Measuring the Internet’s transition to IPv6Vorlesung: Internet Measurement (SoSe 2016, TUB)

Agenda

IPv4 scarcity

IPv6 adoption

Happy eyeballs (dual-stack hosts)

IPv6 usage in a dual-stack ISP

IPv4 scarcity and IPv6 adoption

Allocations (IANA)

Advertisements (BGP)

Server readiness

Client readiness

IPv4 scarcity: prefix allocations

IANA regional Internet registries (RIRs)

- Download daily snapshots- Query with GNU/Linux whois tool

- Whois 104.244.42.1

Source [12]

IPv4 scarcity: advertised prefixes

Route collectors

- RouteViews project- RIPE

A routing gap: Large amounts of address space unrouted

Source [12]

IPv6 adoption: prefix allocations (II)

IPv6 allocations grow

Source [1]

IPv6 adoption: advertised prefixes

IPv6 advertisements grow faster than IPv4Source [1]

IPv6 adoption: Server readiness (I)

Is a domain reachable over IPv6?

- DNS A and AAAA records- GNU/Linux: host, dig, nslookup

What about popular domains?

- Alexa top 1M

IPv4

IPv6

IPv6 adoption: Server readiness (II)

As of 4. June 2016: 55K of top 1M have IPv6 (http://bgp.he.net/ipv6-progress-report.cgi)

Source [1]

Server readiness: Web site complexity

Can we fetch all elements of a Web page over IPv6?

- 27 % website have some failures- 9 % more than 50 %- 6 % fail completely- Sources: same- and cross-origin

Need fall-back mechanisms for dual-stack networks / hosts

Source [4]

Dual-stack hosts: happy eyeballs (RFC 6555)

Problem: naive implementations reduce applications’ performance

www.example.com AAAA?

www.example.com A?

192.0.2.1

2001:db8::1

TCP SYN, IPv6

TCP SYN, IPv4

TCP SYN+ACK, IPv4

TCP ACK, IPv4

Client Name server 192.0.2.1 2001:db8::1

Several seconds [8]

Dual-stack hosts: happy eyeballs (RFC 6555)

Case 1: Broken IPv6

www.example.com AAAA?

www.example.com A?

192.0.2.1

2001:db8::1

TCP SYN, IPv6

TCP SYN+ACK, IPv4

TCP ACK, IPv4

TCP SYN, IPv4

TCP SYN, IPv6

Client Name server

Don’t wait for IPv6

192.0.2.1 2001:db8::1

Dual-stack hosts: happy eyeballs (RFC 6555)

Case 2: both IPv4 and IPv6 work

TCP SYN, IPv6

TCP SYN+ACK, IPv6

TCP ACK, IPv4

TCP SYN, IPv4

TCP SYN, IPv6

Client Name server

TCP SYN+ACK, IPv4

(skipped DNS)

192.0.2.1 2001:db8::1

TCP RST, IPv4

Use IPv6!

IPv6 adoption: Client readiness

What about traffic at the Internet’s core infrastructure?

https://www.google.com/intl/en/ipv6/statistics.html

IPv6 traffic share: European IXP - 795 ASNs

IPv6 traffic share: European IXP - 795 ASNs

Problem

Mismatch with traffic shares and IPv6 adoption metrics

Possible explanations:

- Lagging edge networks - Bad performance (happy eyeballs)- Users mostly consume content that is only available over IPv4

Investigate IPv6 traffic at a dual-stack ISP

Barriers for IPv6 traffic at a dual-stack ISP

Client or server readiness?

(iii) CPE

(i) OS(ii) applications

Home network Dual-stack ISP Service providersIPv4 traffic

IPv6 traffic

(iv) ISP connectivity (v) service availability

Inte

rnet

Methodology

Step 1: Assess connectivity options of two end points.

- Subscribers: - RADIUS protocol: assign IPs to subscriber

- One IPv4 address: CPE performs NAT on devices’ traffic- One IPv6 prefix (e.g., /64): each device get its own IPv6 address

- Content providers- From active measurements: A and AAAA DNS records- These are hostnames, we have IP addresses!

Step 2: Map traffic to content providers. How?

Methodology (II)

Idea:

- Associate DNS requests issued by an IP address to its network flows [6,9]- Update mapping according to the TTL values (Time-To-Live)

Caveats:

- Large TTLs and short traces - Sometimes browsers do not respect TTL values [10]- IP addresses may serve multiple services

Opportunities:

- Many OS do not suppress AAAA in the absence of global IPv6 connectivity- Caveat: negative caching

See example with trace

Dataset

Trace Total TCPv4 TCPv6 UDPv4 UDPv6

# bytes 64.5TB 80.5% 10.7 % 7.4 % 1.1 %

# flows 356.2M 53.1% 4.7 % 18.2 % 21.7 %

Dual-stack ISP with 12.9K subscribers, 45 h trace (winter 15/16)

IPv6 just 11 %. Why IPv4 instead of IPv6?

Categories of subscribers

IPv4-only: 17.3% of the DSLs

- CPEs do not obtain an IPv6 prefix (e.g., old contracts)

IPv6-inactive: 29.9%:

- No IPv6 traffic:- CPE is not configured / IPv6 disabled- Users only fetch IPv4 content

IPv6-active: 52.9%:

Traffic contribution per category

The actual share of IPv6 traffic (11%) lower than upper bounds for the IPv6 share

Service Subscribers Sum

IPv4-only IPv6-inactive IPv6-active

IPv4-only 5.4% 20.1% 22.5% 47.9%

IPv6-ready 3.2% 9.2% 15.4% 27.8%

IPv6-only 0.0% 0.0% < 0.1 % < 0.1%

Unknown 3.4% 8.8% 12.1 % 24.2%

Sum 11.9% 38.1% 49.8%

How much traffic from IPv6-ready providers is IP4 (and why)?

Subscribers connectivity, devices, and happy eyeballs

Source [8]

How much traffic from IPv4 only services could be IPv6?

DNS negative caching and IPv4-only speaking devices, but

A lot of intent by IPv6-active subscribers

Source [8]

IPv4 or IPv6?: Happy eyeballs

Most happy eyeballs will prefer IPv6

Source [8]

Transition to IPv6: what-if scenarios

Optimistic: legacy devices, happy eyeballs, etc.

Source [8]

Summary

IPv4 scarcity:

- leads to deployment of IPv6 and Carrier-grade NATs (CGNATs)

Measuring IPv6 adoption:

- Prefix allocations (RIRs), routing (BGP), server readiness (DNS), etc.- IPv6 adoption grows- Mismatch between connectivity metrics and traffic

Measuring IPv6 usage:

- Devices, CPEs, ISP, peerings, network performance, etc.

Bibliography[1] Czyz et. al. “Measuring IPv6 Adoption”. SIGCOMM ‘14[2] Richter et. al. “Beyond Counting: New Perspectives on the IPv4 Address Space”. arXiv:1606.00360[3] Naylor et. al. “The Cost of the “S” in HTTPS”. CoNEXT ‘14[4] Bajpai et. al. “Measuring Web Similarity from Dual-Stacked Hosts”. RIPE 72[5] Schinazi. “Apple and IPv6 - Happy Eyeballs”. https://www.ietf.org/mail-archive/web/v6ops/current/msg22455.html[6] Bermúdez et. al. “DNS to the rescue: Discerning content and services in a tangled Web”. IMC ‘12[7] Savolainen et. al. “Experiences of host behavior in broken IPv6 networks”. http://www.ietf.org/proceedings/80/slides/v6ops-12.pdf[8] Pujol et. al. “Understanding the share of IPv6 traffic in a dual-stack ISP”. (unpublished)[9] Plonka et. al. “Context-aware clustering of DNS query traffic”. IMC ‘08[10] Callahan et.al. “On modern DNS behavior and properties”. CCR ‘13[11] Fedora. “Current implementation of AI ADDRCONFIG considered harmful”. https://fedoraproject.org/wiki/QA/Networking/NameResolution/ADDRCONFIG?rd=Networking/NameResolution/ADDRCONFIG[12] Richter et. al. “A primer on IPv4 scarcity”. CCR ‘15

Backup slides

IPv6 adoption - Providers’ view

Source [2]

Methodology: Why DNS?

Option 1: parse HTTP requests:

- GET /bar.png- Host: www.foo.com

Challenges:

- Encryption, e.g., HTTPS, QUIC, etc.- Parsers for other protocols