What’s happening

in the Networking Landscape?An overview on contemporary merchant

switching silicon and SDN landscape

Paolo Bianco

GCN Systems Engineer

paolo.bianco@dell.com Windows Server: Power your business

2

º

25 years ago: Compute paradigm shift

Mainframe model

Proprietary CPUs

(e.g. SPARC, PA-RISC,

Alpha)

Proprietary OS

(e.g. Solaris, HP-UX, Ultrix)

Limited apps

Proprietary

architectures & mgmt

tools

X86 Servers

model

Application ecosystem

Orchestration / automation

for dist computing

Industry standard

(X86 CPU)

Standard OS / hypervisors

3

Traditional networking

Proprietary ASICs

Proprietary OS

(e.g., Cisco IOS, Jun OS)

Hundreds of protocols

Proprietary architectures

& mgmt tools

Open Networking

Networking OS

SDN / NVO controller

Standard orchestration

and automation tools

Merchant silicon

Now: Networking paradigm shift

Contemporary

Merchant Silicon

Landscape

Windows Server: Power your business

5

Traditional Approach: Bottom-Up network design

6

New Approach: Top-Down network design

ProgrammableASIC

7

The Programmable Switching Silicon Landscape

• One big Player: Broadcom

– 95% of market share in 2015

– Broader Ecosystem

– First with fastest Programmable Switch Silicon (12.8TB/s)

• Some Emerging Contenders

– Cavium

› First with 25GbE

› XPA programmable architecture on all DC chipset

– Barefoot

› Innovative approach

› PISAprogrammable architecture

› HW+SW «Weapon System» with P4

– Innovium

› Very New company

› Now Sampling 12.8TB/s Teralynx chipset

› Innoflex programmable architecture

8

Broadcom Ecosystem

9

Broadcom Offering

10

Trident 3 chipset Architecture

11

Cavium - XPliant

12

Broadcom Vs XPliant

13

Innovium Teralynx

14

Domain Specific «Weapon Systems»

GraphicsGraphics

SignalProcessing

Machine Learning Networking

DSP

Signal

Java/C++

Compiler

Computers

CPU DSP

Signal

OpenCL

Compiler

GPU

Signal

Matlab

Compiler

DSP

Signal

TensorFlow

Compiler

TPU

Signal

?

Compiler

?

>>>

?

15

P4.org – P4 Language Consortium

16

Barefoot Tofino

17

PISA: Protocol Independent Switch Architecture

9

Ingress EgressBuffer

Pro

gra

mm

able

Pars

er

18

PISA: Protocol Independent Switch Architecture

Mix of SRAM and TCAM for: lookup tables, counters, meters, Bloom filters

ALUs for: Standard Boolean andArithmetic Operations & add/delete fields, hashes

Recirculation

Programmable Packet Generator

Pro

gra

mm

able

Pars

er

19

Copyright © 2016 - Barefoot Networks

Tofino Block Diagram

Rx MACs10/25/40/50/100

Ingress

PipelineTx MAC

10/25/40/50/100

Control & Configuration

Reset /

ClocksPCIe

CPU

MACDMA

engines

Rx MACs10/25/40/50/100

Ingress

PipelineTx MAC

10/25/40/50/100

Rx MACs10/25/40/50/100

Ingress

PipelineTx MAC

10/25/40/50/100

Rx MACs10/25/40/50/100

Ingress

PipelineTx MAC

10/25/40/50/100

Traffic

Manager

Egress

Pipeline

Egress

Pipeline

Egress

Pipeline

Egress

Pipeline

SDN for Merchant Silicon Landscape

Windows Server: Power your business

21

What is Software Defined Networking?

• A software-defined networking (SDN) architecture defines how a networking and computing system can be built

using a combination of open, software-based technologies and commodity networking hardware that separate

the control layer and the data layer of the networking stack.

• Open Networking is the foundation for SDN (but is not SDN!)

• In the SDN architecture, the splitting of the control and data forwarding functions is referred to as

“disaggregation” because these pieces can be sourced separately, rather than deployed as one integrated system.

In turns: SDN focuses on network stack disaggregation

22

Implications of SDN

Controller (N.O.S.)

ApplicationsApplicationsApplications

SouthboundAPI

Switch O.S

Switch HW

Switch O.S

Switch HW

Switch O.S

Switch HW

Global View

Programmatic

Control

Network O.S.

ASIC

ApplicationsApplications

Network O.S.

ASIC

ApplicationsApplications

Network O.S.

ASIC

ApplicationsApplications

Current Networking SDN Enabled Environment

23

From Network of boxes to Networked fabric

Traditional Approach

“Network of boxes”

Modern Approach

“Networked fabric”

• Deployment

• Configuration prone to error

• Time consuming configuration

• Complex topology validation

• Management

• Box by box visibility

• Device level troubleshooting

Deployed, managed,

optimized, and automated

as a single entity

24

Vendor EcosystemContemporary SDN Ecosystem (not necessarily complete!)

25

Big Switch Networks Portfolio

Networking

Architecture

(Hyperscale

Inspired)

SDN Controller

Software

Open Networking

Hardware

Scale-out Fabric

Architecture

The Next-Generation Data Center Networking CompanyBig Switch

Networks

Products/

SolutionsPervasive

Visibility

Big

Monitoring

Fabric

DMZ

SecurityVMware SDDC(vSphere, NSX, vSAN)

Big

Cloud

Fabric

OpenStack(NFV/Private Cloud)

Containers(Docker, Kub, Mesos, RH)

Cloud

Monitoring

Company

Mission

Next-Generation

DC switchingNext-Generation

DC security and monitoring

26

• Pair of Hardware Controllers

• 64 racks

• Maximum 16 Spine Switches

• 10G spine connections to Leaf

Switches

• 40G breakout to 4 * 10G on Dell

Z9100-ON (due to breakout

capabilities available on all ports)

• 32 * 4 = 128• 128 Leaf Switches

• 2 per rack, 64 racks: 2 * 64 = 128

• Whilst 16 Spine Switches can be used, you

must remember that each Spine Switch

must connect to every Leaf Switch.

Therefore each Leaf Switch would require

16 Spine to Leaf connections

Big CloudFabric

Controllers

RACK 64RACK 63RACK 3RACK 2Service Rack1

Big Cloud P Fabric

….. Up to * 16

Total 10G Leaf Compute/Service connectivity:

48 * 64 = 3072 (MLAG)

Big Cloud Fabric Maximum DeploymentMaximum deployment for a single BCF Pod

27

POD-Level deployment in Existing Data Centeroperate with legacy

Data

Center

Core

Routers

Internet/WA

N

Big

Cloud

Fabric

Controll

er

RACK 1

RACK 2

Servic

e

Rack

Big CloudFabric

Big

Cloud

Fabric

Controll

er

RACK N-1 RACK

N

RACK 1

RACK 2

Servic

e

Rack

Big CloudFabric

Big

Cloud

Fabric

Controll

er

RACK N-1 RACK

N

RACK 1

RACK 2

Servic

e

Rack

Big CloudFabric

DC2

(Remote)

BG

P

BG

PBG

P

VXLAN

VX

LAN

Legacy Network(Box-by-Box)

L3

L2

L2

RACK 1 RACK 2 RACK 3 RACK 44

28

Adaptive

Cloud Fabric

Pluribus Adaptive Cloud Fabric

28

Hosts

Multi-chassis

LAG

Existing

Network

Pluribus HA

Clusters

Access

Switches

Hosts

Existing

Network

Access

Switches

Adaptive

Cloud Fabric

Logical

Equivalent

Underlay

Geographically dispersed sites

Overlay

Overlay

29

Fabric Manager

29

Insight Analytics

Pluribus UNUM™ Unified Management, Automation and Analytics

Virtual appliance

Device & Fabric

Management

Automated

Provisioning

Topology

VisualizationAnalytics & Performance

Management

Thank you

Windows Server: Power your business