Quality of Service and Asynchronous Transfer Mode in IP - - PowerPoint PPT Presentation

quality of service and asynchronous transfer mode in ip
SMART_READER_LITE
LIVE PREVIEW

Quality of Service and Asynchronous Transfer Mode in IP - - PowerPoint PPT Presentation

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Bruce A. Mah bmah@CS.Berkeley.EDU http://http.CS.Berkeley.EDU/~bmah/ The Tenet Group Computer Science Division University of California at Berkeley T Y O I F S


slide-1
SLIDE 1

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 1 of 47

Quality of Service and Asynchronous Transfer Mode in IP Internetworks

Bruce A. Mah bmah@CS.Berkeley.EDU http://http.CS.Berkeley.EDU/~bmah/ The Tenet Group Computer Science Division University of California at Berkeley Berkeley Multimedia and Graphics Seminar 27 November 1996

A
  • T

H E

  • U

N I V E R S I T Y

  • O

F

  • C

A L I F O R N I A

  • 1

8 6 8

  • L
E T T H E R E B E L I G H T
slide-2
SLIDE 2

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 2 of 47

Motivation

ATM growing in popularity, but Internet is ubiquitous and heterogeneous. Want to use these types of networks together efficiently.

R R R FDDI Ethernet Ethernet S S S ATM Network

slide-3
SLIDE 3

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 3 of 47

IP and ATM: Critical Differences

Asynchronous Transfer Mode Internet Protocol Connections Yes No Service Model Quality of service support, can provide performance guarantees Best-effort Packets Small, fixed-size cells Variable-sized packets

slide-4
SLIDE 4

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 4 of 47

Problems to be Solved

ATM QOS in an IP Internetwork

How can IP applications benefit from ATM quality of service support?

Multiplexing

What packets should share a virtual circuit?

Virtual Circuit Management

When should virtual circuits be created and torn down?

Other Issues (not addressed here)

Routing Address Resolution Multicast Support

slide-5
SLIDE 5

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 5 of 47

Results

ATM QOS

QOS support can be helpful, if used carefully.

Multiplexing

Multiplexing eliminates virtual circuit setups, can help application performance.

Virtual Circuit Management

Caching idle virtual circuits can improve both network and application performance.

slide-6
SLIDE 6

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 6 of 47

Outline

☞ Design Alternatives An Internet Simulated ATM Networking Environment Methodology Results and Analysis

ATM Quality of Service Multiplexing Virtual Circuit Management

Conclusions

slide-7
SLIDE 7

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 7 of 47

IP over ATM: A Software View

ATM stack treated by IP as a datalink layer. Device driver puts IP packets in AAL frames, establishes virtual circuits as needed.

ATM Adaptation Layer Ethernet IP UDP TCP Signaling FDDI

slide-8
SLIDE 8

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 8 of 47

Policies and Design Alternatives

ATM Quality of Service

Different service disciplines Preference to different applications

Multiplexing

...per conversation ...per application type per host pair ...per router pair

Virtual Circuit Management

Permanent virtual circuits Switched virtual circuits Switched virtual circuits with caching

slide-9
SLIDE 9

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 9 of 47

IP over ATM Policy Space

Policies to address problems: a three-dimensional space.

Quality of Service Policies Virtual Circuit Management Policies Multiplexing Policies

slide-10
SLIDE 10

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 10 of 47

An Internet Simulated ATM Networking Environment (INSANE)

Functional Requirements

ATM (cell transport, adaptation layer, signalling) TCP features (slowstart, congestion control, fast retransmit, etc.) Synthetic workload with application-specific traffic patterns

Logistical Constraints

Easy to configure for large scenarios (1000+ hosts) Fast (hour-long simulations in reasonable time) Batch processing, off-line analysis

slide-11
SLIDE 11

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 11 of 47

INSANE Protocol Stack

Telnet FTP NNTP HTTP SMTP Audio Video TCP UDP IP User Workload Generator ATM Device Driver LAN Device Driver AAL RCAP LAN ATM

Reliable Cells

slide-12
SLIDE 12

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 12 of 47

Performance Notes

300 simulation runs (4000 seconds each) 1500 hours of CPU time on Sun Ultra 1 cluster Estimated 10 GB raw data

Hardware Platform Running Time Sun Ultra 1 3X Sun Sparcstation 10 11X 100 MHz P5 (FreeBSD 2.1.0-RELEASE) 8X DEC Alpha AXP 3000/400 8X

slide-13
SLIDE 13

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 13 of 47

Environment

R R R R R R S W 1.5 Mbps ATM 30 ms × 200 × 2 100 Mbps LAN

slide-14
SLIDE 14

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 14 of 47

Workload

Application Source Interarrival Time (MM:SS, per site) telnet tcplib 0:10 FTP tcplib 0:15 HTTP empirical 0:05 Audio tcplib 10:00 Video empirical 10:00 SMTP tcplib 0:04 NNTP tcplib 3:45

slide-15
SLIDE 15

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 15 of 47

Performance Metrics

Objective: Measure performance effects visible to applications and users.

telnet Setup and round-trip time FTP File and session response times HTTP File, Web page transfer times Audio Loss rate, one-way end-to-end delay Video Loss rate

slide-16
SLIDE 16

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 16 of 47

Outline

Design Alternatives An Internet Simulated ATM Networking Environment Methodology ☞ Results and Analysis

ATM Quality of Service Multiplexing Virtual Circuit Management

Conclusions

slide-17
SLIDE 17

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 17 of 47

ATM Quality of Service

Schedulers

Best Effort (First-Come-First-Served) Static Priority Rate-Controlled Static Priority Rate-Controlled Static Priority (Rate Jitter Control)

QOS parameters assigned based on application type Results

Static priority can give preference, but starvation a danger Rate control for bulk transfers yields inconclusive results Rate jitter control reduces losses in long TCP bulk transfers

slide-18
SLIDE 18

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 18 of 47

High-Priority Telnet

Interactive performance can be improved.

0.05 0.1 0.15 0.2 0.25 Best Effort High-Priority Telnet Telnet Response Time (seconds)

slide-19
SLIDE 19

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 19 of 47

High-Priority FTP

FTPs given high priority take less time.

1 2 3 4 Best Effort High-Priority FTP FTP File Transfer Time (seconds)

slide-20
SLIDE 20

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 20 of 47

High-Priority FTP

High-priority FTP can degrade others’ performance.

2 4 6 8 10 12 Best Effort High-Priority FTP Audio Loss/Overdue Rate (%) Overdue Loss

slide-21
SLIDE 21

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 21 of 47

High-Priority Applications

Signalling starved by higher priority traffi c.

0.2 0.4 0.6 0.8 1 1.2 Best Effort High-Priority Applications Telnet Connect Time (seconds)

slide-22
SLIDE 22

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 22 of 47

RCSP Rate-Limiting of Bulk Transfers

RCSP can be used to constrain bulk transfers.

0.1 1 10 100 Best Effort FTP sent via RCSP FTP File Transfer Time (seconds)

slide-23
SLIDE 23

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 23 of 47

RCSP Rate-Limiting of Bulk Transfers

Rate-control effects on best-effort applications unclear.

0.05 0.1 0.15 0.2 0.25 0.3 Best Effort FTP sent via RCSP Telnet Round-Trip Time (seconds)

slide-24
SLIDE 24

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 24 of 47

Jitter Control Effects

Bursts smoothed out by delay jitter control, fewer losses.

1 10 100 1000 Best Effort FTP sent via RCSP FTP sent via Jitter Control RCSP FTP Session Time (seconds)

slide-25
SLIDE 25

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 25 of 47

Outline

Design Alternatives An Internet Simulated ATM Networking Environment Evaluation Results and Analysis

ATM Quality of Service ☞ Multiplexing Virtual Circuit Management

Conclusions

slide-26
SLIDE 26

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 26 of 47

Multiplexing Policies

Different levels of traffi c aggr egation

Virtual circuit per conversation (e.g. TCP connection) Virtual circuit per application per end-host pair Virtual circuit per router pair

Results

Multiplexing eliminates virtual circuit setups Interaction with policing degrades performance of long FTP, Web transfers Buffer contention causes losses with per-router pair multiplexing

slide-27
SLIDE 27

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 27 of 47

Aggregation Improves Transfer Times

Reduced need for virtual circuit setups shortens fi le transfers.

0.5 1 1.5 2 Conversation Multiplexing Application Multiplexing FTP File Transfer Time (seconds)

slide-28
SLIDE 28

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 28 of 47

Per-Application Interference

Previous gains disappeared or reversed themselves.

5 10 15 20 25 30 Conversation Multiplexing Application Multiplexing FTP File Transfer Time (seconds)

slide-29
SLIDE 29

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 29 of 47

Per-Application Interference

Rate control policing causes cells from one IP conversation to delay cells from another, even if not simultaneous. Per-application multiplexing Per-conversation multiplexing

In Out In Out

slide-30
SLIDE 30

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 30 of 47

Per-Router Multiplexing

Loss rate multiplied by 3–8, due to buffer contention.

1 2 3 4 5 6 7 Conversation Multiplexing Router-Pair Multiplexing Audio Loss/Overdue Rate (%) Overdue Loss

slide-31
SLIDE 31

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 31 of 47

Outline

Design Alternatives Evaluation An Internet Simulated ATM Networking Environment Results and Analysis

ATM Quality of Service Multiplexing ☞ Virtual Circuit Management

Conclusions

slide-32
SLIDE 32

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 32 of 47

Virtual Circuit Management Policies

Different policies for managing virtual circuits

Permanent virtual circuits (set up at network start) Switched virtual circuits (set up on demand, torn down when idle 10 seconds) Switched virtual circuits (set up on demand, cache when idle 10 seconds)

Results

Caching idle virtual circuits can eliminate about 95% of setups Signifi cant speedups for applications doing r epeated TCP connections

slide-33
SLIDE 33

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 33 of 47

Virtual Circuit Cache

Establish virtual circuit when needed After 10 seconds idle

Router “unbinds” virtual circuit from conversation(s) Available for reuse Match destination and QOS parameters

After 300 seconds (5 minutes) idle

Tear down virtual circuit

slide-34
SLIDE 34

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 34 of 47

Cache Statistics

Hit rates

For best-effort traffi c 97–98% For most other schemes 92–95% Worst case 56–69%

Signalling Rates (setups per second, best-effort traffi c) Per-router overhead: 20–40% more virtual circuits open.

No caching Caching Per-application 3.0 0.06 Per-conversation 7.0 0.18

slide-35
SLIDE 35

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 35 of 47

Connection Caching and HTTP

Cache has a benefi cial impact on application performance.

5 10 15 SVC SVC with Caching HTTP Page Retrieval Time (seconds)

slide-36
SLIDE 36

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 36 of 47

Summary of Results

ATM Quality of Service

Static priority effective for giving preference to any application, but starvation a danger Overall effects of rate control inconclusive Delay jitter control reduces losses for long bulk transfers

Multiplexing (traffi c aggr egation)

Multiplexing shortens small transfers by eliminating VC setups Rate control effects cause long transfers to interfere with each other Contention for buffering increases loss

Virtual Circuit Management

Caching unused connections eliminates vast majority of VC setups Improves application performance

slide-37
SLIDE 37

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 37 of 47

My Related Work

Network Traffi c Measur ement and Models

World Wide Web Internet Video IP Multicast

The Tenet Real-Time Protocol Suite

First set of network protocols to give performance guarantees over a packet-switched internetwork

slide-38
SLIDE 38

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 38 of 47

Conclusions

ATM quality of service can help Internet applications, if used carefully. Multiplexing of traffi c can help applications by r educing virtual circuit setups. Virtual circuit caching improves application and network performance.

slide-39
SLIDE 39

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 39 of 47

END OF TALK

(remaining slides used for answering questions)

slide-40
SLIDE 40

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 40 of 47

Real-Time Guarantees

At connection setup time...

“User” provides traffi c specifi cation, performance r equirements Network performs admission control tests, returns “accept” or “deny”

During data transmission...

Queues in network queue according to a particular service discipline Best-effort traffi c fi lls “gaps” between cells for guaranteed connections

slide-41
SLIDE 41

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 41 of 47

Rate-Controlled Static Priority

Per-Connection Rate Controllers Priority Queues Input Output

slide-42
SLIDE 42

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 42 of 47

Static Priority Scheduling

Level Type of Traffi c 1 Telnet 2 FTP control, Audio 3 FTP Data, HTTP, Video Signalling Best-effort

slide-43
SLIDE 43

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 43 of 47

Work-Conserving RCSP

Application Dir Local Delay (ms) Peak Rate (Kbps) Average Rate (Kbps) Avg. Interval (ms) telnet up 10 19.2 7.68 500 down 10 38.4 19.2 1000 FTP (control) up 20 3.84 0.77 5000 down 20 3.84 0.77 5000 FTP (data) up 80 3.84 3.84 2000 down 80 19.2 19.2 2000 HTTP up 40 7.68 3.84 10000 down 40 96.0 48.0 1000 audio any 20 96.0 76.8 100 video any 50 148 110 2000

slide-44
SLIDE 44

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 44 of 47

RCSP Rate-Limiting of Bulk Transfers

In most cases, no statistically signifi cant ef fects.

2 4 6 8 10 12 Best Effort FTP sent via RCSP Audio Loss/Overdue Rate (%) Overdue Loss

slide-45
SLIDE 45

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 45 of 47

ATM Switch Composite Object

Composite objects created by invoking Tcl scripts

SwitchFifoN switch 8 1024 1000000 Sig GoBackN SwitchModule CellQueueFifo CellQueueFifo CellQueueFifo CellQueueFifo CellInputPort CellInputPort CellInputPort CellInputPort

slide-46
SLIDE 46

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 46 of 47

Interaction with Objects

Command line interface like debugger Interact with objects and examine state

] router1.ip netstat -r Destination Mask Gateway Flags 128.32.150.0 255.255.255.0 128.32.150.254 128.32.131.0 255.255.255.0 128.32.131.254 127.0.0.1 255.0.0.0 127.0.0.1 ] router1.ip netstat -i Name Address Netmask Ipkts Ierrs Opkts Oerrs router1.lan1 128.32.150.254 255.255.255.0 0 0 1 0 router1.lan0 128.32.131.254 255.255.255.0 1 0 0 0 router1.lo0 127.0.0.1 255.0.0.0 0 0 0 0

slide-47
SLIDE 47

Quality of Service and Asynchronous Transfer Mode in IP Internetworks Last Change: November 20, 1996 Page 47 of 47

“ State of the Art”

XUNET II

Best-effort Per-router-pair multiplexing Permanent virtual circuits

FORE Systems ATM LAN

Best-effort Per-router-pair multiplexing Permanent or switched virtual circuits (no caching)

Ipsilon IP Switching

Best-effort (RSVP support in future) Various multiplexing policies Permanent or switched virtual circuits (not end-to-end, no caching)