Skip to content

IPv6-only VPS — has anyone actually run one as their main box?

VPS Hosting by GeorgeNmp 15 replies 1.2K views
#11
rhinodigital said:
Which kernel was openvz 7 running

5.4.203-1-pve on that particular node. But the issue is not the kernel version, it is the OpenVZ kernel patchset. They carry a massive out-of-tree netfilter patch that predates nftables flowtable support for v6. So every packet hits conntrack userspace helper lookup even when no helper is registered.

I tested on Proxmox VE 8 with 6.5 kernel, same hardware, virtio-net: 1.9 Mpps at 14% CPU. The difference is the bridge path, not the hypervisor.

OVHcloud's resolve what? Their DNS64 resolvers are 2001:41d0:3:163::1 and ::2. I have not tested them under load.

virsh list --all | wc -l: 47
#12
prague1983 said:
Which game server has hardcoded v4 bind

Valheim dedicated server. There is a github issue open since 2021, marked wontfix. The binary has no -bind parameter that accepts v6.

I tried the LD_PRELOAD with libipv6nat but it segfaults on Unity's network stack. Socat works for tcp but the steam query protocol over udp is stateful and breaks.

Currently running on a €5 Hetzner cx11 with dual stack in Singapore. Latency to Vietnam is 45ms. Not great. V6-only in Singapore would be 45ms but then I have this v4 bind problem.

Might just pay for a small Vultr instance in Osaka as v4 frontend and tunnel.

phở at 3AM, deploy at 4
#13
Zurich1984 said:
/24 will cost you a kidney by 2028

The broker market is overheated but not irrational. RIPE last /8 policy means you cannot get a new allocation smaller than /22, and only if you are an LIR. End users buy /24 because it is the minimum routable prefix on the global table. Most transit filters reject /25 and longer.

OVHcloud charging €2/month for a single v4 is actually cheap compared to amortized purchase + SWIP fees. A /24 at $48k over 5 years is $800/month. If you need 256 addresses. Most people need one.

NAT64 + v6-only is economically rational for anything that does not need inbound v4. The problem is the long tail of v4-only services. Not the big ones. Google, Cloudflare, Facebook have had v6 for a decade. It is the B2B API from some bank in Germany that breaks.

mitigated 800Gbps before breakfast
3 #14
GeorgeNmp said:
NAT64/DNS64 reliability for legacy IPv4 reachability

To close the loop on my earlier post: the 15 ms relay latency is a red herring. The real email problem is reputation. IPv6-only sending IP has no history. Gmail treats it as neutral, Outlook as slightly negative. You need 30 days of consistent low-volume sending before they stop soft-failing you.

My InterServer relay has 4 years of history. That is worth more than the €2 OVHcloud charges for v4.

For monitoring, I ended up moving Prometheus to v6-only and keeping Alertmanager on dual-stack. Alerts need to reach PagerDuty and they are still v4-only on some endpoints. Same pattern: push from v6, pull to v4.

PMTUD: I did see the 1220 byte MSS issue. Specifically with Stripe's API via Cloudflare. Their v6 path sends ICMPv6 PTB but the NAT64 gateway at Vultr was not relaying ICMPv4 PTB back correctly. Fixed by setting

net.ipv6.tcp_mtu_probing=1
globally. Not ideal but works.

It's always DNS. Always.
#15
danfra said:
A /24 at $48k over 5 years is $800/month

You do not amortize IP addresses like a car. They appreciate. That /24 bought in 2020 for $22k sells today for $48k. It is an asset, not a cost.

The people paying €2/month to OVHcloud are renting. The people buying /24s are holding. Different markets, same squeeze.

I predict we see v4-as-a-service from hyperscalers soon. AWS already does this with Elastic IP pricing that scales nonlinearly. Azure will follow. The hobbyist gets priced out first, then the SME, then everyone except Fortune 500 with legacy compliance requirements.

IPv6-only is not the future. It is the present for anyone who cannot pass through the v4 toll booth.

/24 for sale. No lowballs.
#16
FlowSana said:
Alerts need to reach PagerDuty and they are still v4-only on some endpoints.

Thank you, this is exactly the kind of operational detail I needed. I had considered tcp_mtu_probing but was concerned about the 10-second black hole detection delay on initial SYN. For interactive workloads that is noticeable.

danfra said:
OVHcloud charging €2/month for a single v4 is actually cheap compared to amortized purchase + SWIP fees.

Cogent's v6 peering disputes are legendary. They depeered HE.net in 2014 and still do not peer with Google over v6 in some regions. Any NAT64 gateway using Cogent for v4 but not v6 creates exactly this kind of asymmetry.

I am going to provision a test instance in OVHcloud Gravelines and run tayga + local unbound with DNS64. Will report back with PMTUD findings across a few target ASNs.

Has anyone here actually gotten OVHcloud's v6-only VPS to route to their Singapore datacenter? The pricing page says "Europe only" but their v6 anycast seems to announce globally. I used https://www.ovhcloud.com/en/vps/ to check the plans.

iBGP, eBGP, don't care, just peer

Post a reply

You need an account to reply. Log in or register to join the conversation.

Post reply Preview Save draft