This document discusses carrier Ethernet service density and the Xelerated HX family of network processors. It defines service density as the amount of network services simultaneously supported by a packet processing device at wire speed. The two key components that determine service density are service processing and service classification/lookups. The presentation examines Xelerated's evolution in these areas over time and how its new HX330 and HX320 network processors achieve the highest levels of service density and efficiency in the industry. Details are provided on the architecture and capabilities of these 100Gbps HX processors.

1. Service Density for Carrier Ethernet Beyond 2010
Carrier Ethernet Design, Linley Tech Seminar
San Jose January 28, 2010
Thomas Eklund
Founder
VP Business Development
Slide 1 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

2. Agenda
 Service Density
 What It Is
 Can we measure it?
 Xelerated Service Density EvoluNon
 The HX 100 Gbps Network Processor
Slide 2 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

3. Leader in Carrier Ethernet Network Processing
 Global Semiconductor Company
 Stockholm, Beijing, Santa Clara and Tel Aviv
 Programmable Chips for Growing
Carrier Ethernet Market
 Network Processors &
Programmable Ethernet Switches
 Data plane soWware
 Technology Leadership
 Patented Dataflow Architecture
 Wirespeed by Design
 Customer AYracNon
 8 out 10 of top Carrier Ethernet
system vendors
Outstanding Financial
Performance 2009
Slide 3 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

4. More Services – Greater Demands
Video Transport
ProtecNon Switching
MPLS‐TP
Y.1731
VPLS Synchronous Ethernet
Mirroring
Link and Service OAM
PPPoE,
DHCP Proxy Relay PTPv1 & v2
PBB, PBT Session Processing
L3 VPN DPI
L2VPN L2VPN
IPv6 IPv6
MPLS MPLS
VLAN VLAN VLAN
ACL ACL ACL ACL
MAC Learning MAC Learning MAC Learning MAC Learning
IPv4 IPv4 IPv4 IPv4
Ethernet Forwarding Ethernet Forwarding Ethernet Forwarding Ethernet Forwarding
1996 1998 2000 2002 2004 2006 2008 2010 2012
Slide 4 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

5. Service Density DefiniNon
“The amount of network services
simultaneously supported by the packet
processing device at wirespeed operaNon.”
But it is hard to measure:
 Every service requires unique set of processing and
classificaNon/memory lookup resources
 Device architecture and instrucNon set differs
 Subject to implementaNon variaNons and opNmizaNon skills
Slide 5 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

6. The Two Components of Service Density
 Service Processing
# Processor Cores
x Clock frequency
x Opera4ons per clock cycle 1
x Synchroniza4on and VLIW factor 2
x Instruc4on efficiency 3
1. VLIW instrucNons enables
4 or 5 operaNons per
 Service ClassificaNon, 2.
instrucNon
Raw capacity is less
# Lookups per second 4 interesNng than realisNc
scenarios
x Raw Memory Bandwidth 3. Network opNmized
instrucNons, e.g. bit field
x Lookup opera4on factor operaNons, addressing,
x Instruc4on efficiency switching
4. Depend on memory BW,
type and operaNon type
Slide 6 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

7. Service Density and Service Efficiency
Power dissipaNon (Energy Efficiency)
 Keeps cost down (OPEX)
 Higher density systems
 Increasingly important
Service Processing Efficiency = ”MOPS per WaY @Wirespeed”
Slide 7 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

8. Xelerated Service Density EvoluNon
Million Lookups per second
8,000
HX
6,000 100 Gbps NPU
4,000 X11 X11
X11
New GeneraNon d240t
d200d240 InstrucNon
factor
48 Gbps NPU
2,000 X10q X10q
X10q
‐e ‐m ‐w
40 Gbps NPU
25 000 50,000 100,000 150,000 Processing
(RealisNc MOPS)
Slide 8 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

9. The Industry Benchmark for Service Processing Density
Service Efficiency (RealisNc MOPS/W)
5,000 HX330
4,000
3,000
Please do the maths
X11
(This is at least twice as good
2,000
as for any compeNtor)
1,000
25 000 50,000 100,000 200,000
Service Density (RealisNc MOPS)
Slide 9 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

10. The Industry Benchmark for Service ClassificaNon Density
Million Lookups/s
8,000
High number of lookups enable
greater service density and HX320
increased flexibility HX330
6,000
Please do the maths…
4,000
2,000
200 400 600 800
External Memory BW (Gbps)
Slide 10 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

11. HX Under the Hood
Slide 11 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

12. Carrier Ethernet 2.0
Metro, Unified Access and Mobile Backhaul
”a pefect fit for dataflow architecture”
Unified Fiber Access Metro Core
Data Center
Enterprise
Enterprise Backbone Gigabit Data Center
VirtualizaNon
Deep Packet Processing
Switch Access Virtualisation in Datacenter Cloud CompuNng
Router Virtualized Servers
Unified Fiber Access vNIC
GPON, EPON, 10 G PON and AE
MDU FTTH VEPA
Point‐to‐Point
Applicances Metro Ethernet
Triple Play, VoD, IPTV
Business and ResidenNal Services Dataflow = Carrier Ethernet
FTTB
Ethernet Switch/Router
ONU/ONT GPON OLT
Fiber Access EPON
GPON, EPON, 10 G PON and AE Dataflow Card Integration MSPP Edge Router
Line = Carrier Ethernet Core
Triple Play, VoD, IPTV FTTN Router
VDSL
Business and Residential Services
Mobile Backhaul Mobile Backhaul MAC Processor
PHY Packet Service Router
Eth
TM
Switch
3/4G
PTP & Synchronous Ethernet
X10q
PTN, MEF22 WiMAX X11
Video Router
AX/HX
LTE, WiMAX, Microwave Metro Ethernet Packet
Copper‐DSL Transport
AggregaNon Switch
Cable
Slide 12 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

13. Under the Hood: The HX Family of Network Processor
– 100 Gbps with Integrated Traffic Manager
The HX Family
of Metro Ethernet NPUs
HX330 & HX230 HX320 & HX326
Integrated Traffic Manager
Deep‐buffer
Product Processing Performance Interfaces Processors
Traffic Manager
100 Gbps 56xGE, 12x10GE
HX330 Yes 512
150 Mpps 10/20/50G Interlaken
100 Gbps 56xGE, 12x10GE
HX320 ‐ 512
150 Mpps 10/20/50G Interlaken
100 Gbps 56xGE, 12x10GE
HX326 Yes 512
150 Mpps 100G Interlaken
50 Gbps 28xGE, 6x10GE
HX230 ‐ 256
75 Mpps 10/20G Interlaken
Slide 13 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

14. The HX NPU Block Diagram
’Line‐Card‐On‐a‐Chip’
TCAM DDR2+, QDR2+ DDR3 DDR3
NSE i/f SRAM i/f DRAM i/f DRAM i/f
Line Interfaces MACs Programmable Look‐aside Engines
100/1000 Base‐X 10/100/
/GE/2.5GE
2500 Base‐X Hash SRAM TCAM Status Drop
SGMII, QSGMII Engine Engine Engine Engine Engine
1.25‐6.25 Gbaud SerDes
XAUI XG MACs
Interlaken 10G/20G
Programmable Pipeline
EAP
EAP
Interlaken
10/25/50G
System Interfaces Packet
Generator
XAUI Interlaken
XAUI+ 100G
SPAUI Shared Memory Switch
Interlaken
Flow Ctrl
PCIe
CPU i/f Traffic Manager
CPU
DDR
CPU i/f
Slide 14 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

15. Key New Features
 Hiearchical Traffic Manager
 Shared Memory Switch
 Stacking/Meshing
 Precision Time SoluNons
 Increased Service Density
Slide 15 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

16. Wirespeed HX330 TM Features
 Single Scheduler
 Queues freely allocated between ports and ingress/egress
 5 levels Scheduling Hierarchy
 Propagated priority scheduling and deficit shaping
 Hierarchical scheduling nodes flexibly configured
 RR, DWRR and SPQ scheduling at all levels
 Dual Rate Shaping
 Packet Buffer
 Guaranteed BW using 4 or 2 LA Interfaces DDR3
 33M packets
 Dynamic bandwidth sharing between read/write
 WRED and drop precedence watermarks
 Buffer ReservaNon
Slide 16 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

17. TM Service Mapping Example
1 x ’Private’ Subscriber with 4 – 8 Services
SPQ
/WFQ
Shaping Scheduling
Voice HP1 Shaping Backpressure
IPTV HP2
VOD MP2
Best Effort TCP To
LP1 Ports
”Surfing”
Service Subscribers User Classes E.g. Ports
Queues Queues
Provider
Per Service SPQ
CIR/EIR /WFQ
shaping Scheduling
Per User
CIR/EIR
Voice HP1 shaping
IPTV HP2
VOD MP2
Best Effort TCP
LP1
”Surfing”
Flows
VLANs
Service levels Hierarchy Interfaces
Flows
Services
CoS Users
Interfaces
ApplicaNons
Users
User Groups
Tunnels
Interfaces
Slide 17 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

18. Integrated Traffic Manager
No Trade‐off Between Performance and Cost
800
700
600 Xelerated’s TM is equally
cost‐effecNve for small and
500 large packet sizes.
Cost (USD)
400
AlternaNve Designs tend to be
non‐wirespeed, sNll at a higher
300
cost level
DRAM+SRAM TM
200 Xelerated DRAM TM
100
0
8 192 4 096 2 048 1 024 512 256 128 64
Packet size (bytes)
Slide 18 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

19. Pointer Storage in DRAM
Wirespeed Performance for Internet Services
500
450
400
350
Packet buffer
required to
300 support Internet
Packet buffer (ms)
round trip
250 delays
200
150
2 million pointers
100
50 Xelerated
0
8192 4096 2048 1024 512 256 128 64
Packet size (bytes)
Slide 19 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

20. Integrated Shared Memory Switch
High‐bandwidth Switching
 Single Main Packet Buffer
Shared Memory Switch (SMS)
EAP
 Fine‐grained programmable
TI resource reservaNon for e.g.
asymmetric configuraNons
Programmable Pipeline
PP
Input  3‐level hierarchical shaping/scheduling
TM Arbiter
 MulNple watermarks for
Traffic Schedulers
Schedulers
Schedulers
drop presedence
Inter‐
Faces  400 Mpps read/write performance
 Supports over‐speed transit traffic in
stacked configuraNons
Main Packet Buffer
TI (Shared Memory) PP EAP  Logical and SpaNal MulNcast
TM
Traffic Manager
Slide 20 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

21. The HX and Service Density …
“The Proof Is in the Pudding”
Slide 21 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

22. Single Chip Service Density Example:
Xelerated’s Metro Ethernet ApplicaNon
 Shipping since 2003 Simultaneously supported features @ wirespeed
 Robust C API FuncTons MEA Key Features
 Data structures in ANSI C for L2 forwarding L2 switching, IEEE 802.1Q
consistency between control PB IEEE 802.1ad
and data planes Provider Bridging PBB, IEEE 802.1ah
PBB‐TE, IEEE 802.1ay
 Test Suites L3 forwarding IPv4/IPv6
PWE3
MPLS
 DiagnosNc SW MPLS LER/LSR
Tunneling IP‐IP, GRE, MPLS, VPLS
 Aricent CP integrated
L2/L3 mulNcast
MulTcast
IGMP snooping
RestoraTon RSPT, MSTP (IEEE 802.1w/IEEE 802.1s)
Filtering ACL
Sole data plane MAC learning, distributed MAC learning
OAM CFM 802.1ag
DiagnosNcs
Metering Single and dual rate metering (MEF & IETF)
Counter and StaTsTcs MIB2, RMON, SMON
Precision Time SoluTon PTP v1, v2, Synchronous Ethernet
Slide 22 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

23. 100 Gbps Is Just A Good Start
Service Density is What Makes The Difference
 Meet the ever‐increasing demand for processing
 Improved ROI for Service Providers
 Longer LifeNme
 Support real producTon NPU applicaNons
 Not limited to lab test
 Future headroom
 Increased flexability (easier to add new funcNonality in a producNon
code base)
Avoid Being Caught by Service Density LimitaNons
12 months Into the Engineering Project
Slide 23 4 mars 2010 Copyright Xelerated 2010. All rights reserved.

24. Thomas Eklund
thomas.eklund@xelerated.com
Slide 24 4 mars 2010 Copyright Xelerated 2010. All rights reserved.