EXTREME Lippis AA Report Spring 2013

8/13/2019 EXTREME Lippis AA Report Spring 2013

http://slidepdf.com/reader/full/extreme-lippis-aa-report-spring-2013 1/36

Lippis Open Industry Active-Active Cloud Network Fabric Test

for

Two-Tier Ethernet Network Architecture

April, 2013

© Lippis Enterprises, Inc. 2013

A Report on the:

Extreme Networks Open Fabric




for Two-Tier Ethernet Network Architecture

2 © Lippis Enterprises, Inc. 2013 Evaluation conducted at Ixia’s iSimCity Lab on Ixia test equipment www.lippisreport.co

Acknowledgements

We thank the ollowing people or their help and support in making possible this first industry

evaluation o active-active Ethernet Fabrics:

Victor Alston, CEO o Ixia—or his continuing support o these industry initiatives

throughout its execution since 2009.

Leviton—or the use o fiber optic cables equipped with optical SFP+ connectors to link Ixia

test equipment to 10GbE switches under test.

Siemon—or the use o copper and fiber optic cables equipped with QSFP+ connectors to link

Ixia test equipment to 40GbE switches under test.

Michael Githens, Lab Program Manager at Ixia—or his technical competence, extra effort

and dedication to airness as he executed test week and worked with participating vendors to

answer their many questions.

Henry He, echnical Product Manager at Ixia—or his insights into active-active networking

protocols and multiple test design interactions with the vendor community.

Jim Smith, VP o Marketing at Ixia—or his support and contribution to creating a successulindustry event.

Mike Elias, Photography and Videography at Ixia—or video podcast editing.

All participating vendors and their technical plus marketing teams—or their support o not

only the test event and multiple test configuration file iterations but to each other, as many

provided helping hands on many levels to competitors.

Bill Nicholson or his graphic artist skills that make this report look amazing.

Jeannette ibbetts or her editing that makes this report read as smooth as a test report can.

License Rights

© 2013 Lippis Enterprises, Inc. All rights reserved. Tis report is being sold solely to the purchaser set orth below. Te

report, including the written text, graphics, data, images, illustrations, marks, logos, sound or video clips, photographs

and/or other works (singly or collectively, the “Content”), may be used only by such purchaser or inormational purpos-

es and such purchaser may not (and may not authorize a third party to) copy, transmit, reproduce, cite, publicly display,host, post, perorm, distribute, alter, transmit or create derivative works o any Content or any portion o or excerpts

rom the Content in any ashion (either externally or internally to other individuals within a corporate structure) unless

specifically authorized in writing by Lippis Enterprises. Te purchaser agrees to maintain all copyright, trademark and

other notices on the Content. Te Report and all o the Content are protected by U.S. and/or international copyright laws

and conventions, and belong to Lippis Enterprises, its licensors or third parties. No right, title or interest in any Content

is transerred to the purchaser.

Purchaser Name:

Note that, currently Ixia’s statistics do not support the combination o Multicast traffic running over

ports within a LAG. Tereore, packet loss or this scenario was not accurately calculated and is

thereore not valid.

mailto:Bill%20Nicholson%20%[email protected]%3e

mailto:Jtypeplus%20%[email protected]%3e

mailto:Jtypeplus%20%[email protected]%3e

mailto:Bill%20Nicholson%20%[email protected]%3e






Table of Contents

Executive Summary .............................................................................................................................................4

Market Background .............................................................................................................................................7

Active-Active Fabric est Methodology ............................................................................................................8

Ethernet Fabrics Evaluated ................................................................................................................................12

Vendor ested:

Extreme Networks Open Fabric ........................................................................................................12

Active-Active Vendor Cross-Vendor Analysis ...............................................................................................23

Cross Vendor Analysis

Fabric est .............................................................................................................................................24

Server-oR Reliability est .................................................................................................................30

Cloud Simulation est .........................................................................................................................31

Ethernet Fabric Industry Recommendations .................................................................................................34

erms o Use ........................................................................................................................................................35

About Nick Lippis ..............................................................................................................................................36

© Lippis Enterprises, Inc. 2013 Evaluation conducted at Ixia’s iSimCity Santa Clara Lab on Ixia test equipment www.lippisreport.co






Executive Summary

o assist I business leaders with the design and procurement o their private or public data center

cloud abrics, the Lippis Report and Ixia have conducted an open industry evaluation o Active-Active

Ethernet Fabrics consisting o 10GbE (Gigabit Ethernet) and 40GbE data center switches. In thisreport, I architects are provided the first comparative Ethernet Fabric perormance and reliability

inormation to assist them in purchase decisions and product differentiation.

Te Lippis test report based on independent validation at Ixia’s iSimCity laboratory communicates

credibility, competence, openness and trust to potential buyers o Ethernet Fabrics based upon

active-active protocols, such as RILL, or ransparent Interconnection o Lots o Links, and SPBM,

or Shortest Path Bridging MAC mode, and configured with 10GbE and 40GbE data center switching

equipment. Most suppliers utilized MLAG, or Multi-System Link Aggregation, or some version o it,

to create a two-tier abric without an IS-IS (Intermediate System to Intermediate System) protocol be-

tween switches. Te Lippis/Ixia tests are open to all suppliers and are air, thanks to well-vetted customEthernet Fabric tests scripts that are repeatable. Te Lippis/Ixia Active-Active Ethernet Fabric est was

ree or vendors to participate and open to all industry suppliers o 10GbE, 40GbE and 100GbE switch-

ing equipment, both modular and fixed configurations.

Tis report communicates test results that took place during the late autumn and early winter o

2012/2013 in the modern Ixia test lab, iSimCity, located in Santa Clara, CA. Ixia supplied all test equip

ment needed to conduct the tests while Leviton provided optical SPF+ connectors and optical cabling.

Siemon provided copper and optical cables equipped with QSFP+ connectors or 40GbE connections.

Each Ethernet Fabric supplier was allocated lab time to run the test with the assistance o an Ixia engi-

neer. Each switch vendor configured its equipment while Ixia engineers ran the test and logged result-

ing data.

Te tests conducted were an industry first set o Ethernet Fabric est scripts that were vetted over a six-

month period with participating vendors, Ixia and Lippis Enterprises. We call this test suite the Lippis

Fabric Benchmark, and it consisted o a single- and dual-homed abric configuration which three tra-

fic types, including multicast, many-to-many or unicast mesh and unicast rom multicast returns. Te

Lippis Fabric Benchmark test suite measured Ethernet Fabric latency in both packet size iterations and

Lippis Cloud Simulation modes. Reliability or packet loss and packet loss duration was also measured

at various points in the Fabric. Te new Lippis Cloud Simulation measured latency o the abric as tra-

fic load increased rom 50% to 100% consisting o north-to-south plus east-to-west traffic flows.






Ethernet Fabrics evaluated were:

Arista Sofware-Defined Cloud Network (SDCN) consisting o its 7050S-64 10/40G

Data Center Switch oR with 7508 Core Switches

Avaya Virtual Enterprise Network Architecture or VENA Fabric Connect consisting o its

VSP 7000

Brocade Virtual Cluster Switching or VCS consisting o its VDXM 6720 oR and

VDXM 8770 Core Switch

Extreme Networks Open Fabric consisting o its Summit® X670V oR and

BlackDiamond X8 Core Switch






The following lists our top ten findings:

1. New Fabric Latency Metrics: Te industry is amiliar

with switch latency metrics measured via cut-through

or store and orward. Te industry is not amiliarwith abric latency. It’s anticipated that abric latency

metrics reported here will take some time or the

industry to digest. Te more abrics that are tested, the

greater utility o this metric.

2. Consistent Fabric Perormance: We ound, as

expected, that abric latency or each vendor was

consistent, meaning that as packet sizes and loads

increased, so did required processing and thus latency.

Also, we ound that non-blocking and ully mesh

configurations offered zero abric packet loss providing

consistency o operations.

3. No Dual-Homing Perormance Penalty: In ully

meshed, non-blocking abric designs, we ound no

material abric latency difference as servers were

dual homed to different op-o-Racks (oRs). Fabric

latency measurement in dual homed were the same as

single-homed configuration (as expected) even though

increased reliability or availability was introduced to

the design via dual-homing server ports to two oRs

plus adding MLAGs between oRs. Tis was true in

both Arista and Extreme’s test results.

4. CLI-Less Provisioning: Brocade’s VCS, which is

RILL based, offered several unique attributes, such as

adding a switch to its Ethernet Fabric plus bandwidth

between switches without CLI (Command Line

Interace) provisioning. Fabric configuration was

surprisingly simple, thanks to its ISL, or Inter-Switch

Link runking.

5. Different Fabric Approaches: In this Lippis/

Ixia Active-Active Ethernet Fabric est, different

approaches to abrics were encountered. Avaya tested

its Distributed oR, or doR, as part o its Fabric

Connect offering, which stacks o oRs horizontally

and can offer advantages or smaller data centers with

dominate east-west flows. Extreme Network’s Open

Fabric utilizes its high perormance and port dense

BlackDiamond X8 switch, which enables a abric to be

built with just a ew devices.

6. ECMP n-way Scales: MLAG at Layer 2 and Equal

Cost Multi-Path, or ECMP, at Layer 3 are dominant

approaches to increasing bandwidth between switches.Arista demonstrated that an ECMP-based abric scale

to 32 with bandwidth consistency among links that is

bandwidth is evenly distributed between the 32 10GbE

7. Balanced Hashing: We ound that vendors utilized

slightly different hashing algorithms, yet we ound

no difference in hash results. Tat is, we ound evenly

distributed traffic load between links within a LAG

(Link Aggregation) during different traffic load

scenarios.

8. vCenter Integration: All abric vendors offer vCenterintegration so that virtualization and network

operations teams can view each other’s administrative

domains to address vMotion within L2 confines.

9. MLAG the Path to Two Tier Data Center Networks:

MLAG provides a path to both two-tier data center

networking plus RILL and/or SPB (Shortest Path

Bridging) in the uture. MLAG takes traditional link

aggregation and extends it by allowing one device to

essentially dual home into two different devices, thus

adding limited multipath capability to traditional LAG

10. More TRILL and SPB: It’s disappointing that

there are not more vendors prepared to publicly

test RILL and SPB implementations as Avaya and

Brocade demonstrated their ease o deployment and

multipathing value in this Lippis/Ixia Active-Active

Ethernet Fabric est.






Market Background

Data center network design has been undergoing rapid

changes in a ew short years afer VMware launched VM

(Virtual Machine) Virtual Center in 2003. Server virtualiza-tion enabled not only efficiency o compute, but a new I

delivery model through private and public cloud computing

plus new business models to emerge. Fundamental to mod-

ern data center networking is that traffic patterns have shif-

ed rom once dominant north-south or client-to-server to

now a combination o north-south plus east-west or server-

server and server-storage. In many public and private cloud

acilities, east-west traffic dominates flows. Tere are many

drivers contributing to this change, in addition to server

virtualization, such as increased compute density scale,

hyperlinked servers, mobile computing, cloud economics,etc. Tis simple act o traffic pattern change has given rise

to the need or ew network switches or tiers, lower latency,

higher perormance, higher reliability and lower power

consumption in the design o data center networks.

In addition to traffic shifs and changes, service providers

and enterprise I organizations have been under increasing

pressure to reduce network operational cost and enable sel-

service so that customers and business units may provision

I needed. At the February 13, 2013, Open Networking

User Group in Boston at Fidelity Investments, hosted by the

Lippis Report, Fidelity showed the beginning o exponential

growth in VM Creation/Deletion by business unit manag-

ers since August 2012. Reduced OpEx and sel-service are

driving a undamental need or networking to be included

in application, VM, storage, compute and workload auto

provisioning.

o address these industry realities, various non-profit oun-

dations have ormed, including the Open Compute Project

Foundation, Te OpenStack Foundation, Te Open Net-

working Foundation, etc. Tese oundations seek to open

up I markets to lower acquisition cost or CapEx and inject

innovation, especially around auto provisioning to lower

OpEx. While the oundations are developing open source

sofware and standards, the vendor community has been in-

novating through traditional mechanism, including product

development and standards organizations, such as the IEF,

IEEE and others.

Data center networks have been ramping up to build private

and public cloud inrastructure with 10/40GbE and soon

100GbE data center switches with 400GbE on the horizon.At the center o next generation data center/cloud network-

ing design are active-active protocols to increase application

perormance, thanks to its lower latency and increased reli-

ability via dual-homed, ull-meshed non-blocking network

abric plus CLI-less bandwidth provisioning.

o deliver a two-tier network, spanning tree protocol (SP)

is eliminated and replaced with active-active multipath links

between servers and oR switches, oR Core switches and

between, Core switches. Te industry is offering multiple

active-active protocol options, such as Brocade’s VCS Fab-ric, Cisco’s FabricPath, Juniper’s QFabric, RILL, SPBM and

LAG Protocol. MLAG and ECMP are design approaches to

limited active-active multipathing; they are widely used and

central to many vendors’ SP alternative strategies, but they

lack CLI-less provisioning.

Ethernet abrics are promoted to be the optimal platorm to

address a range o data center design requirements, includ-

ing converged storage/network, network virtualization,

Open Networking, including Sofware-Defined Networking

or SDN, and simply the way to keep up with ever-increasingapplication and traffic load.

In this industry first Lippis/Ixia Active-Active Ethernet

Fabric est , we provide I architects with comparative

active-active protocol perormance inormation to assist

them in purchase decisions and product differentiation.

New data center Ethernet abric design requires automated

configuration to support VM moves across L3 boundaries,

low latency, high perormance and resiliency under north-

south plus east-west flows, and minimum number o

network tiers.

Te goal o the evaluation is to provide the industry with

comparative perormance and reliability test data across all

active-active protocols. Both modular switching (Core plus

End-o-Row, or EoR) products and fixed oR configuration

switches were welcome.






Lippis Active-Active Fabric Test Methodology

Tere are two active-active test configurations—single and

dual homed—used in the Lippis/Ixia Active-Active Ethernet

Fabric est. Tese configurations consisted o two or ouroRs and two Core switches to test Ethernet abric latency,

throughput and reliability. For those vendors that offer

RILL, SPBM or FabricPath Core switches were not needed

as Ixia’s IxNetwork simulated these active-active protocols

where latency, throughput and reliability can be measured.

Most companies ran both simulated and non-simulated test

runs or IS-IS based active-active and MLAG, respectively.

Te single- and dual-homed configurations and traffic pro-

files are detailed below. Tese configurations were used or

those vendors wishing to test MLAG and/or RILL, SPBMand FabricPath without the use o Ixia’s simulation.

Single-Homed Configuration: In the single-homed con-

figuration, two oR and two Core switches made up an

Ethernet abric. Tirty-two-10GbE links connected Ixia test

equipment to two oR, which are divided into eight, our-

port LAGs. Each oR connected to two Core switches with

16-10GbE or our-40GbE links. Tereore, the load placed

on this Ethernet abric was 32-10GbE ports, or 320Gbs.

A mix o unicast, multicast and mesh or any-to-any flows

to represent the Brownian motion typical in modern data

center networks was placed upon this abric while latencyand throughput were measured rom ingress to egress; that

is, rom oR-to-Core-oR, representing abric latency and

throughput.

Dual-Homed Configuration: In the dual-homed configu-

ration, our oRs and two Core switches created the Ether-

net abric. Each Ixia port, acting as a virtual server, was dual

homed to separate oR switches, which is a best practice in

high availability data centers and cloud computing acili-

ties. Te load placed on this Ethernet abric was the same

32 10GbE, or 320Gbs, as in the single-homed configuration,

with a mix o unicast, multicast and mesh or any-to-any

flows placed upon the abric. Each oR was configured

with eight-10GbE server ports. Eight-10GbE or two-40GbE

lagged ports connect oRs. Finally the oRs were con-

nected to Core switches via eight-10GbE or dual-40GbE

port MLAGs. Latency and throughput were measured rom

ingress to egress; that is, rom oR-to-Core-oR, represent-

ing abric latency and throughput rather than device.

L1 L2 L5 L6 L3 L4 L7 L8

10G

Ixia Port LAG AssignmentIxia Port LAG Assignment

40G

4x40G

Core Core

ToR ToR

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32

Single-Homed Topology

XXXX

ToR ToR

40G

4x40G

Core Core

1 4 17 31 3 10 22 32

Dual-Homed Topology

ToR ToR






Test Traffic Profiles: Te logical network was an iMix o

unicast traffic in a many-to-many or mesh configuration,

multicast traffic and unicast return or multicast peers

where LAGs are segmented into traffic types. LAGs 1, 2,

3 and 4 were used or unicast traffic. LAGs 5 and 6 weremulticast sources distributing to multicast groups in LAGs

7 and 8. LAGs 7 and 8 were unicast returns or multicast

peers within LAGs 5 and 6.

Lippis Cloud Performance Test

In addition to testing the abric with unicast, multicast andmany-to-many traffic flows at varying packet sizes, the

Lippis Cloud Perormance est iMix was used to generate

traffic and measure system latency and throughput rom

ingress to egress. o understand the perormance o the

Ethernet abric under load, six iterations o the Lippis Cloud

Perormance est at traffic loads o 50%, 60%, 70%, 80%,

90% and 100% were perormed, measuring latency and

throughput on the oR switch. Te oR was connected to

Ixia test gear via 28-10GbE links.

Te Lippis Cloud Perormance est iMix consisted o east-

west database, iSCSI (Internet Small Computer System

Interace) and Microsof Exchange traffic, plus north-south

HP (Hypertext ranser Protocol) and Youube traffic.

Each traffic type is explained below:

East-West database traffic was set up as a request/response.

A single 64-byte request was sent out and three different-

sized responses were returned (64, 1518 and 9216 bytes).

A total o eight ports were used or east-west traffic. Four

ports were set as east and our ports were set as west. Tese

eight ports were not used in any other part o the test. Te

transmit rate was a total 70% o line rate in each direction.

Te response traffic was urther broken down with weightso 1/2/1 or 64/1518/9216 byte rames or the three response

sizes. Tat is, the weight specifies what proportion rom the

rate set per direction will be applied to the corresponding

x ports rom the traffic profile.

East-West iSCSI traffic was set up as a request/response

with our east and west ports used in each direction. Each

direction was sending at 70% o line rate. Te request was

64 bytes and the response was 9216 bytes.

East-West Microsof Exchange traffic was set up on two

east and two west ports. Te request and response were

both 1518 and set at 70% o line rate.

Te ollowing summarizes the east-west flows:

Database: 4 East (requestors) to 4 West (responders)

iSCSI: 4 East (requestors) to 4 West (responders)

MS Exchange: 2 East (requestors) to 2 West

(responders)

Database/iSCSI/MS Exchange Weights: 1/2/1, i.e.,

25%/50%/25% o rate set per direction and applicable

on selected ports. East rate: 70% = West rate: 70%.

North-South HTTP traffic was set up on our north and

our south ports. Te request was 83 bytes and the response

was 305 bytes. Te line rate on these ports was 46.667% line

rate in each direction.

North-South YouTube traffic was using the same our

north and south ports as the HP traffic. Te request was500 bytes at line rate o 23.333%. Tere were three responses

totaling 23.333% in a 5/2/1 percentage breakdown o 1518,

512 and 64 bytes.

Traffic Profiles

Unicast(Many to Many)

Switch 1 Switch 2

1

2

5

6

3

4

7

8

Multicast

Switch 1 Switch 2

1

2

5

6

3

4

7

8

Unicast(For Multicast Peers)

Switch 1 Switch 2

1

2

5

6

3

4

7

8






Reliability

Fabric reliability was tested in three critical areas: 1) be-

tween Ixia test gear and the oR, 2) between oR and core

and 3) with the loss o an entire core. Te system under test(SU) was configured in the “single-homed” test design.

Only the Ixia-to-oR reliability test was required, all other

reliability tests were optional.

Server to ToR Reliability Test: A stream o unicast many-

to-many flows at 128-byte size packets was sent to the

network abric. While the abric was processing this load,

a 10GbE link was disconnected in LAGs 3, 4, 7 and 8 with

packet loss and packet loss duration being measured and

reported. Note that packet loss duration can vary as the link

ailure detection is based on a polled cycle. Repeated tests

may show results in the nanosecond range or slightly higher

numbers in the millisecond range. Te poll interval or link

detection is not configurable.

ToR to Core Reliability Test: Tere were two parts to this

Lippis/Ixia Reliability est. First, a link that connected oR

switches to Core switches while unicast many-to-many tra-

fic flows were being processed was pulled with the resulting

packet loss plus packet loss duration recorded by Ixia test

equipment. Ten the link was restored, and the resulting

packet loss plus packet loss duration was recorded by Ixiatest equipment. A stream o unicast many-to-many flows at

128-byte size packets was sent to the Ethernet abric. While

the abric was processing this load, a link was disconnected,

and packet loss plus packet loss duration was measured.

When the link was restored, the abric reconfigured itsel

while packet loss and packet loss duration was measured.

Again note that packet loss duration can vary as the link

ailure detection is based on a polled cycle. Repeated tests

may show results in the nanosecond range or slightly higher

numbers in the millisecond range. Te poll interval or linkdetection is not configurable.

Core Switch Shut Down Reliability Test: As above, packet

loss and packet loss duration was measured when the abric

was orced to re-configure due to a link being shut down

and restored while a 128-byte size packet o many-to-many

unicast traffic flowed through the abric. Tis Reliability

est measured the result o an entire Core switch being shut

down and restored. Again note that packet loss duration can

vary as the link ailure detection is based on a polled cycle.

Repeated tests may show results in the nanosecond range or

slightly higher numbers in the millisecond range. Te poll

interval or link detection is not configurable.

Active-Active Simulation Mode Test

For those wishing to test their abric with RILL, SPBM and

FabricPath, the Lippis/Ixia Active-Active est offered a sim-

ulated core option. Te simulated core option eliminated

the need or Core switches in the configuration, requiring

only oRs or the single- and dual-homed configurations.

IxNetwork IS-IS simulated a ully meshed, non-blocking

core. wo and then our oR switches connected to the sim-

ulated core. Note that an equal number o server and core

acing links were required to achieve non-blocking. oR

switches were connected in an “n-way” active-active config-

uration between oR and Ixia test equipment with Ixia gear

configured or the specific abric protocol. N is a maximum

o 32. Te DU was connected to Ixia test equipment with

L1 L2 L5 L6 L3 L4 L7 L8

10G

Ixia Port LAG AssignmentIxia Port LAG Assignment

40G

4x40G

Core Core

ToR ToR

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32


XXXX






enough ports to drive traffic equal to the “n-way” active-

active links. Te active-active links were connected to Ixia

test equipment in core simulation mode where throughput,

packet loss and latency or unicast rom multicast returns,

multicast and unicast many-to-many traffic were measured.

Te same LAG and traffic profiles as detailed above wereapplied to the simulation mode configuration while latency

and throughput were measured or RILL, or SPBM and/or

FabricPath.

Optional Fabric Tests

Te ollowing were optional tests, which were designed to

demonstrate areas where abric engineering investment and

differentiation has been made. Te acquisition o the 2013

Active-Active Fabric est report distribution license was

required to participate in these optional tests; please contact

[email protected] to request a copy.

Most o these optional tests, i not all, were short 10-minute

video podcast demonstrations as they were ocused upon

optional cost reduction via either ease o use or automation

o network configuration.

VM Fabric Join/Remove and Live Migration

Demonstration: One o the most pressing issues or I

operations is or an Ethernet abric to support VMs. By

support, we mean the level o difficulty o a VM to join or

be removed rom the abric. In addition to live VM joins

and removes, the ability or the abric to support the live

migration o VMs across L3 boundaries without the needor network re-configuration is a undamental requirement.

Tereore, the objective or this test was to observe and

measure how the network abric responded during VM

join/remove plus live migration. In this optional test, the

vendor demonstrated the process in which a VM joins and

is removed rom the abric. In addition, a VM was migrated

live while we observed needed, i any, network abric

configuration changes. Tis demonstration was captured

on video and edited into a (maximum) 10-minute videopodcast. A link to the video podcast is included in the

final test report. Vendors may use SPB, RILL, FabricPath,

VXLAN (Virtual Extensible Local Area Network) over

ECMP, etc. wo servers, each with 30 VMs, were available

or this test. Vendors were responsible or NIC (Network

Interace Controller) cards plus other equipment necessary

to perorm this test.

East-West Traffic Flow Perormance Test: Networking

oR switches so that east-west traffic flow does not need to

traverse a Core switch is being proposed by various vendors

as part o their network abric. As such, an optional test ran

RFC 2544 across three interconnected oR switches with

bidirectional L2/L3 traffic ingress at oR switch 1 and egress

at oR 3. Troughput, latency and packet loss was measured

with horizontal oR latency compared to traditional oR-

Core-oR.

Video feature: Click to view a discussion on the

Lippis Report Test Methodology

Virtual Entities withinIxia’s IxNetworkSimulated Fully Meshed,

Non-Blocking Core

Core

DUT

Core

CoreCore

Edge Edge

Host/servers

Host/servers

DUT

Active-Active Links

Active-Active Links

mailto:nick%40lippis.com?subject=Active-Active%20Fabric%20Test%20Report

http://www.lippisreport.com/?p=4101

mailto:nick%40lippis.com?subject=Active-Active%20Fabric%20Test%20Report







Extreme Networks X670V and BD-X8 Active-Active Test Configuration

Hardware Configuration/Ports Test Scripts Software Version

Devices undertest

Summit-X670V-48x with VIM4-40G4Xhttp://extremenetworks.com/products/summit-x670.aspx

EXOS 15.3

BD-X8 with 1 BDXA-40G24X, 2 BDX-MM1, 3 BDXA-FM20T

http://extremenetworks.com/products/blackdiamond-x.aspxEXOS 15.3

Single Homed 2x X670V ToR 32 10GbE/16 Each ToR + Each ToR 4-40GbE to Core EXOS 15.3

2x BD X8 Cores 8-40GbE connect ToRs EXOS 15.3

4-40GbE MLAG between Cores EXOS 15.3

Dual Homed 4x X670V ToR 32 10GbE/8 Each ToR + Each ToR 2-40GbE to Core EXOS 15.3

2x BD X8 Core 2-40GbE MLAG between ToRs in each compute zone EXOS 15.3

4-40GbE MLAG between Cores EXOS 15.3

Test Equipment Ixia XG12 High Performance Chassis Single Home Active-Active IxOS 6.30 EA SP2

Dual Home Active-Active IxNetwork 6.10 EA

Cloud Performance Test IxNetwork 7.0 EA &

IxOS 6.40 EA

Ixia Line Cards Xcellon Flex AP10G16S (16 port 10G module)

Xcellon Flex Combo 10/40GE AP (16 port 10G / 4 port 40G)

Xcellon Flex 4x40GEQSFP+ (4 port 40G)

www.ixiacom.com

Cabling 10GbE Optical SFP+ connectors. Laser optimized duplex lc-lc 50 micron mm fiber, 850nm SPF+ transceivers

www.leviton.com

Siemon QSFP+ Passive Copper Cable 40 GbE 3 meter copper QSFP-FA-010

http://www.siemon.com/sis/store/cca_qsfp+passive-copper-assemblies.asp

Siemon Moray Low Power Active Optical Cable Assemblies Single Mode, QSFP+ 40GbE optical cable QSFP30-03 7 meters


For the Lippis/Ixia Active-Active Fabric est, Extreme

Networks built a ault tolerant, two-tier Ethernet data

center Fabric with its Open Fabric solution, consisting o

10/40GbE switches, including the Summit® X670V-48X oR

switches and high-perormance 20 terabit capacity Black-

Diamond® X8 10/40GbE Core switches.

Te combination o these two products in the design o a

data center Fabric is unique on multiple levels. First, the

BlackDiamond® X8 Core switch is the astest, highest capac-

ity, lowest latency Core switch Lippis/Ixia has tested to date

with the highest density, wire-speed support o 10GbE or

40GbE ports and just 5W per 10GbE port o power.

As a point o differentiation, the high-port density o the

BlackDiamond® X8 Core switch enables a data center Fabric

to be built with just a ew devices, and as its latency is in

the order o a microsecond, this design makes it the highest

perormer in the field.

Video feature: Click to view Lippis/Mendis

Extreme Open Fabric Video Podcast

http://extremenetworks.com/products/summit-x670.aspx

http://extremenetworks.com/products/blackdiamond-x.aspx

http://www.ixiacom.com/

http://www.leviton.com/



http://lippisreport.com/?p=7888



http://www.leviton.com/

http://www.ixiacom.com/

http://extremenetworks.com/products/blackdiamond-x.aspx

http://extremenetworks.com/products/summit-x670.aspx






Te combination o the Summit® X670V oR and BlackDia-

mond® X8 Core orms the basis o Extreme’s Open Fabric

architecture utilized in virtualized multi-tenant data cen-

ters, Internet exchanges, data centers with SDN, converged

I/O data centers and cloud network designs.

Te basic configuration tested consisted o Summit® X670V-

48X oR switches connecting Ixia test gear as server acing

devices and BlackDiamond® X8 switches in the Core. Te

Summit® X670V-48X oR connected to Ixia test gear at

16-10GbE links via our-10GbE LAGs. Each Summit®

X670V-48X oR switch connected to BlackDiamond® X8s

via 40GbE in a ull mesh configuration via its VIM4-40G4X

module. Te BlackDiamond® X8s were connected via our-

40GbE trunks leveraging MLAG.

Te logical network was a mix o unicast traffic in a many-

to-many or mesh configuration, multicast traffic and unicast

return or multicast peers where the LAGs were segmented

into traffic types. For example, LAGs 1, 2, 3 and 4 were used

or unicast traffic. LAGs 5 and 6 were multicast sources

distributing to multicast groups in LAGs 7 and 8. LAGs 7

and 8 were unicast returns or multicast peers within LAGs

5 and 6.

Traffic Profiles

Unicast(Many to Many)

Switch 1 Switch 2

1

2

5

6

3

4

7

8

Multicast

Switch 1 Switch 2

1

2

5

6

3

4

7

8

Unicast Return(For Multicast Peers)

Switch 1 Switch 2

1

2

5

6

3

4

7

8

We tested this data center Fabric in both single-homed and

dual-homed configurations and measured its overall system

latency and throughput. We also tested or reliability, which

is paramount, since Extreme Networks data center Fabric

architecture gives the option to place much o the process-

ing on a ew high-density BlackDiamond® X8 Core switches

instead o dozens o switches. Another configuration uti-

lized in high-density server data centers or cloud comput-

ing acilities taking advantage o the BlackDiamond® X8’s

port density is to connect servers directly into a network oBlackDiamond® X8 Core switches.

Extreme Networks demonstrated its Fabric with MLAG as

the active-active protocol and thus, eliminated the slower,

more rudimentary SP and created a highly efficient

two-tier network design. In addition to current support o

MLAG or L2 and L3, as well as ECMP or L3 based active-

active topologies, Extreme Networks also has committed to

support RILL and other active-active protocols within its

ExtremeXOS OS in uture releases.

Single Homed

For the Single-Homed Server Lippis/Ixia est, Extreme con-

figured two Summit® X670Vs and two BlackDiamond® X8s

or its Ethernet Fabric.

Tirty-two-10GbE links connected Ixia test equipment

to two Summit® X670Vs, which were divided into eight,

our-port LAGs. Each Summit® X670V connected to two

BlackDiamond® X8s with our-40GbE links. Tereore, the

load placed on this Ethernet Fabric was 32-10GbE ports, or

320Gbs, with a mix o unicast, multicast and mesh or any-

to-any flows to represent the Brownian motion typical in

modern data center networks.

Latency and throughput were measured rom ingress to

egress rom Summit® X670Vs to BlackDiamond® X8s to Sum-

mit® X670Vs, representing Fabric latency and throughput.






L1 L2 L5 L6 L3 L4 L7 L8

10G

40G

4x40G BD X8

Summit X670

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32


XXXX

Unicast Results

Extreme Networks’ Open Fabric system latency or unicast

traffic varied rom a low o 2.2 microseconds to a high o 28

microseconds. As expected, latency increased with packetsize where 9216 size packets experienced the largest system

delay. Zero packet loss was observed across all packet sizes o

unicast traffic.

0

5000

10000

15000

0

10000

20000

30000

0

2000

4000

6000

8000

10000

128 256 512 1522 9216

Max Latency 2900 3320 4220 7780 28240

Avg Latency 2445 2591 2923 4366 14117

Min Latency 2280 2360 2580 3460 9620

ns

Avg Latency

Max Latency

Min Latency

Extreme 2-X670V ToR & 2-BD X8 Core

Single Homed Test

Unicast Return From Multicast Flows(min, max, avg latency)






Multicast Results

Te Extreme Open Fabric latency measured or multicast

traffic varied rom a low o 2.4 microseconds to a high o 38

microseconds. As expected, latency increased with packetsize where 9216 size packets experienced the largest system

delay. Note that currently Ixia’s statistics do not support the

combination o multicast traffic running over ports within a

LAG. Tereore, packet loss or this scenario was not accu-

rately calculated and is, thereore, not valid.

0

5000

10000

15000

20000

25000

0

10000

20000

30000

40000

0

2000

4000

6000

8000

10000

128 256 512 1522 9216

Max Latency 3360 4140 5500 11340 38180

Avg Latency 2691 3053 3517 5700 22382

Min Latency 2400 2480 2700 3580 9740

ns

Avg Latency

Max Latency

Min Latency


Single Homed Test

Multicast Traffic

(min, max, avg latency)

Many-to-Many Results

Te Extreme Open Fabric system latency or many-to-many

unicast traffic in a mesh configuration varied rom a high o

45 microseconds to a low o 830 nanoseconds. As expected,latency increased with packet size where 9216 size packets

experienced the largest system delay. Zero packet loss was

observed across all packet sizes o unicast traffic.

0

5000

10000

15000

20000

0

10000

20000

30000

40000

50000

0

400

800

1200

1600

128 256 512 1522 9216

Max Latency 3220 3820 5140 10460 45420Avg Latency 2065 2297 2887 4789 19115

Min Latency 820 860 1000 1240 1240

ns

Avg Latency

Max Latency

Min Latency


Single Homed Test

Many to Many Full Mesh Flows

(min, max, avg latency)






Te table below illustrates the average system latency across

packet sizes 128 to 9216 or unicast, multicast and many-to-

many traffic flows through the Extreme Networks single-

homed configuration. All traffic flows perormed in a tight

range between 2 and 5.8 microseconds except or the large

packet size o 9216 where latency increased by an approxi-

mate actor o five. Tis is mostly due to the act that the

9216 packet size was six times larger than the previous 1522

size and required substantially more time to pass through

the Fabric.

0

5000

10000

15000

20000

25000

0

5000

10000

15000

0

5000

10000

15000

20000

128 256 512 1522 9216

Unicast 2445 2591 2923 4366 14117

Multicast 2691 3053 3517 5700 22382

Many-to-Many 2065 2297 2887 4789 19115

Unicast Return From

Multicast Flows Avg Latency

Multicast Traffic Avg Latency

Many-to-Many

Full Mesh Flows Avg Latency

ns

Extreme 2-X670V ToR 2-BD X8 Core

Single Homed Test

Unicast Return From Multicast Flows,

Multicast Traffic, Many-to-Many FullMesh Flows (avg latency)

Dual Homed

For the Dual-Homed Server Lippis/Ixia est, Extreme

configured our Summit® X670Vs and two BlackDiamond®

X8s or its Ethernet Fabric. Each Ixia port, acting as a vir-tual server, was dual homed to separate Summit® X670Vs,

which is a best practice in high availability data centers and

cloud computing acilities. Te load placed on this Ether-

net Fabric was the same 32 10GbE, or 320Gbs, with a mix

o unicast, multicast and mesh or any-to-any flows. Each

Summit® X670V was configured with eight-10GbE server

ports. Ten two-40GbE lagged ports connected the X670Vs

Finally the top X670Vs were connected to BlackDiamond®

X8 Core switches via dual-40GbE port MLAGs. Latency

and throughput were measured rom ingress to egressrom Summit® X670Vs to BlackDiamond® X8s to Summit®

X670Vs, representing Fabric latency and throughput rather

than device.

40G

4x40G

1 4 17 31 3 10 22 32

Dual-Homed Topology

BD X8

Summit X670






Unicast Results

Te Extreme Open Fabric system latency or unicast tra-

fic varied rom a low o 2.2 microseconds to a high o 28

microseconds. Te result was the same as the single-homed

configuration (as expected) even though increased reli-

ability or availability was introduced to the design via

dual-homing server ports to two Summit® X670V oRs plus

MLAG between oRs. In line with expectations, latency

increased with packet size where 9216 size packets experi-

enced the largest system delay. Zero packet loss was ob-

served across all packet sizes o unicast traffic.

0

5000

10000

15000

0

10000

20000

30000

0

2000

4000

6000

8000

10000

128 256 512 1522 9216

Max Latency 2960 3320 4280 6720 28220

Avg Latency 2438 2578 2923 4256 14094

Min Latency 2280 2360 2580 3460 9600

ns

Avg Latency

Max Latency

Min Latency


Dual Homed TestUnicast Return From Multicast Flows(min, max, avg latency)

Multicast Results

Te Extreme Open Fabric system latency or multicast traffic

varied rom a low o 2.4 microseconds to a high o 38 mi-

croseconds. Te results were the same as the single-homed

configuration even though increased reliability or availability

was introduced to the design via dual-homing server ports

to two Summit® X670V oRs plus MLAG between oRs. As

expected, latency increased with packet size where 9216 size

packets experienced the largest system delay. Again, 75%

packet loss was expected and observed across all packet sizes

o multicast traffic. Note that currently Ixia’s statistics do not

support the combination o multicast traffic running over

ports within a LAG. Tereore, packet loss or this scenario

was not accurately calculated and is, thereore, not valid.

0

5000

10000

15000

20000

25000

0

10000

20000

30000

40000

0

2000

4000

6000

8000

10000

128 256 512 1522 9216

Max Latency 3440 4140 5540 11260 38300

Avg Latency 2697 3042 3565 5818 22533

Min Latency 2420 2480 2720 3600 9740

ns

Avg Latency

Max Latency

Min Latency


Dual Homed Test

Multicast Traffic(min, max, avg latency)






Many-to-Many Results

Te Extreme Open Fabric system latency or many-to-many

unicast traffic in a mesh configuration varied rom a low o

830 nanoseconds to a high o 45 microseconds. Again, theresult was the same as the single-homed configuration, even

though increased reliability or availability was introduced

to the design via dual-homing server ports to two Sum-

mit® X670V oRs plus MLAG between oRs. As expected,

latency increased with packet size where 9216 size packets

experienced the largest system delay. Zero packet loss was

observed across all packet sizes o unicast traffic.

0

5000

10000

15000

20000

0

10000

20000

30000

40000

50000

0

400

800

1200

1600

128 256 512 1522 9216

Max Latency 3160 3720 5220 9360 45440

Avg Latency 2045 2223 2730 4882 19121

Min Latency 820 860 1000 1220 1220

ns

Avg Latency

Max Latency

Min Latency


Dual Homed TestMany to Many Full Mesh Flows(min, max, avg latency)

Te table below illustrates the average system latency across

packet sizes 128 to 9216 or unicast, multicast and many-to-

many traffic flows through the Extreme Networks dual-homed configuration. All traffic flows perormed in a tight

range between 2 and 5.8 microseconds except or the large

packet size o 9216 where latency increased by an approxi-

mate actor o five. Tis was mostly due to the act that the

9216 packet size was six times larger than the previous 1522

size and required substantially more time to pass through

the Fabric.

0

5000

10000

15000

20000

25000

0

5000

10000

15000

0

5000

10000

15000

20000

128 256 512 1522 9216

Unicast 2438 2578 2923 4256 14094

Multicast 2697 3042 3565 5818 22533

Many-to-Many 2045 2223 2730 4882 19121

Unicast Return From

Multicast Flows Avg Latency

Multicast Traffic Avg Latency

Many-to-Many

Full Mesh Flows Avg Latency

ns

Extreme 2-X670V ToR 2-BD X8 Core

Dual Homed TestUnicast Return From Multicast Flows,Multicast Traffic, Many-to-Many FullMesh Flows (avg latency)






Cloud Performance Test

In addition to testing the Extreme Open Fabric with uni-

cast, multicast and many-to-many traffic flows at varying

packet sizes, the Lippis Cloud Perormance est iMix wasalso used to generate traffic and measure system latency

and throughput rom ingress to egress. o understand the

perormance o Extreme’s Open Fabric system under load,

we ran six iterations o the Lippis Cloud Perormance est

at traffic loads o 50%, 60%, 70%, 80%, 90% and 100%, mea-

suring latency and throughput on the X670V oR switch.

See the methodology section or a ull explanation o this

test. Te X670V oR was connected to Ixia test gear via 28-

10GbE links.

Lippis Cloud Performance

Te Extreme X670V perormed flawlessly over the six Lip-

pis Cloud Perormance iterations. Not a single packet was

dropped as the mix o east-west and north-south traffic

increased in load rom 50% to 100% o link capacity. Te

average latency was stubbornly consistent as aggregate

traffic load was increased. Microsof Exchange, Youube

and database traffic were the longest to process with 100

nanoseconds more latency than HP and iSCSI flows. Te

difference in latency measurements between 50% and 100%o load across protocols was 181ns, 146ns, 87ns, 93ns and

85ns, respectively, or HP, Youube, iSCSI, Database and

Microsof Exchange traffic. Tis was a very tight range with

impressive results as it signifies the ability o the Fabric to

deliver consistent perormance under varying load.

Reliability

Extreme Networks went above and beyond what was

required in the Lippis/Ixia Active-Active Fabric est to

demonstrate and test the reliability o a data center Fabric

built with its Summit® X670V and BlackDiamond® X8.

Within Extreme’s two-tier switch architecture, reliability

was tested in three critical areas: 1) between Ixia test gear

and the Summit® X670V, 2) between Summit® X670V and

BlackDiamond® X8 and 3) with the loss o an entire Black-

Diamond® X8. Te Summit® X670V and BlackDiamond® X8

were configured in the “single-homed” test design.

0

200

400

600

800

1000

1200

1400

NS_HTTP YouTube iSCSI DB MS Exchange

50% Load 60% Load 70% Load 80% Load 90% Load 100% Loadns

Summit X670V ToR IxCloud

Performance Test

28 ports of 10GbE

(avg latency)

Server to Summit® X670V Reliability Test

A stream o unicast many-to-many flows at 128-byte size

packets was sent to the Extreme Open Fabric. While the

Fabric was processing this load, a 10GbE link was discon-

nected in LAGs 3, 4, 7 and 8. For many-to-many unicast

traffic, 0.139% packet loss was observed over 87 millisec-

onds packet loss duration.

L1 L2 L5 L6 L3 L4 L7 L8

10G

40G

4x40G BD X8

Summit X670

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32


XXXX






Summit® X670V to BlackDiamond® X8Reliability Test

Tere were two parts to the Lippis/Ixia Reliability est.

First, a QSFP+ 40GbE optical cable that connected Sum-mit® X670V switches to BlackDiamond® X8 switches while

unicast many-to-many traffic flows were being processed is

pulled, and the resulting packet loss plus packet loss dura-

tion was recorded by Ixia test equipment. Ten the 40GbE

link was restored and the resulting packet loss plus packet

loss duration was recorded by Ixia test equipment. Tere are

our-40GbE links that connected a Summit® X670V to two

BlackDiamond® X8s.

L1 L2 L5 L 6 L3 L4 L 7 L8

10G

40G

4x40G BD X8

Summit X670

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32


XXX

X

Once again, a stream o unicast many-to-many flows at

128-byte size packets was sent to the Extreme Open Fabric.

While the Fabric was processing this load, a 40GbE link was

disconnected between Summit® X670V and BlackDiamond®

X8. For many-to-many unicast traffic, 0.003% packet loss was

observed over 3 milliseconds packet loss duration. When the40GbE link was restored, the Fabric reconfigured itsel in 56

milliseconds and lost 0.026% o packets in that time.

BlackDiamond® X8 Shut Down Reliability Test

Tere were two parts to this Reliability est o losing an en-

tire Core switch. First, the BlackDiamond® X8 Core switch

was shut down while unicast many-to-many traffic flows

were being processed, and the resulting packet loss plus

packet loss duration was recorded by Ixia test equipment.

0.00%

0.50%

1.00%

1.50%

2.00%

2.50%

3.00%

MLAG Link from

ToR to Core -

Cable Pull

MLAG Link from

ToR to Core -

Restoring the Link

0

10

20

30

40

50

60

ms

MLAG Core,

entire BD 8X

Goes Down

All 4-40GbE links -

Switch Shutdown

Event

MLAG Core,

entire BD 8X

Goes Down

All 4-40GbE links -

Switch

Restoration


Single Homed Configuration Reliability Test

Based upon 128 Byte size packet

Ten the BlackDiamond® X8 Core switch was restored,

and the resulting packet loss plus packet loss duration was

recorded by Ixia test equipment.

For many-to-many unicast traffic, 0.01% packet loss was

observed over 4.3 milliseconds packet loss duration. When

the BlackDiamond® X8 was restored, the Fabric reconfig-ured itsel in 42 milliseconds and lost 0.021% o packets in

that time.

L1 L2 L 5 L6 L3 L 4 L7 L8

10G

40G

4x40G BD X8

Summit X670

2 4 5 6 17 20 21 24 9 12 15 16 25 28 29 32


X






Discussion

Te Extreme Networks Open Fabric data center network, built with its Summit®

X670V oR switches and BlackDiamond® X8 Core switches, has undergone the

most comprehensive public test or data center network Fabrics, and has achievedoutstanding results in each o the key aspects o networking. While the Open

Fabric architecture supports multiple, high availability network designs and con-

figurations, during the Lippis/Ixia Active-Active Ethernet Fabric est, a two-tier

network was implemented or both single homed and dual homed topologies.

We ound that its system latency was very consistent under different packet sizes,

payloads and traffic types. It perormed to expectations; that is, large-size packet

streams consume more time to pass through the switches resulting in greater

latency. Tere was no packet loss in either single- or dual-homed configurations,

while it was supplied 320 Gbs o unicast, multicast and many-to-many traffictypes to process. In addition to processing a mix o different traffic types, Ex-

treme Networks Open Fabric perormed just as outstandingly during the Lippis

Cloud Perormance est, processing a mix o HP, Youube, iSCSI, Database

and Microsof Exchange traffic that increased in load rom 50% to 100% o ca-

pacity. Here, too, its latency was consistently low with zero packet loss and 100%

throughput achieved.

Extreme used MLAG to implement its active-active protocol. Extreme’s MLAG

implementation perormed flawlessly, proving that a two-tier data center net-

work architecture built with its Summit® X670V oR and BlackDiamond® X8Core switches will scale with perormance.

Regarding scale, Extreme Networks tested two tier topologies with Open Fabric

in these tests, demonstrating versatility and robustness in the single and dual

homed topologies. Note that due to the capacity o the BlackDiamond X8 core

switches, these designs can scale-up to support large network implementations

with small oversubscription ratios at the oR level. Similarly, i needed large

topologies can be deployed without oversubscription by deploying dual Black-

Diamond X8 core switches in a single tier architecture, yet supporting active-

active connectivity rom servers. Tis design can scale up to 750+ 10GbE or 190+40GbE attached active-active servers. Tese designs make Extreme Networks

stand out in their ability to deploy highly scalable designs or active-active net-

works.

Extreme Networks solutions, as tested, ocused on Layer 2 active-active archi-

tectures, its Open Fabric hardware and sofware, as tested, also supports Layer

3 orwarding or routing and redundancy protocols. As such, Equal Cost Multi

Path (ECMP) based designs are commonly deployed with these products or

active-active Layer 3 connectivity.

In this video podcast, Extreme

demonstrated the ease in which VMs are

moved between data centers when its OpenFabric was connected via VPLS or Virtual

Private LAN Service.

Video feature: Click to view

VM Migration Video Podcast

http://lippisreport.com/?p=8013






One o the big and very positive surprises o Extreme’s set o active-active test

results concerned its reliability results. Links between Ixia gear and Summit®

X670V oR switches were disconnected and re-established to measure packet

loss and packet loss duration. In addition, an MLAG link between Summit®

X670V oR switches and BlackDiamond® X8 Core switches were disconnectedand re-established to measure packet loss and packet loss duration. Finally, an

entire BlackDiamond® X8 Core switch was shut down and powered back up

to measure packet loss and packet loss duration. Tese tests proved that the

Extreme Networks Open Fabric architecture is resilient and reliable as packet

losses in the range o .266% to .003% were observed or different traffic types and

points o disconnection. Packet loss duration was measured in the hundreds o

milliseconds, well under the threshold o CP (ransmission Control Protocol)

timeouts, so that users would not be disconnected rom their applications during

a network disruption.

Data center networking is moving in multiple directions o efficiency. Converged

I/O hopes to reduce cabling and storage switch cost by combining both storage

and datagram traffic over one Ethernet Fabric. Te Open Networking standards

approach to networking looks to reduce operational spend by centralizing net-

work control where northbound APIs abstract network services so that applica-

tions and data center orchestration systems can automate network configuration.

As these trends develop and grow, a stable, high perorming data center network

inrastructure that scales becomes ever so important. Extreme Networks Open

Fabric demonstrated high perormance and reliability under various loads and

conditions during this Lippis/Ixia Active-Active Fabric est. Its Open Fabric roadmap includes 100GbE, Open Flow and OpenStack, a set o northbound APIs, a

new unified orwarding table to support L2/L3 and flows, application tie-ins and

much more. We find that Extreme Networks Open Fabric is an excellent choice

to consider or modern data center networking requirements.






Active-Active Fabric Cross-Vendor Analysis

o deliver the industry’s first test suite o Ethernet abrics Ixia, Lippis Enterprises

and all vendors provided engineering resources to assure that the configura-

tion files are suitable or MLAG, RILL and SPB configurations. In addition,the results are repeatable, a undamental principal in testing. Tis test was more

challenging thanks to the device under test being a abric or “system” versus a

single product. Tese participating vendors are:

Arista Sofware-Defined Cloud Network

Avaya Virtual Enterprise Network Architecture Fabric Connect

Brocade Virtual Cluster Switching


Brocade was the first company to be tested. Te test suite evolved afer its first

test, thus we do not include Brocade’s single and dual homed results in the cross-

vendor section due to a different traffic mix utilized or Brocade’s VCS. Further,

each vendor was offered optional testing opportunities which some accepted and

some declined. For the cross-vendor analysis we report on only required aspects

o the Lippis/Ixia Active-Active Ethernet Fabric est.

Tese abric configurations represent the state-o-the-art in two-tier networking.

Pricing per abric varies rom a low o $75K to a high o $150K. oR and Core

switch port density impact pricing as does 10GbE vs 40GbE. Price points on a

10GbE per port basis are a low o $351 to a high o $670. 40GbE oR switch price

per port is as low as $625 to $2,250 per port. In the Core 10GbE price per port is

as low as $1,200 while 40GbE ports are as low as $6,000 per port.

We compared each o the above firms’ abrics in terms o their ability to orward

packets: quickly (i.e., latency), without loss o their throughput at ull line rate or

three types o traffic, unicast mesh in a many-to-many configuration, multicast

and unicast returns rom multicast peers. We compare Server-oR reliability and

how each oR perorms during the Lippis Cloud simulation test.






128-bytes 256-bytes 512-bytes 1522-bytes 9,216-bytes *

Single Homed Dual Homed


Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Avaya VENA2

Fabric Connect

4-VSP 7000s ToR

Brocade VCS3,4

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Avaya VENA2

Fabric Connect

4-VSP 7000s ToR

Brocade VCS3,4

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

Arista Avaya BrocadeExtreme Arista Avaya Brocade Extreme

0

10000

20000

30000

40000

0

30000

60000

90000

120000

150000

ns ns

Fabric Test, Many-to-Many Full Mesh Unicast - Average Fabric Latency

All Switches Performed at Zero Frame Loss

1Software-Defined Cloud Network2Virtual Enterprise Network Architecture3Virtual Cluster Switching4

Brocade’s test was based on slightly different traffic profile and thus is not included here

* The latency measurement was unexpectedly high andhighly probable that it was the result of buffering andnot a true measure of fabric latency.

1Software-Defined Cloud Network2Virtual Enterprise Network Architecture3Virtual Cluster Switching4Brocade’s test was based on slightly different traffic profile and thus is not included here

Single Homed Double Homed

Framesize

(bytes)

Arista SDCN1 2-7050S-64 ToR &

2-7508 Core

Avaya VENA2 Fabric Connect

4-VSP 7000s ToR

Brocade VCS3,4 2-VDX & 6720 ToR2-VDX8770 Core

Extreme NetworksOpen Fabric

2-X670V ToR &2-BD X8 Core


2-7508 Core


4-VSP 7000s ToR




128 6993 2741 2065 6446 2356 2045

256 10189 3447 2297 9167 2356 2223

512 16375 4813 2887 14375 2428 2730

1,522 39780 10491 4789 34007 2724 4882

9,216 143946 53222 19115 142800 6258 19121

Brocade’s test

was based on a

slighty different

traffic profile

and thus is not

included here.

Brocade’s test

was based on a

slighty different

traffic profile

and thus is not

included here.

Jumbo rame 9216 size packet size tra-

fic requires significantly more time to

pass through the abric, thanks to seri-

alization, thereore, its plotted on a sep-

arate graphic so that smaller size packet

traffic can be more easily viewed.

Extreme Networks Open Fabric de-

livered the lowest latency or singled

homed ully meshed unicast traffic

Note: Avaya VSP 7000 s wereconfigured in Store andForward Mode whileall others were in CutThrough which mayimpact comparison ofthe data directly.






o packet sizes 128-1522 ollowed by

Avaya Fabric Connect and Arista’s

SDCN. Avaya’s Fabric Connect deliv-

ered the lowest latency or dual homed

ully meshed unicast traffic o packetsizes 128-1522. Te Extreme Net-

works Open Fabric and Avaya’s Fabric

Connect Dual homed results or ully

meshed unicast traffic o packet sizes

128-1522 were nearly identical. Note

that Avaya’s Fabric Connect was con-

figured with oR switches while Arista

and Extreme provided oR and Core

Switches, thereore, there is significant-

ly more network capacity with theseconfigurations.

Both Arista and Extreme dual homed

results are the same as single homed,

as expected. Avaya’s dual homed re-

sult is lower than its single homed and

is due to increased bandwidth between

the oRs. Further, Core switches do

take longer to process packets, thanksto their higher port densities and inter-

module system abrics. Avaya’s lack o

a Core switch provided it an advantage

in this test.

We don’t believe that IxNetwork la-

tency measurement o a abric in Cut-

through (C) or Store and Forward

(SF) is material. As the SF RFC 1242 la-

tency measurement method is the timeinterval starting when the last bit o the

input rame reaches the input port and

ending when the first bit o the output

rame is seen on the output port (LIFO)

while C RFC 1242 latency measure-

ment method is the time interval start-

ing when the end o the first bit o the

input rame reaches the input port and

ending when the start o the first bit othe output rame is see on the output

port (FIFO). Te measurement differ-

ence between C vs. SF in a abric un-

der test is the size o one packet rom

the starting point; in essence it’s the se-

rialization delay o one packet size. C

vs. SF on device latency measuremen

are material. Given the above and the

act that Avaya’s VSP 7000s were con-

figured in SF while all other switcheswere configured or C, we cannot rule

out an anomaly that may impact Ava-

ya’s abric latency measurement during

this industry test.






128-bytes 256-bytes 512-bytes 1522-bytes 9,216-bytes *



Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Arista Avaya BrocadeExtreme Arista Avaya Brocade ExtremeAvaya VENA2

Fabric Connect

4-VSP 7000s ToR

Brocade VCS3,4

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Avaya VENA2

Fabric Connect

4-VSP 7000s ToR

Brocade VCS3,4

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

ns ns

0

10000

20000

30000

40000

0

50000

100000

150000

200000

250000

300000

350000

400000

450000

Fabric Test, Multicast Traffic - Average Fabric Latency


1Software-Defined Cloud Network2Virtual Enterprise Network Architecture3Virtual Cluster Switching

4Brocade’s test was based on slightly different traffic profile and thus is not included here

1Software-Defined Cloud Network2Virtual Enterprise Network Architecture3Virtual Cluster Switching4Brocade’s test was based on slightly different traffic profile and thus is not included here


Framesize

(bytes)


2-7508 Core


4-VSP 7000s ToR





2-7508 Core


4-VSP 7000s ToR




128 7693 3439 2691 8104 2892 2697

256 7792 4312 3053 11329 2884 3042

512 10302 6082 3517 16931 3036 3565

1,522 19983 13594 5700 38407 3951 5818

9,216 415230 53670 22382 402035 15362 22533

Brocade’s test

was based on a

slighty different

traffic profile

and thus is not

included here.

Brocade’s test

was based on a

slighty different

traffic profile

and thus is not

included here.

* The latency measurement was unexpectedly high andhighly probable that it was the result of buffering andnot a true measure of fabric latency.


livered the lowest latency or singled

homed multicast traffic o packet sizes

128-1522 ollowed by Avaya Fabric

Connect and Arista’s SDCN. Notice

that or all vendors there is little di-

erence in latency between 128-512,

which provides consistent perormance

or applications running in that range

Even as the packet size increased by a

actor o two so does latency in most

cases.

Note: Avaya VSP 7000 s were configuredin Store and Forward Mode while allothers were in Cut Through which mayimpact comparison of the data directly.






Avaya’s Fabric Connect delivered the

lowest latency or dual homed multicast

traffic o packet sizes 256-9216. Te

Extreme Networks Open Fabric and

Avaya’s Fabric Connect Dual homed re-sults or multicast traffic o packet sizes

128-1522 were nearly identical. Note

that Avaya’s Fabric Connect was con-



switches; thereore, there is significant-

ly more network capacity with these

configurations.

Both Arista and Extreme dual homedresults are the same as single homed,

as expected. Avaya’s dual homed re-

sult is lower than its single homed and

is due to increased bandwidth between

the oRs. Further, Core switches do

take longer to process packets, thanks

to their higher port densities and inter-

module system abrics. Avaya’s lack o

a Core switch provided it an advantage

in this test.

Arista’s multicast latency in both single

and dual homed is anomalistically high.

Te reason is that the Arista 7500 has a

unique architecture with a credit-based

Fabric scheduler. Tis design allowsor airness across flows throughout

the system and allows or efficient uti-

lization o the Fabric bandwidth. Uni-

cast traffic is buffered on ingress using

VOQs. Tere are over 100,000 VOQs in

the system divided into eight classes o

traffic. By design multicast traffic can-

not be buffered on ingress as multiple

ports could be members o a multicast

group, and the packets must be trans-mitted to all destination ports without

dependencies on each other.

Te Arista 7500 provides a very large

amount o multicast bandwidth by

replicating on the ingress silicon, at

the Fabric and also at the egress. Tis

three-stage replication tree allows the

platorm to deliver wire speed multicast

to all 384 ports simultaneously. I there

is congestion on the egress port, mul-

ticast packets destined to that port are

buffered at egress. Other ports are not

affected by this congestion and head o

line blocking is avoided. When thereare multiple multicast sources, and 9K

packets are used, traffic is burstier and

it’s possible to overflow the egress bu-

ers. Such a burst could result in drop-

ping a small percentage o the overal

traffic thus increasing measured laten-

cy. Tis separation o buffering in the

ingress or unicast traffic and egress or

multicast traffic allows the 7500 to per-

orm well under real-world scenarioswith mixed unicast and multicast tra-

fic patterns. o obtain an accurate ab-

ric latency measurement, Arista recom-

mends that multicast traffic is run at a

no-drop rate across all nodes.






128-bytes 256-bytes 512-bytes 1522-bytes 9,216-bytes



Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Arista Avaya BrocadeExtreme Arista Avaya Brocade ExtremeAvaya VENA2,3

Fabric Connect

4-VSP 7000s ToR

Brocade VCS4,5

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

Arista SDCN1

2-7050S-64 ToR &

2-7508 Core

Avaya VENA2,3

Fabric Connect

4-VSP 7000s ToR

Brocade VCS4,5

2-VDX6720 ToR &

2-VDX8770 Core

Extreme Network

Open Fabric

2-X670V ToR &

2-BD X8 Core

ns ns

0

2000

4000

6000

8000

10000

12000

0

10000

20000

30000

40000

Fabric Test, Unicast Return From Multicast Flows - Average Fabric Latency


1Software-Defined Cloud Network2,Virtual Enterprise Network Architecture3Avaya’s multicast traffic was simulated by forwarding4Virtual Cluster Switching5Brocade did not support multicast at test time.

1Software-Defined Cloud Network2,Virtual Enterprise Network Architecture3Avaya’s multicast traffic was simulated by forwarding4

Virtual Cluster Switching5Brocade did not support multicast at test time.


Framesize

(bytes)


2-7508 Core

Avaya VENA2,3 Fabric Connect

4-VSP 7000s ToR





2-7508 Core


4-VSP 7000s ToR




128 5437 2775 2445 5346 2510 2438

256 6183 3181 2591 6062 2561 2578

512 7358 4055 2923 7111 2675 2923

1,522 11059 7883 4366 10292 3041 4256

9,216 35693 39197 14117 35661 5803 14094

Brocade did not

support multicast

at test time

Brocade did not

support multicast

at test time


livered the overall lowest latency or

singled and dual homed unicast return

rom multicast flows traffic. Avaya

Fabric Connect offered lower latency at

packet sizes 1522-9216 or dual homed

unicast return rom multicast flows,

however it delivered the largest latency

or single homed jumbo packet sizes o

9216, which is anomalistic

Note: Avaya VSP 700 0s wereconfigured in Store andForward Mode whileall others were in CutThrough which mayimpact comparison ofthe data directly.






Notice that or all vendors there is little

difference in latency between 128-512,

on the order o hundreds o nanosec-

onds, which provides consistent peror-

mance or applications running in thatrange. Even as the packet size increased

by a actor o two so does latency in

most cases.

Arista’s SDCN and Extreme Networks

Open Fabric perormed consistently

within single and dual homed with

each set o data being nearly identical,

which is expected and desired. Notethat Avaya’s Fabric Connect was con-



switches; thereore, there is significant-

ly more network capacity with these

configurations.

Avaya’s dual homed result is lower than

its single homed and is due to increased

bandwidth between the oRs. Further

Core switches do take longer to process

packets, thanks to their higher portdensities and inter-module system ab-

rics. Avaya’s lack o a Core switch pro-

vided it an advantage in this test.






0.00%

0.05%

0.10%

0.15%

0.20%

0.25%

0.30%

Arista

SDCN1

Avaya

VENA2

Fabric Connect

Brocade4

VCS3

Packet Loss %

Extreme Networks

Open Fabric

0

50

100

150

200

250

300

% ms

Packet Loss Duration (ms)

Server-ToR Reliability Test

128 Byte Size Packet In Many-to-Many Full Mesh Flow Through Fabric

One 10GbE Cable in LAGs 3, 4, 7, and 8 Between Ixia-to-ToR is Pulled

Server-ToR Reliability Test

128 Byte Size Packet In Many-to-Many Full Mesh Flow Through Fabric

One 10GbE Cable in LAGs 3, 4, 7, and 8 Between Ixia-to-ToR is Pulled

1Software-Defined Cloud Network2,Virtual Enterprise Network Architecture3Virtual Cluster Switching4Brocade did not test for reliability

1Software-Defined Cloud Network2,Virtual Enterprise Network Architecture3Virtual Cluster Switching4Brocade did not test for reliability

Company Fabric Name Fabric Products Packet Loss % Packet Loss Duration (ms)

Arista SDCN1 2-7050S-64 ToR & 2-7508 Core 0.118% 70.179

Avaya VENA2 Fabric Connect 4-VSP 7000s ToR 0.283% 278.827

Brocade4 VCS3 2-VDX6720 ToR & 2-VDX8770 Core

Extreme Networks Open Fabric 2-X670V ToR & 2-BD X8 Core 0.139% 87.018

Arista’s SDCN delivered the lowest

packet loss and shortest packet loss du-

ration in the server-oR reliability test

ollowed by Extreme Network’s Open

Fabric, then ollowed by Avaya’s Fab-

ric Connect. Te difference between

Arista and Extreme’s results or this re-

liability test is 17 milliseconds o packet

loss duration and .024% packet loss; a

narrow difference. Avaya’s Server-oR

packet loss duration is approximately

our times that o Arista.






NS HTTP NS Youtube EW iSCSI EWDB EW MSExchange

Arista SDCN1

7050S-64 ToR

Avaya4 VENA2 Fabric Connect

VSP 7000s ToR

Brocade VCS3

VDX6720 ToR


X670V ToR

0

500

1,000

1,500

2,000

ns

1Software-Defined Cloud Network2,Virtual Enterprise Network Architecture3Virtual Cluster Switching4Avaya did not test for Cloud Simulation


Company Fabric Name Fabric Products NS HTTP NS Youtube EW iSCSI EW DB EW MSExchange

Arista SDCN1 7050S-64 ToR 1121 1124 987 1036 1116

Avaya4 VENA2 Fabric Connect VSP 7000s ToR

Brocade VCS3 VDX6720 ToR 1929 1859 1815 1857 597

Extreme Networks Open Fabric X670V ToR 1149 1196 1108 1206 1292

Cloud Simulation ToR Switches At 50% Aggregate Traffic Load

Zero Packet Loss: Latency Measured in ns

28-10GbE Configuration Between Ixia-ToR SwitchTested While In Single Homed Configuration









NS HTTP NS Youtube EW iSCSI EWDB EW MSExchange

Arista SDCN1

7050S-64 ToR

Avaya4 VENA2 Fabric Connect

VSP 7000s ToR

Brocade VCS3

VDX6720 ToR


X670V ToR

0

2,000

4,000

6,000

8,000

10,000

ns



Company Fabric Name Fabric Products NS HTTP NS Youtube EW iSCSI EW DB EW MSExchange

Arista SDCN1 7050S-64 ToR 1187 1184 1033 1083 1156

Avaya4 VENA2 Fabric Connect VSP 7000s ToR

Brocade VCS3 VDX6720 ToR 4740 3793 10590 9065 602

Extreme Networks Open Fabric X670V ToR 1330 1342 1195 1300 1376












Arista Network’s 7050S-64 delivered

the lowest latency measurement or the

Lippis Cloud simulation test at 50% and

100% load, ollowed by Extreme Net-

works’ X670V and Brocade’s VCS VDX6720. Both the Arista Networks’ 7050S-

64 and Extreme Networks’ X670V de-

livered nearly consistent perormance

under 50% and 100% load with varia-

tion o a ew hundred nanoseconds per

protocol, meaning that there is plentyo internal processing and bandwidth

capacity to support this traffic load

Te Brocade’s VCS VDX 6720 was

slightly more variable. All products de-

livered 100% throughput, meaning that

not a single packet was dropped as load varied.






Ethernet Fabric Industry Recommendations

Te ollowing provides a set o recommendations to I

business leaders and network architects or their consider-

ation as they seek to design and build their private/publicdata center cloud network abric. Most o the recommenda-

tions are based upon our observations and analysis o the

test data. For a ew recommendations, we extrapolate rom

this baseline o test data to incorporate key trends and how

these Ethernet Fabrics may be used in their support or

corporate advantage.

Consider Full Mesh Non Blocking: Most o the abric

configurations were ully meshed and non-blocking which

provided a highly reliable and stable inrastructure. Tisarchitecture scales, thanks to active-active protocols, and

enables two-tier design, which lowers equipment cost plus

latency. In addition to being highly reliable, it also enables

dual-homed server support at no perormance cost.

Consider Two-Tier Network Fabric: o reduce equip-

ment cost, support a smaller number o network devices

and increase application perormance, it’s recommended to

implement a two-tier lea-spine Ethernet Fabric. Tis Lip-

pis/Ixia Active-Active Ethernet Fabric est demonstrated

that two-tier networks are not only ready or prime-time

deployment, they are the preerred architecture.

MLAG and ECMP Proven and Scales But: It was proven

that a two-tier network can scale, thanks to ECMP up to

32-way links. In addition, ECMP offers multipathing, too,

at scale. What’s missing rom ECMP is auto provisioning

o links between switches. In short, ECMP requires manual

configuration.

Consider Utilizing TRILL and/or SPB: Over time, most

vendors will support RILL and SPB in addition to MLAG

and ECMP. Both RILL and SPB offer unique auto-provi-

sioning eatures that simpliy network design. It’s recom-

mended that network architects experiment with both

active-active protocols to best understand its utility within

your data center network environment.

Strong Underlay or a Dynamic Overlay: Te combination

o ully meshed, non-blocking two-tier network build with

standard active-active protocols constructs a strong under-lay to support a highly dynamic overlay. With the velocity o

change in highly virtualized data centers ushering in virtu-

alized networks or overlays, a stable and scalable underlay is

the best solution to support the rapid build-up o tunneled

traffic running through Ethernet Fabrics. Tis huge demand

in overlay traffic is yet another good reason to consider a

two-tier active-active Ethernet Fabric or data center and

cloud networking.

Be Open to Different Fabric Architectures: Not all datacenters support 10,000 or 100,000 servers and require

enormous scale. Tere are different approaches to building

Ethernet Fabrics that are ocused on converged I/O or sim-

plicity o deployment, auto provisioning, keeping east-west

traffic at the oR tier, etc. Many vendors offering Ethernet

Fabrics offer product strategies to scale up as requirements

demand.

Get Ready or Open Networking: In this Lippis/Ixia est,

we ocused on the active-active protocols or all the rea-

sons previously mentioned. When considering an Ethernet

Fabric, it’s important to ocus on open networking, such as

the integration o the network operating system with Open-

Stack, or how oR and Cores support various SDN control-

lers, do the switches support OpenFlow or have a road map

or its support. Tere are three types o networks in data

centers today, L2/L3, Network Virtualization overlays that

tunnel traffic through L2/L3 and soon OpenFlow flows.

Consider those vendors that support all types o networking

as this is a ast-moving target. Auto provisioning o net-

working with compute and storage is increasingly impor-

tant; thereore, look or networking vendors that support

network configuration via SDN controllers plus virtualiza-

tion and cloud orchestration systems.






Terms of Use

Tis document is provided to help you understand whether

a given product, technology or service merits additional

investigation or your particular needs. Any decision topurchase a product must be based on your own assessment

o suitability based on your needs. Te document should

never be used as a substitute or advice rom a qualified

I or business proessional. Tis evaluation was ocused

on illustrating specific eatures and/or perormance o the

product(s) and was conducted under controlled, laboratory

conditions. Certain tests may have been tailored to reflect

perormance under ideal conditions; perormance may vary

under real-world conditions. Users should run tests based

on their own real-world scenarios to validate perormanceor their own networks.

Reasonable efforts were made to ensure the accuracy o

the data contained herein but errors and/or oversights can

occur. Te test/ audit documented herein may also rely

on various test tools, the accuracy o which is beyond our

control. Furthermore, the document relies on certain rep-

resentations by the vendors that are beyond our control to

veriy. Among these is that the sofware/ hardware tested is

production or production track and is, or will be, avail¬able

in equivalent or better orm to commercial customers.Accordingly, this document is provided “as is,” and Lippis

Enterprises, Inc. (Lippis), gives no warranty, representation

or undertaking, whether express or implied, and accepts no

legal responsibility, whether direct or indirect, or the accu-

racy, completeness, useulness or suitability o any inorma-

tion contained herein.

By reviewing this document, you agree that your use o any

inormation contained herein is at your own risk, and you

accept all risks and responsibility or losses, damages, costsand other consequences resulting directly or indirectly rom

any inormation or material available on it. Lippis is not

responsible or, and you agree to hold Lippis and its related

affiliates harmless rom any loss, harm, injury or damage

resulting rom or arising out o your use o or reliance on

any o the inormation provided herein.

Lippis makes no claim as to whether any product or compa-

ny described herein is suitable or in¬vestment. You should

obtain your own independent proessional advice, whether

legal, accounting or otherwise, beore proceeding with any

investment or project related to any inormation, products

or companies described herein. When oreign translations

exist, the English document is considered authoritative. o

assure accuracy, only use documents downloaded directly

rom www.lippisreport.com .

No part o any document may be reproduced, in whole or

in part, without the specific written permission o Lippis.

All trademarks used in the document are owned by their

respective owners. You agree not to use any trademark inor as the whole or part o your own trademarks in connec-

tion with any activities, products or services which are not

ours, or in a manner which may be conusing, misleading or

deceptive or in a manner that disparages us or our inorma-

tion, projects or developments.





About Nick Lippis

Nicholas J. Lippis III is a world-renowned authority on advanced IP networks,

communications and their benefits to business objectives. He is the publisher

o the Lippis Report, a resource or network and I business decision mak-ers to which over 35,000 executive I business leaders subscribe. Its Lippis

Report podcasts have been downloaded over 200,000 times; Iunes reports

that listeners also download the Wall Street Journal’s Money Matters, Business

Week’s Climbing the Ladder, Te Economist and Te Harvard Business Review’s

IdeaCast. He is also the co-ounder and conerence chair o the Open Networking User Group, which

sponsors a bi-annual meeting o over 200 I business leaders o large enterprises. Mr. Lippis is cur-

rently working with clients to design their private and public virtualized data center cloud comput-

ing network architectures with open networking technologies to reap maximum business value and

outcome.

He has advised numerous Global 2000 firms on network architecture, design, implementation, ven-

dor selection and budgeting, with clients including Barclays Bank, Eastman Kodak Company, Federal

Deposit Insurance Corporation (FDIC), Hughes Aerospace, Liberty Mutual, Schering-Plough, Camp

Dresser McKee, the state o Alaska, Microsof, Kaiser Permanente, Sprint, Worldcom, Cisco Systems,

Hewlett Packet, IBM, Avaya and many others. He works exclusively with CIOs and their direct reports

Mr. Lippis possesses a unique perspective o market orces and trends occurring within the computer

networking industry derived rom his experience with both supply- and demand-side clients.

Mr. Lippis received the prestigious Boston University College o Engineering Alumni award or ad-

vancing the proession. He has been named one o the top 40 most powerul and influential people in

the networking industry by Network World . echarget , an industry on-line publication, has named

him a network design guru while Network Computing Magazine has called him a star I guru.

Mr. Lippis ounded Strategic Networks Consulting, Inc., a well-respected and influential computer

networking industry-consulting concern, which was purchased by Sofbank/Ziff-Davis in 1996. He

is a requent keynote speaker at industry events and is widely quoted in the business and industry

press. He serves on the Dean o Boston University’s College o Engineering Board o Advisors as well

as many start-up venture firms’ advisory boards. He delivered the commencement speech to Boston

University College o Engineering graduates in 2007. Mr. Lippis received his Bachelor o Science in

Electrical Engineering and his Master o Science in Systems Engineering rom Boston University. His

Masters’ thesis work included selected technical courses and advisors rom Massachusetts Institute o

echnology on optical communications and computing.

EXTREME Lippis AA Report Spring 2013

Documents

Transcript of EXTREME Lippis AA Report Spring 2013