Hi, Hal-- On Jun 28, 2016, at 8:41 PM, Hal Murray <[email protected]> wrote: > [email protected] said: >> If a machine has both IPv4 and IPv6 networking available, then prioritizing >> IPv4 or 6-in-4 over IPv6 is very likely to improve timekeeping performance. >> >> (I've seen a few counterexamples where IPv6 worked better, but that is >> rare.) > > Could you please say more? What sort of connectivity to the clients have > and/or what servers are they using?
Sure. I have two major sources of data, one internal from managing a 4-digit number of machines (well, mostly VMs but some bare metal) which are located in various data centers which commonly have 10 GB ethernet to the blade chassis. There are multiple 10 GB links channel-bonded together via LACP for inter-building or intra-datacenter links, and I believe OC-48 and OC-192 links for WAN connectivity between datacenters located in different geographical regions. The clients are mostly dual-stacked Linux systems and especially the OEL 5.x systems have problems with heavy IPv6 traffic (meaning hundreds to thousands of connections per second). Most of that is service traffic, not NTP, but having the network stack start to choke under high load affects NTP traffic along with everything else. [1] The other major source of data comes from the various CDNs which Apple uses, such as Akamai's EdgeKey/EdgeSuite and NetStorage, or ChinaCache in China, or Apple's own CDN. The details are proprietary, but suffice it to say that I've had pretty extensive insight into IPv4 versus IPv6 traffic such as network latency, packet loss and retransmissions, etc...broken down by country if I want to view the data that way, for ~40 countries worldwide. There are some places in Europe where IPv6 works about as well as IPv4 does-- initially parts of France, but now Germany, Belgium, and perhaps one or two others. Likewise, a few places in APAC like Taiwan and parts of Japan seem to have IPv6 work as well as IPv4 at least some of the time. > Is the problem that there aren't many good IPv6 ntp servers? Admittedly, the CDN stuff is mostly HTTP and not NTP, but one can also draw reasonable conclusions from the fact that these queries return IPv4 A records: dig -t a time.apple.com. dig -t a time.windows.com. ...whereas the following queries do not return IPv6 AAAA records: dig -t aaaa time.apple.com. dig -t aaaa time.windows.com. Regards, -- -Chuck [1]: When running on MacOS X on XServes some years ago, I didn't need to prefer IPv4 over IPv6, but I do with Linux-- or else make launch arg changes like adding -Djava.net.preferIPv4Stack=true for Java-based stuff. _______________________________________________ pool mailing list [email protected] http://lists.ntp.org/listinfo/pool
