a tutorial overview with a voice over ip slant · 2007-03-05 · cisco systems, inc. 1 of 24 etsi...

1 of 24

IP

es:

r IP Slant

Cisco Systems, Inc.

ETSI Workhop on Voice overJune 9, 1999

Kathleen [email protected]

Differentiated Servic

A Tutorial Overview with a Voice ove

2 of 24ervices

Differentiated Services

del is an approach lly deployable way

s

ges maximal

rking on the

de up of of which is

Kathleen Nichols — ETSI Workshop: Differentiated S

Cisco Systems, Inc.

• The differentiated services architectural moto delivering QoS in a scalable, incrementathat:

› keeps control of QoS “local”

› pushes work to the edges and boundarie

› requires minimal standardization, encourainnovation

• The IETF Differentiated Services WG is wo“minimal standardization” part of this

• Diffserv’s model is based on an Internet maindependently administered domains, eachconnected to at least one other

3 of 24ervices

An Architectural Framework based on Clouds and

s are regions of control,

ay’


Cisco Systems, Inc.

Boundaries

Follows the structure of today’s Internet: Cloudrelative homogeneity in terms of administrativetechnology, bandwidth, etc.

‘Me’

‘Not-me-1’

‘Not-me-2’

‘Far Aw

4 of 24ervices

Clouds within Clouds


Cisco Systems, Inc.

Telecommuters

‘Me’

Training

Tele-commuters

Remote Site

5 of 24ervices

QoS and Clouds

to some locally

aries of clouds and

ndary

s is confined to ents where clouds


Cisco Systems, Inc.

• Within a cloud, QoS is allocated according determined set of rules

• Almost all the work is confined to the boundcovered by a set of rules

• Rules might not be symmetric across a bou

• QoS information exchanged between cloudboundaries and covered by bilateral agreemhave different owners

6 of 24ervices

Advantages of Model Based on Clouds

nistered regions of

technology might

p and evolve

ility

esn’t need to be ost nodes can be ection-oriented state for every yable or scalable


Cisco Systems, Inc.

• Clouds can map to the independently admithe Internet

• Architecturally agnostic: within a cloud any be used to deliver QoS

• Signaling agnostic and signaling can develo

• Possibility of multiple paths increases reliab

• QoS can be deployed in only one cloud, dosignaled per connection, and the state in mreduced considerably as compared to connapproaches which tie up resources, requireconnection and are not incrementally deplo

7 of 24ervices

Scalability through Aggregation

s to handle

ing QoS may be of rules which

they are to receive

al with the small track of every h


Cisco Systems, Inc.

• Fundamental to the diffserv approach is:

› there are a relatively small number of waypackets in the forwarding path

› the number of traffic conversations requirquite large and subject to a wide range devolve from policy

• Packets are grouped by the forwarding behaviorwithin a cloud

• Nodes in the center of a network only have to denumber of traffic aggregates rather than keepingseparate traffic conversation that passes throug

8 of 24ervices

Aggregation and Conversations

gates and are

dual conversations ame marking. Any tion

t the behavior performance for regate


Cisco Systems, Inc.

• The per-conversation state is kept at the edges

• Flows or conversations are classified into aggre“conditioned” to meet the rules of that aggregate

• Packets are not marked for the “services” indivimay be receiving. Many services may use the sviable service must make sense under aggrega

• Don’t distinguish between flows, so the treatmenaggregate receives should not result in differentdifferent traffic compositions of the behavior agg

9 of 24vices

“Minimal Standardization”: IETF RFC 2474

e packet’s et in IPv4 and the its 0-5 as a

into a table of

arding treatment o that behavior from, for

ally assigned and point to behavior

000000-111000

Kathleen Nichols — ETSI Workshop: Differentiated Ser

Cisco Systems, Inc.

• A bit-field in the packet header determines thforwarding treatment. DS field is the TOS octTraffic Class octet in IPv6; within that uses b“codepoint” field (DSCP)

• Codepoints should be looked at as an index packet forwarding treatments at each router.

• This table maps a DSCP to a particular forwor “per-hop behavior” (PHB) that is applied taggregate. PHBs are constructed by vendorsexample, particular queue schedulers

• Behavior for only a few codepoints to be globdiffserv-capable equipment must make codemapping flexible and accessible

• Class Selector Compliant PHBs get DSCPs

10 of 24ervices

The DSCP and PHBs in Use

lassifiers (both DS and configurable )markers. These can

a particular queue

odepoint ected to occur at

same behavior a “behavior

n, the packet ingress microflow or


Cisco Systems, Inc.

• Diffserv-capable equipment will include packet cfield and multiple fields of the IP packet header)traffic conditioners such as shapers, policers, (rebe used to control both selection and entry into

• A packet’s DS field may be marked with a c“anywhere” in the network (but marking expedges and boundaries)

• All packets with the same codepoint get thethus providing aggregation and scalability (aggregate”)

• Marking can be based on microflow identificatiolink, the measured temporal characteristics of aaggregate, etc.

11 of 24ervices

The EF Per-Hop Behavior (RFC 2598)

ut it is also the one

, can be ism

ets from any urable rate.

dependent of the ansit the node.

rate when r longer than the ed packet at the


Cisco Systems, Inc.

This is a forwarding behavior of general use, bmost useful for doing VoIP

In simple terms, it is a rough equivalent of PQimplemented by PQ with some safety mechan

More precisely (from RFC 2598)

“the departure rate of the aggregate’s packdiffserv node must equal or exceed a config

The EF traffic SHOULD receive this rate inintensity of any other traffic attempting to tr

It SHOULD average at least the configuredmeasured over any time interval equal to otime it takes to send an output link MTU sizconfigured rate. “

12 of 24ervices

Example Implementations of EF PHB

andwidth. BE is

one-half the MTU.

An MTU at thesubscribed rate

Arrival pattern

Priority Queueing

WRR/DRR


Cisco Systems, Inc.

Arrival rate is twice the output link bandwidth

EF queue configured to get 25% of the output link bconfigured to get 75% of the output link bandwidth.

Here packets come in only two sizes, the MTU and

EF EF

EF

EF EFBE BE

BE BE

BE EF BE

EF BE BE EF

EF

BE BE BE

13 of 24vices

Performance of VoIP using Diffserv’s EF PHB

eams. 10% of traffic f the EF packets, sults due to Y. Kim

24

FQ

R

Kathleen Nichols — ETSI Workshop: Differentiated Ser

Cisco Systems, Inc.

Median Jitter (|(aj-ai)-(dj-di)|) for voice type packet stris EF-marked, 60% gets other “special” treatment. Ohalf are long (1500 bytes), half short (100 bytes). (Reof Cisco)

5 ms

10 ms

15 ms

20 ms

4 8 12 16 20

PQCBWW RRMDR

number of congest T1 hops

14 of 24ervices

Performance of VoIP using EF PHB (cont)

24

W FQRR

RR


Cisco Systems, Inc.

95th percentile of jitter

10 ms

20 ms

30 ms

40 ms

50 ms

60 ms

4 8 12 16 20

PQCBWMD

number of congest T1 hops

15 of 24ervices

Services in the Diffserv Framework

bout building

behavior

boundaries

daries

e same aggregate

under aggregation


Cisco Systems, Inc.

• That covers the forwarding path, but what aservices?

• Services are built by adding rules to governaggregates:

› initial packet marking

› how particular aggregates are treated at

› temporal behavior of aggregates at boun

• Different user-visible services can share th

• Services must be sensible and quantifiable

16 of 24ices

Example: Putting Together an EF-based Service

routers know

Kathleen Nichols — ETSI Workshop: Differentiated Serv

Cisco Systems, Inc.

Natural Question: How do the border and edgewhat to mark?

17 of 24ervices

Controlling the Boundaries

track of priorities users, projects,

, consult and n information to

an Jacobson).

must authenticate also be

tations remain up


Cisco Systems, Inc.

• A repository of of policy is needed to keep and limits on QoS allocations for individual and/or departments.

• An entity needs to receive requests for QoSupdate the database, and send configuratiothe routers, where indicated.

• Call this entity a “bandwidth broker” (BB) (V

• BB is part of the network infrastructure andrequests from users. Some information canconfigured.

• Intradomain policy decisions and implemento each domain.

18 of 24ervices

Bandwidth Broker in the Enterprise

n, BB configures n (COPS-PR?)


Cisco Systems, Inc.

In a static or configured implementatioleaf/edges with soft state informatio

19 of 24ervices

Requests to a BB and the Result

requests


Cisco Systems, Inc.

Here the BB responds to user

20 of 24ervices

Requests Can Come from Many Sources

g”


Cisco Systems, Inc.

“Agnostic about signalin

21 of 24ervices

What about Voice?

trivial and not

ode sufficient to

ation is within the

router to ensure


Cisco Systems, Inc.

Inside a high-bandwidth cloud, VoIP flows are worth tracking

› Assume configured EF at each network nhandle all calls

› When a call is initiated, check if the destincloud

› If so, just admit the call

› May want to set up a classifer on the edgeno “spoofing”

22 of 24ervices

Voice between (Administratively Same) Clouds

esources, h-rich clouds

red for an EF rate f “busy”

ble - call_bw) >=0?

ed


Cisco Systems, Inc.

Track connections only in the areas of limited rboundaries between clouds. For two bandwidtconnected by a low-bandwidth link

› Assume the low-bandwidth link is configuthat gives a sufficiently low probability o(bw_available)

› When a call is initiated, check destination

› If it’s in the other cloud, check: (bw_availa

› If not, refuse call

› If yes, bw_available -= call_bw and proce

23 of 24ervices

Voice across Clouds

limited resources,

te that sufficient to (bw_available)

l_bw) >=0?

ait for reply

and proceed


Cisco Systems, Inc.

Locally, track connections only in the areas of tail to “next cloud”.

› Assume the tail is configured for an EF rahandle all outside calls most of the time

› When a call is initiated, check destination

› If it’s “not-me”, check: (bw_available - cal

› If not, refuse call

› If yes, signal/message “next cloud” and w

› If reply is positive, bw_available -= call_bw

24 of 24ervices

Pointers

ervices Working erv-charter.html. page.

Van Jacobson ilable at: proceedings from s/may98Workshop/

ce about BBs .lbl.gov/papers/


Cisco Systems, Inc.

• For information on the IETF Differentiated SGroup, see www.ietf.org/html.charters/diffsThe diffserv RFCs are at the bottom of the

• A talk on diffserv and bandwidth brokers bygiven at the Internet2 QoS workshop is avawww.internet2.edu/media/qos8.ram and thethat workshop are at www.internet2.edu/qo9805-Proceedings.pdf

• A “historical” document that still has relevan(Nichols, Jacobson, and Zhang): ftp://ftp.eedsarch.pdf

a tutorial overview with a voice over ip slant · 2007-03-05 · cisco systems, inc. 1 of 24 etsi...

Documents