Next-generation Internet, an open source projectgdt/presentations/2000-11-25-aoss-ngi/aoss.…• QoS requires authentication to limit ... – Put into worst queue – Drop packets

Next-generation Internet, an open source project

Glen [email protected]

Australian Academic & Research Network

Australian Open Source Symposium2000-11-25

UniSA, Adelaide

http://www.aarnet.edu.au/

aarnet

Internet protocol development,

an open source project

Part A

aarnet

Internet Engineering Task Force

• Protocol development– Internet Architecture Board sets

framework

• A strong academic heritage– Openness– Correctness

aarnet

Open source

• RFCs are freely available– IEEE's ethernet standard costs $2,500pa

• RFCs are the IETF's "publications of record"

• Not all RFCs are standards– "Standards track" versus

"Informational"– Marketing may not recognise this :-)

aarnet

Open source

• Anyone can participate– Participation in Working Group mail lists

is open to all– Participation at meetings is by payment

to cover secretariat costs– Volunteer meeting organiser

participation is free– Elections of office bearers occur at

meetings

aarnet

Open source

• Anyone can initiate– Write an Internet Draft and mail it to

[email protected]– IEEE requires a project sponsor, and

that sponsoring organisation incurs significant costs

aarnet

Open source

• Conflict is allowed– The market will decide between IFC and

iSCSI protocols for accessing remote storage

– Propietary protocols can be documented as Informational RFCs• NFS, TACACS• But no requirement to document future

versions of the protocol (TACACS+)

aarnet

Rough concensus

• RFCs and Drafts are not approved by vote, but by a consenus– "rough": some objectors might be

ignored– Versus "absolute": ISO TP1

aarnet

Running code

• Running code is the final arbiter of technical disputes– "show me the source"– Versus OSI's committee politics

• For an RFC to become a standard, two distinct interoperable implementations have to be shown

aarnet

Process failures

• CIFS– can't be implemented from poor-quality

RFC

• Kerberos extensions– Interoperabilty denied

• Vanity publishing

aarnet

Operating systems

• BSD Unix is most used protocol development platform– Networking code is best– License

• Linux is being used more often– License– Application availability– Corporate approval

aarnet

IETF projects and their consequences for open

source coders

Part B

aarnet

Quality of service

Part BTopic 1

aarnet

Differentiated Services

• Differentiated Services Code Point– A number, 0 to 63, in the IP packet

header– Determines the network's treatment of

the packet• Eg: news might have DSCP=1 and be

routed over slow path• Eg: VoIP might have DSCP=40 and have

priority

aarnet

Assigning DSCPs

• DSCP is assigned by ISP– At customer-ISP network boundary

• "classification"

– DSCPs may be mapped at ISP-ISP boundaries• "policing"

aarnet

Classification must be cheap

– Use well-known port number• This is what is technically wrong with

Napster's use of the Internet– Can't assign Napster a 'worst-effort' forwarding

behaviour

– Use IP addresses• To build different network performances for

different classes of users– Gold, Silver, Bronze products

aarnet

Implications

• Applications must expect to receive a different DSCP then they send

• Applications can be designed for networks that are not best effort– Napster & USENET can use idle

bandwidth– Voice can be queued before everything

else for low-latency

aarnet

Authentication

• QoS requires authentication to limit network misuse– Imagine a DoS attack that uses the

voice DSCP

aarnet

Unreachability and performance

– Currently per-destination– Becoming

• Per-user– Due to authentication

• Per application– Due to QoS

aarnet

Ping

• ICMP Echo and Echo Reply– Was designed to 'indicate' reachability

• Some applications use it to monitor reachability

• Other applications use it to launch denial of service attacks

aarnet

Ping

• So classify 'unnecessary' ICMP– Assign it a DSCP

• Configure per-hop behaviour to limit DoS impact– Put into worst queue– Drop packets if input rate is >0.1Mbps

aarnet

Ping

• The best networks now appear to have the worst performance

• Applications that want to measure reachability need to do this in their application protocol– Don't do this until after authentication

aarnet

Internet Protocolversion six

Part BTopic 2

aarnet

IP version 6

• An extended socket interface– Standard across all platforms

• Address format– The standard textual format is complex

• a0cd:345s:0::0:34:128.22.232.3

– Can't determine if a string is IPv6 address using a regular expression (whoops)

aarnet

IP version 6

• Addresses can be renumbered on the fly– Interfaces hold multiple addresses whilst

this is happening– Applications shouldn't know about IP

addresses, use DNS names• FTP

– 'Authentication' by DNS name pattern needs to for any IP address from the host

aarnet

Internet Protocolsecurity

Part BTopic 3

aarnet

IPsec

• End-to-end authentication of IP packets

• Modification of the packet causes it to be rejected

• So network address translation, or IP masquerading, is a problem

aarnet

Latency

Part BTopic 4

aarnet

Bandwidth is not longer a technical problem

• 100Mbps ethernet is cheap• 1Gbps is a retail product• 10Gbps is being standardised• 40Gbps is in the lab

aarnet

Latency can't be changed

• Speed of light in fibre or copper• 70ms one-way to the USA

aarnet

Buffering

• You can have 1GB in transit waiting to be acknowleged

• Kernel versus user space memory allocation– Gigabit leads to big spikes in kernel

memory usage, not a gradual allocation of small amount of memory

aarnet

Protocol design

• Avoid me-you-me-you transactions– Session establishment

• Authentication• Negotiations

– Database cursors become less efficient than SQL

• Use big counters– Can easily have 216 packets in transit

aarnet

Redesigning HTTP

• Use one TCP• Push images• No redirects• Automatic distributed caching• Next page links become unfriendly

aarnet

Congestion control

Part BTopic 4

aarnet

Multimedia

• Congestion meltdown has happened in the past– Which is why TCP backs off upon

congestion

• Some real multimedia protocols don't avoid congestion

aarnet

Congestion manager

• Path congestion information needs to be shared between transport layers

• Some applications can reduce their data rates– Chunkier video

• Congestion manager in kernel– Extension to sockets API

aarnet

Denial of service

Part BTopic 5

aarnet

Denial of service

• Not addressed well in protocol design

• Too much done before authentication– Move towards authentication to get

access to link layer• COPS• PPP isn't adequate for multi-user machines

aarnet

For unauthenticated transactions

• Send a smaller packet then you receive– Poor eg: DNS query for aol.com

• Examine source addresses for reasonableness

aarnet

Distributed attacks

• AARNet monitoring shows huge numbers of hacked machines– Mainly Linux machines

• downside of Linux being successful for Internet server applications

• Can only be defeated by quality system administration

aarnet

Quality of system administation

• Not harder, smarter• "Enterprise" sysadms are familiar

with the solutions• Standard software configuration• Automated software updates• Strong authentication & cyptography• Centralised user management• Linux does this better than Windows, but

still worse then NetWare

aarnet

ISP actions

• QoS– Limit suspect traffic– Sell only 10Mbps throughput to the

average retail customer

• Intrusion detection and removal– Easier with fiber

• Passive optical fiber tap• Inject data to cause CRC errors in incoming

traffic

aarnet

Life critical applications

Part BTopic 6

aarnet

Don't

• But these are surprisingly common– eg: weather

• Possiblilty of a hack causing death– AARNet has had people take control of

cranes and NMR scanners

aarnet

If you have to

• Use end-to-end crypto• If you are hacked you can still trust your

data• Key management needs to ensure this• Lots of turnkey systems don't do crypto!

• Lock the box down• Use a VPN

aarnet

Sysadm v lusers

• IT areas try to control too much• Unfriendliness is inevitable

• Not everyone's user name can be "fred"

• But one unmaintained machine...– Insist that all machines have sysadms

aarnet

Closing comments

Part BTopic 7

aarnet

Don't invent a protocol by accident

• struct my_t m;send(m);

• No interoperatbility– Between versions– Between CPU architectures

• Sizes, byte order, padding

aarnet

Do it right

• Representations– IETF text or XML– XDR or ASN.1– TLV: tag, length, value

aarnet

Do it right

• Don't talk to unknown people– Be paranoid

• Connection sources• Overflows

– Allow binding to loopback interface only– Log weirdness

aarnet

Do it right

• Keep statistics• Document what they count

– Does "packets in" include bad packets?

aarnet

Protocols happen in unexpected places

• Disk formats– ext2

• Process-daemon communication– Good: Zebra– Bad: lpd

aarnet

Global storage

Extra

aarnet

ISPs want to do more

• "Value added" business strategy– ASP

• Relevance of a new Microsoft revenue model isn't clear for free software users :-)

– Eight layer• Helpdesk application support• Programming, web design and consulting

– Servers• Web, e-mail• Storage

aarnet

Storage

• I back up my laptop– It has all my work for the past ten

years

• This is "expensive"– Multiple DAT-2 tapes– $700 worth of second-hand hardware

• Want to use a shared tape drive– A network-connected tape library

aarnet

Network storage

• Network file systems don't cut it for all applications– Differing operating systems– Differing file system formats– Need to restore with full meta-data

aarnet

iSCSI

• Small Computer Systems Interface– Tape, CD, disk, CD-W

• Over TCP/IP

aarnet

Design issues

• SCSI allows multiple outstanding requests– So losing a packet will stall all requests– But TCP is a stream protocol, so it will

wait for the lost packet to be retransmitted before handing on the command

aarnet

Design issues

• TCP design tops out at about 10Gbps– This is only about 1GBps

• SCSI layer in Linux confused– Intercepting SCSI protocol data units is

hard– Need to emulate control signals of a

SCSI parallel interface