Top Banner
QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University http://www.ece.rice.edu/~knightly (and without bankruptcy along the way)
19

QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Jan 16, 2016

Download

Documents

Eunice Elliott
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

QoS: Where, When, and Why?

Ed Knightly

ECE/CS Departments

Rice University

http://www.ece.rice.edu/~knightly

(and without bankruptcy along the way)

Page 2: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

The Big Questions

• When do clients/customers want QoS?– Always! – No one wants unpredictable, poor performance

• How to deliver QoS?– Capacity planning?– Admission control (flow or aggregates)?– Adaptive algorithms and applications?– Marking/scheduling/AQM packet differentiation?

• Are we there yet?– NO! End-to-End service is lousy

Even the undergrads won’t use VoIP Video is useless match-box sized

– Applications can adapt, but users will not!

Page 3: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Answers

Required mechanisms depend on

– QoS stringency: voice, live streaming media, …

– Traffic: predictability, aggregation level, …

– Economics: how can provider make $$ (or lose less)

– Deployability: ten-year scale for layer 3, six-months for layer 2/7

Case studies: Core, Enterprise, Metro Edge, Internet Data Centers, Wireless access

Page 4: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

I. Why QoS in NOT Needed In Today’s Core

1. Traffic is highly predictable and not bursty in the core

Operate here? Even QoS cannot save you

Page 5: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Why QoS is NOT Needed in Today’s Core

1. Traffic is highly predictable and not bursty in the core because users are capped by– Abundance of low-rate access links

Ex. OC-3, T3, T1, DSL, modems, …– Abundance of rate limiters

Ex. Throttle dorm traffic; reduce ISP bill

2. There is no “production” real-time traffic– No VoIP, high-quality video, …

Page 6: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

With the Current Evolutionary Path

• Capacity planning is simple and sufficient – Traffic prediction = Time-of-day prediction (easy!)

• There are no research issues in the core – QoS or otherwise (why worry about ECN, AQM, multicast, …)

• More bankruptcies and junk bonds to come

Page 7: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Disruptive Changes for the Future of the Core

• Economics? – Need to re-think;real costs to over-provisioning in line cards

• Integration of voice/video/new apps? – It’s happening at the edge

• Burstable high-speed packet access? – 1 and 10 Gigabit Ethernet access

Clients will burst, the traffic will change Contrast to today’s relatively low speed circuit access

Page 8: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

II. The Network Edge (Myth)

• Common, but unrealistic view– Hosts, network of switches (or routers), edge Internet

router

Internet B ac kb o neR o uter

S w itc hH o s t

...

Page 9: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

The Network Edge (Reality)

• Ring metro backbone– Rings are the dominant configuration for their fault tolerant

properties– Size ~ 100 nodes, ~ 10 km

InternetB ac kb o neM etro B ackbo ne

Page 10: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

SONET: the Dominant Metro Ring Technology

• Circuits between pairs of nodes. Problems:– Cannot re-use unused capacity (no statistical multiplexing)– Cannot burst to full link rate– Coarse bandwidth granularity (155 Mbps)– Up to N2 circuits

Inefficient!

InternetB ac kb o ne

Page 11: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Emerging Approach: Packet Rings (GigE and RPR)

• Advantages– Statistical multiplexing, burstable access, efficiency

• Example Metro Provider: Phonoscope – 2 GigE rings spanning Houston with 400+ customers– 400 Mbps of data– 50 Mbps VoIP– 800 Mbps of streaming MPEG

InternetB ac kb o ne

Page 12: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Challenges in Packet Rings

• Fairness– Closest node to the gateway gets the most bandwidth– Extent of unfairness depends on protocol (TCP/UDP), topology (RTT and

number of nodes) and traffic inputs– Goal: inter-node performance isolation

• Guaranteed QoS– Voice/Video require performance isolation (without circuits)

InternetB ac kb o ne

Page 13: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Research Issues in Packet Rings

• Distributed priority scheduling and MAC– How to throttle at ingress, use simple transit path scheduling,

and achieve ring-wide QoS/fairness goals?

– Hot topic in ad hoc networking, equally challenging (and more relevant?) in packet rings

• Admission control, faster-time-scale capacity planning– Challenge in low-aggregation regime with (true) bursty traffic

• Potential impact is high– Aggressive and innovative edge service providers– Significant enterprise market– Fast innovation/deployment time at layer 2

Page 14: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

III. The Edge of the Network Edge:Internet Data Centers

• Many hosted sites at a shared physical location

– Share network bandwidth

– Share management facilities

Page 15: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Problem with Today’s IDC’s

• Static Resource Model – Number of servers per

website is fixed

• Performance losses– Peak allocation poor

resource utilization– Mean allocation

performance degradation– Inability to deal with flash

crowds

• Result: inefficiency, poor QoS, bankruptcy

E-news.com

E-mart.com

Workload

Workload

Un-utilized servers

Page 16: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

QoS Research Problems in IDC’s: Infrastructure-on-Demand

• Migrate servers to hot spots

• Challenge: how to exploit efficiencies of resource sharing while providing QoS?

– When to migrate– By how many– Load prediction– SLA provisioning

Workload

Workload

Un-utilized servers

Server Migration

E-news.com

E-mart.com

Page 17: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

IV. QoS in the Other Edge: Wireless

• Yes, WiFi is in your future

• Applications: VoIP + multimedia + web + …

• Constraint: random access MAC’s are essential for low cost and simplicity

Page 18: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

QoS Challenges at the Wireless Edge

• Low aggregation regime with high burstiness: capacity planning won’t solve it

• Challenging QoS-aware MAC design problems– Unsolved under realistic channel models

exploit multi-rate physical layer directional antennas power aware

• Admission control/bandwidth allocation– Requires modeling/prediction of a complex system

(random access, multi-priority MAC, wireless channel, …)

Page 19: QoS: Where, When, and Why? Ed Knightly ECE/CS Departments Rice University knightly (and without bankruptcy along the way)

Ed Knightly

Summary

• QoS in the core is not needed in evolutionary path– Watch out: revolutions are possible in search of profit

• Challenging hot spots– Metro edge, IDC, wireless edge, Enterprise

• Why? – Traffic less predictable, not highly aggregated, not rate

limited– Can be more aggressive/innovative at layer 2/7