3 ms·

I can’t see any difference in latency in the services I use between IPv4 and IPv6. What I want to know is do all these studies take into consideration the “hap

I can’t see any difference in latency in the services I use between IPv4 and IPv6. What I want to know is do all these studies take into consideration the “happy eyeballs” algorithm that prefers IPv6 over IPv4 in web browser requests, thus delaying the IPv4 request?

betterunix2 5y ago
The latency in question has to do with routers. IPv6 packets have a simpler header that does not force routers to recompute a checksum for each hop, and IPv6 allows better route aggregation which reduces the CPU overhead of route selection in some cases.
- 2fastpost 5y ago
  I am unaware of any router forwarding IPv4 packets slower than IPv6 packets. Can you back up this claim with any sources? Route aggregation or selection does not impact forwarding performance on a router.
  - betterunix2 5y ago
    A few years old but here you go: https://www.retevia.net/fast/ https://www.retevia.net/fast/ In some cases, yes, the routing table size will impact performance; for example if the FIB is too big to fit in TCAM: https://labs.apnic.net/?p=520 https://labs.apnic.net/?p=520 Note that this particular incident was caused by deaggregation of routes announced by Verizon, and that had the routes been properly aggregated it could have been avoided (though normal routing table growth would have eventually caused the problem anyway).
    - ipv6_or_nat 5y ago
      Oh, come on! Geoff’s article on TCAM exhaustion is almost 10 years old. No BGP router in the default free zone has had a 512k route limit in years. Modern routers typically scale to millions of routes. The incident in question was the result of misconfiguration and/or ISPs trying to run old routers past their usable life. The whole thing was completely avoidable and not related to IPv4 vs IPv6 forwarding performance. As to the IPv6 latency article, it can best be summarized as IPv6 sometimes has lower latency than IPv4, except when it doesn’t. In no shape, way or form does the article claim that IPv6 forwarding performance is better than IPv4 forwarding performance. There’s some hand waving about NAT, but it also notes that increasing the number of NAT levels improved latency! The simplest explanation is that the differences come down to different routing policies in IPv4 and IPv6. Thus it depends on where you stand as to what you see.
  - Dagger2 5y ago
    I've heard this claim too: apparently Facebook tested a lot of routers, and they consistently routed v6 ~500us faster than v4. v6 has no checksum, so routers don't need to recalculate it when they decrement the TTL header, and it does make sense that removing a step would help. Unfortunately my source for this is "I heard it from someone that works at Facebook". In practice this is only a small part of final performance, so you need to do end-to-end measurements to get a proper idea of what people will experience. Both Facebook[1] and Apple[2] have done that, showing 10% faster page loads and 40% faster connection establishment respectively on v6. Even if that was solely down to routing policies, it's still a measurable difference. [1] https://engineering.fb.com/2015/09/14/networking-traffic/ipv6-it-s-time-to-get-on-board/ https://engineering.fb.com/2015/09/14/networking-traffic/ipv... [2] https://www.sidn.nl/en/news-and-blogs/apple-connections-established-40-per-cent-faster-with-ipv6-than-with-ipv4 https://www.sidn.nl/en/news-and-blogs/apple-connections-esta...

betterunix2 5y ago

The latency in question has to do with routers. IPv6 packets have a simpler header that does not force routers to recompute a checksum for each hop, and IPv6 allows better route aggregation which reduces the CPU overhead of route selection in some cases.

2fastpost 5y ago
I am unaware of any router forwarding IPv4 packets slower than IPv6 packets. Can you back up this claim with any sources? Route aggregation or selection does not impact forwarding performance on a router.
- betterunix2 5y ago
  A few years old but here you go: https://www.retevia.net/fast/ https://www.retevia.net/fast/ In some cases, yes, the routing table size will impact performance; for example if the FIB is too big to fit in TCAM: https://labs.apnic.net/?p=520 https://labs.apnic.net/?p=520 Note that this particular incident was caused by deaggregation of routes announced by Verizon, and that had the routes been properly aggregated it could have been avoided (though normal routing table growth would have eventually caused the problem anyway).
  - ipv6_or_nat 5y ago
    Oh, come on! Geoff’s article on TCAM exhaustion is almost 10 years old. No BGP router in the default free zone has had a 512k route limit in years. Modern routers typically scale to millions of routes. The incident in question was the result of misconfiguration and/or ISPs trying to run old routers past their usable life. The whole thing was completely avoidable and not related to IPv4 vs IPv6 forwarding performance. As to the IPv6 latency article, it can best be summarized as IPv6 sometimes has lower latency than IPv4, except when it doesn’t. In no shape, way or form does the article claim that IPv6 forwarding performance is better than IPv4 forwarding performance. There’s some hand waving about NAT, but it also notes that increasing the number of NAT levels improved latency! The simplest explanation is that the differences come down to different routing policies in IPv4 and IPv6. Thus it depends on where you stand as to what you see.
- Dagger2 5y ago
  I've heard this claim too: apparently Facebook tested a lot of routers, and they consistently routed v6 ~500us faster than v4. v6 has no checksum, so routers don't need to recalculate it when they decrement the TTL header, and it does make sense that removing a step would help. Unfortunately my source for this is "I heard it from someone that works at Facebook". In practice this is only a small part of final performance, so you need to do end-to-end measurements to get a proper idea of what people will experience. Both Facebook[1] and Apple[2] have done that, showing 10% faster page loads and 40% faster connection establishment respectively on v6. Even if that was solely down to routing policies, it's still a measurable difference. [1] https://engineering.fb.com/2015/09/14/networking-traffic/ipv6-it-s-time-to-get-on-board/ https://engineering.fb.com/2015/09/14/networking-traffic/ipv... [2] https://www.sidn.nl/en/news-and-blogs/apple-connections-established-40-per-cent-faster-with-ipv6-than-with-ipv4 https://www.sidn.nl/en/news-and-blogs/apple-connections-esta...