• No results found

Pre$SDN era: network trends in data centre networking

N/A
N/A
Protected

Academic year: 2021

Share "Pre$SDN era: network trends in data centre networking"

Copied!
19
0
0

Loading.... (view fulltext now)

Full text

(1)

Pre$SDN era: network trends

in data centre networking

Zaheer Chothia – 27.02.2015

(2)

Outline

2

Challenges and New Requirements

History of Programmable Networks

Spanning Tree Protocol HSTPI

Network Overlay Technologies

Examples: SPB, TRILL, FabricPath, MPLS

Emerging Next$Generation Protocols

(3)

Networks have not kept pace

“In the modern data center, traditional technologies

are

limiting the speed, flexibility, scalability, and

manageability

of application deployments.”

Cloud data center networks must contend with:

• 

huge numbers of attached devices Hboth physical and virtualI

• 

large numbers of isolated independent subnetworks

• 

multitenancy Hdifferent tenants collocated on a single hostI

• 

automated creation, deletion, and migration of virtual machines

(4)

Requirements of Modern Data Center

4

GDeHcoupling

Scalability

Mobility

Others

•  Virtualized networks •  Optimized forwarding •  Cloud integration •  … and many more

(5)

Requirements of Modern Data Center

5

GDeHcoupling

Physical/logical connectivity Want to deploy and expand workloads anywhere

Can extend VLAN domain but this affects availability Hlarger fault domainI + reconfiguration and administrative overhead

Infrastructure and policy Common practice:

•  Group entities with like membership into a VLAN •  IP addressing schemes based

on subnet boundaries Results in many inefficiencies and limitations Hcyclic changeI

Scalability

More end hosts and isolated subnetworks

Forwarding tables Network uses end`host

information HIP/MAC addressI to make forwarding decisions Need to propagate this state

across entire data center fabric Network segments

Space limitations: 802.1Q supports at most 4,094 VLANs

Also desired:

•  traffic management •  secure segmentation •  performance isolation

Mobility

Use case: live migration of VMs Need to retain adequate

network state

Address of end host should be

independent of location in the network

Others

•  Virtualized networks •  Optimized forwarding •  Cloud integration •  … and many more

(6)

6

(7)

20 years of development

• 

Active Networking

–  Mid`90s: Internet took off; standardization is too slow! –  Approach: programmable functions in the network

–  Capsule model Hcode to execute at nodes carried in`band in data packetsI

• 

Separating Control and Data Planes

–  Early 2000s: increasing traffic volumes and greater emphasis on network reliability, predictability, and performance

–  Desired: better control of paths used to deliver traffic Htraffic engineeringI –  Logically centralized control + open interfaces to routers and switches

• 

OpenFlow API and Network OSes

–  Network experimentation at scale Hencouraged by successes of PlanetLab/EmuLabI –  Pragmatism: limit flexibility Hbuild on existing switchesI, but immediate deployability –  First widespread adoption of an open interface

(8)

The Road to SDN: An Intellectual


History of Programmable Networks

8

https://www.cs.princeton.edu/~jrex/papers/queue14.pdf https://www.youtube.com/watch?v=dkUDUb9GtH0

(9)

Spanning Tree Protocol GSTPH

Purpose: Ensures loop`free topology by blocking redundant paths

9

http://www.cisco.com/c/en/us/support/docs/lan`switching/spanning`tree`protocol/24248`147.html

Common Spanning Tree ?CST@

•  No load balancing possible; need to block one uplink for all VLANs

•  CPU is spared; only one instance needs to be computed

PerAVLAN Spanning Tree ?PVST@

•  Optimum load balancing Hodd`evenI

•  1000 separate instances Hfor each VLANI even though only two different final topologies

•  Considerably wastes CPU cycles for all of the switches in the network

(10)

STP drawbacks

• 

Designed with maximum

stability and safety

in mind

• 

Blocking redundant ports:

squandered

aggregate bandwidth

• 

Cannot HeasilyI

segregate

into smaller domains


Desirable for scalability, fault isolation, multi`tenancy

• 

Topology change: network

halts

while STP recalculates [see link below]

–  Bounded by Max_Age+2xForward_Time Htypical: 20 + 2*15 = 50 seconds!I

10

http://blog.ine.com/wp`content/uploads/2011/11/understanding`stp`rstp`convergence.pdf

• 

Patchwork:

– 

Rapid STP: faster convergence after a topology

change Hwith active confirmationI

– 

Multiple STP: per`VLAN spanning tree + block all

but one of alternate paths within each

(11)

Network Overlays

Benefits

•  Simplified management

•  Multi`tenancy: scalable beyond 4000 VLANs

•  Workload`anywhere Hmobility and reachabilityI

•  Arbitrary forwarding topologies over fixed underlay

Challenges

•  Decreased visibility: traceroute in overlay will not report individual underlay hop counts

•  Troubleshooting complexity:
 need to investigate mapping of
 virtual to physical topology

11

(12)

Bag of protocols

• 

Designed to address shortcomings mentioned earlier

• 

Typically employ

encapsulation

for transparency from host

Hrather than for logical separationI

• 

Surveyed: layer`2 ECMP designs with

multipath

routing

– 

Shortest Path Bridging HSPBI

– 

Transparent Interconnection of Lots of Links HTRILLI

– 

Cisco FabricPath

Brocade: Virtual Cluster Switching HVCSI


Based on TRILL Hdata planeI; doesn’t use IS`IS core

Juniper: QFabric Hshelved?I

(13)

Shortest Path Bridging GSPBH

• 

Consolidate into a one link state protocol:

–  STP, Multiple STP, Rapid STP –  Multichassis Link Aggregation

–  Multiple MAC Registration Protocol HMMRPI

• 

Service identifiers HI`SIDI for independence from backbone MAC

address / VLAN IDs

• 

New device: learn its immediate neighbourhood, compute shortest

bidirectional paths using link metrics Hsuch as ECMPI

• 

End points are fully aware of entire traffic path Hunlike TRILLI

• 

Two modes

–  SPBV: use VLAN ID for delineation and load balancing
 Learn MAC addresses on all bridges along shortest path

–  SPBM: uses B`MAC+B`VID HbackboneI to designate reachability


C`MAC HclientI addresses are never learned or looked up in the core

(14)

Transparent Interconnection of

Lots of Links GTRILLH

• 

Control plane: IS`IS for discovery and to distribute link`state

database Heach node has state of entire networkI

• 

Uses Routing Bridges HRBridgesI to provide:

– 

Shortest unicast paths Hunlike STP: no single tree constraintI

– 

Faster convergence times

– 

Load splitting over multiple paths

– 

Loop mitigation

– 

MAC scalability: relegated learning to the edge RBs

14

•  Several possible multicast trees •  More active paths Hefficient use

of bandwidthI

(15)

Cisco FabricPath

• 

Proprietary implementation of TRILL

– 

Control plane: utilizes TRILL Hincluding IS`IS for L2 multipath routingI

– 

Data plane: non`interoperable

• 

Forwarding tag GFTAGH

– 

for multi`destination, unicast frames;

– 

assigned on the edge port, honored throughout;

– 

selects one of multiple paths Hmax 1024I that the packet traverses

• 

Conversational MAC learning

– 

interface learns source MAC of ingress frame only if destination MAC is

already present in the table

– 

i.e. only learn if remote device is having a bidirectional conversation with a

locally connected device

– 

unknown unicast frames being flooded: no learning on edge switches

(16)

Framing formats compared

16

TRILL

Ethernet frame

FabricPath

Extended segment ID G24$bitsH


Supports over 16 million virtual networks

Variant 1 Variant 2

(17)

Bag of Protocols

• 

Keynote at 1

st

Open Networking Summit in 2011


“The Future of Networking, and the Past of Protocols” – Scott Schenker

• 

Why does networking lag behind?

“… because of your great ability to master complexity”

• 

Future of networking lies in finding right abstractions

The era of ‘a new protocol per problem’ is over

• 

SDN is defined precisely by these three abstractions

“Distribution, forwarding, configuration”

• 

SDN not just a random good idea…

“... can be “derived” from decomposing network control”

17

http://opennetsummit.org/archives/oct11/shenker`tue.pdf

(18)

My take

18

1.  Thankful I am not a protocol designer!

2.  Decoupling: limited integration between overlay and underlay

3.  Are the original problems fully addressed?


HHard to judge from the fencepostI

Discussion

Uninitiated / new format for me

6 critique / 2 in defense

(19)

19

T2: SDN basics and the Open Flow protocol

A Survey of Software`Defined Networking: Past,

Present, and Future of Programmable Networks

References

Related documents

2002-2004 Community Services Administrator, Refugee & Asylee Program, Lutheran Family Services of the Rockies, Denver, Colorado. 1999-2002 Research Director, Puget Sound

It is referred to as a: ‘Consent form for the donation of eyes, heart, body parts/ whole body’ and what the donor is actually signing up for is only slightly more specific: I

18 SW / HW Planning Process SW / HW Requirements Process Incremental Development Process Demonstrator (Prototype) System RIG Testing Flight Testing Equipment Specification

v  Identity and Access Intelligence is about leveraging identity information to enable better business

The studied Nursing home showed excellent compliance with planning and making residents’ nursing care plans avail- able for staff, referring residents after a fall,

S pomočjo aktualnih podatkov zemljiškega katastra in podatkov o dejanski rabi zemljišč Ministrstva za kmetijstvo in okolje Republike Slovenije (MKO) iz let 2002 in 2014 bomo

This integrative review of literature will explore and synthesize the national evidence and knowledge base currently available to guide research, practice and policy in the area

This is obvious in the special case where the linear variety collapses to a single point: the closed-loop adaptive system has then an exponentially stable