
C
C
C
C
time
7Solution 1 Point-to-Point Bandwidth of Cbetween
all Nodes
- Requires Transport Network with N(N-1)C
Bandwidth - Logical Point-to-Point can be provided through
SONET network - - Very inefficient for bursty traffic
- Transport Network Static
- No Packet Switching Necessary in Intermediate
Nodes (Single Hop Routing)
C
C
C
1
2
C
3
4
C
C
C
C
gt Preferred Solution When Bandwidth is Cheap or
Traffic is Static
8Solution 2 Multi-Hop Packet Switching
- Nodes Are Packet Routers
- Each Node is Only Connected to Nearest Neighbors
- Requires Transport Network with NC Bandwidth
- Very efficient for bursty traffic
- Transport Network Static
- - Intermediate Nodes Complex and Costly
C
C
C
1
2
C
3
4
C
C
C
C
gt Preferred Solution When Bandwidth is Expensive
and Traffic is Bursty
9Solution 3 Burst Switching
Switch Transport bandwidth the Follow Traffic
Pattern gt Transport network Dynamic Needs NC
Transport Bandwidth Needs Global Control Plane
(MPLS, ATM, .)
C
C
- Complex Network control
- Cannot follow traffic faster than time
- of-flight changes (ms)
- Needs large buffers
- Latency and probability of loss issues
C
C
1
2
3
4
C
C
C
C
- Complexity of control plane and intrinsic slow
response - is a problem for this technique
10Solution 4 Load-Balancing
- Send C/N traffic to every node
- Needs 2NC bandwidth
- Packet routing Less Complex than in IP-router
Network - Transport network static
- will work for any traffic pattern
C
C
C
1
2
C
3
4
C
C
C
C
- New Solution that eliminates packet switching,
- no new control planes needed.
11Load-Balancing Capacity Comparison
Uni-Directional Ring
Mesh Network
C
C
1
2
C
C/N
C
C/N
3
4
C
C/N
C
C
C
12Conclusions Network Architectures
- Load balancing
- scales more favorably in transmission capacity
than point-to-point (by factor 2/N) - scales more favorably in packet switching
requirements than IP (by factor N) - scales more favorably in circuit switching
capacity than point-to-point (by factor 2/N) - TDM network sufficient
- Some other benefits
- No global control plane, topologic maps, etc.
- No node ever sees full information (high degree
of data security) - Easy restoration possible (overhead packets,
similar to FEC)
- Some open questions
- Latency (more time of flight propagation delay,
but single-hop routing) - Required resequencing buffer sizes
13The Next Step
- Load-balancing takes THRU traffic from layer 2/3
into layer 1 (SONET/SDH/MPLS) gt simplified
switches, but still needs O-E-O for all THRU
traffic - Can we take the THRU traffic from layer 1 into
physical layer gt no O-E-O on THRU traffic?
14The SONET/SDH Ring Scalability Problem
- SONET ring
- Good granularity (STS-1) but
- O-E-O, data processing at each node
- Amount of thru-traffic increases withnumber of
nodes - ? limited scalability
- WDM ring
- Connect nodes with static wavelengths
- Good scalability, but
- Granularity too coarse
- N-1 transponders at each node
- ? Expensive solution
15Time Multiplexed WDM (T-WDM) TDM Aggregation
Traffic De- aggregator
Traffic Aggregator
10 Gb/s
16Time Multiplexed WDM (T-WDM) Coloring the Packets
Traffic Aggregator
Filter
10 Gb/s
- T-WDM
- Allows to connect in all-optical domain
sub-wavelength granularity channels - Routes channels with simple, passive optical
filters - Scalability of WDM
- Granularity of SONET
17Solution Time-Multiplexed WDM
- T-WDM SONET ring
- Drop single wavelengthat each node
- Tunable transmitters provide WDM connections
without intermediate O-E-Os - Time-multiplexed packet transmission (TDM)
provides fine bandwidth granularity - Periodic global schedule provisions capacity
between nodes and ensures 100 throughput and
fixed low latencies
18T- WDM Ring
19Performance Comparison for N-Node Ring
- T-WDM has Bandwidth Granularity of SONET
- T-WDM has Scalability and Efficiency of WDM
20Experimental Setup
- OC-12 generator produces scrambled SONET frames
with PRBS 231-1 payload - Unused 10 Gb/s time slots are filled with dummy
packets - All Txs switch packets based on a global,
periodic schedule - De-aggreators reassemble packets with SONET data
back into OC-12 stream - BER measurements on OC-12 outputs
- Span 1 21 dB, 400 ps/nm. Span 2 22 dB,
-100 ps/nm
21Fast Tunable Laser Module
3
2
SG -DBR Laser
Gain Source
Wavelocker
TEC Control
- Tunes between 64 ITU channels in less than 45 ns
- Wavelocker feedback improves frequency accuracy
and prevents long-term drift
Without Locker
With Locker
22Aggregator / De-Aggregator Hardware
10 Gb/sPackets out
10 Gb/sPackets in
10 Gb/s Mux(161) board
10 Gb/s Demux(116) board
AggregatorFPGA
DeaggregatorFPGA
15 x OC-12(622 Mb/s)Data out
15 x OC-12(622 Mb/s)Data in
23Aggregator creates Optical Data Packets
- Aggregator FPGA compresses622 Mb/s chunks by
factor 16 in time and adds dead time and training
pattern - 15 packets in 16 time slots? overhead 1/16 (lt
7)? dead time training pattern equals
1/16 of payload - Currently, the aggregator receives one OC-12
stream and creates 14 dummy packets for every
real data packet - Extension to multiple OC-12 channels to be
programmedin the near future
1.4 ms
64 ns
25 ns
24BER measurements
- BER of OC-12 data versus optical power of 10 Gb/s
packets - small penalty (lt 0.5 dB) between back-to-back and
node-to-node transmission - T-WDM can tolerate 24 dB span loss and 400
ps/nm and -100 ps/nm dispersion
25Scheduling Similarity of T-WDM and PON
Time-multiplexed WDM
Tx
Tx
Tx
Tx
Tx
Tx
Tx
Tx
Tx
Rx
Upstream Passive Optical network
Tx
Tx
Tx
Nx1 Passive Combiner
Tx
Tx
Rx
Tx
Tx
26Summary T-WDM
- T-WDM has bandwidth granularity of SONET
- Grooming in optical domain
- Aggregation/deaggregation similar to ATM
- Single 10Gtransmitter shared over multiple
end-nodes - T-WDM has scalability and cost-efficiency of WDM
- Multiple wavelengths simultaneously in fiber
- Optical Add/drop
- Periodic scheduling assures 100 efficiency and
SONET like QoS - Key technologies
- Fast tunable transmitters
- Burst-mode receivers
- Aggregator/Deaggregator
- Global scheduler
27Benefits of Load-Balancing in Packet Routers
- Removes Central Scheduler
- Only local decisions per port-card
- Much better scalability for scheduling
- Guarantees close to 100 throughput
- Replaces Switch Fabric by Hard-Wired Bandwidth
- Allows to reuse STS1 fabric of a SONET xconnect
- Fabric does not need to be strictly non-blocking
- Fabric complexity of O(N) instead of O(N2)
- 1N protection scheme
- Graceful upgrade
28Request by US Defense Agency A Scalable Optical
Router
- Optical Router Architecture Scalable to gt100Tb/s
- Believe is that current router architecture will
not scale to more than a few Tb/s throughput - Main Bottleneck is electronics
- No Electronics in Data-Path
- Complicated functions like regeneration,
buffering, header recognition must be implemented
in optics - Innovative All-optical Signal Processing and
Buffering Techniques - Currently known techniques do not allow for above
processing functions - Push Density of Monolithic Integration
- DARPA believes that optical technologies must
follow path of electronics to achieve high levels
of functionality
29Bell Labs won 12.5M Research Contract with IRIS
Proposal
- Optical Three Stage Load-Balanced Architecture
- Solves Scheduling problem
- Optimally Designed to Take Advantage of Simple
Buffers - High Scalability Because Does Not Require
Strictly Non-Blocking Stages - Photonic Integrated Circuit (PIC) with more than
100 Semiconductor Optical Amplifier - Regenerative Multi-Wavelength Switch
- Optical Time Buffer
- Multi-Wavelength Optical Header Look-Up
- Total Throughput Scales To 256 Tb/s, 6400 Ports
at 40Gb/s - Exploit Space and Wavelength Dimension
- Wavelength Routing in a 80x80 AWG on 80
Wavelength Channels - Wavelength Switching Below 1 ns
30Wavelength Switching
Buffer
From Input Port
Output
T-Tx
40G Rx
retiming
T-Tx
40G Rx
Sche- duler
T-Tx
40G Rx
T-Tx
40G Rx
Clock
31N2 Scalability of the AWG
wavelength
insensitive
Combiners
10
Arrayed
Waveguide
Output Ports
Input Ports
Grating
32Conclusions
- Scalability of current networks architecture for
data traffic is an issue - SONET/SDH inefficient because not statistical
muxing - Packet switched networks very complex
- WDM networks have too coarse bandwidth
granularity - Load-balancing allows for more scalability
- Static SONET/SDH network can now be used for
statistical muxing across network. - Removes scheduling bottleneck from packet
switching - Allows for optical switching technology
- T-WDM for scaling SONET/SDH
- Maintains bandwidth granularity of SONET/SDH
- Scalability of WDM