Increase Tier1 requirements and limit to 1 per country (or less).
| Task Info (Flyspray) | |
|---|---|
| Opened By | Niklas Edmundsson (nikke) |
| Task ID | 71616 |
| Type | General Gripe |
| Project | Arch Linux |
| Category | Mirrors |
| Version | None |
| OS | All |
| Opened | 2021-07-25 22:02:58 UTC |
| Status | Researching |
| Assignee | Anton Hvornum (Torxed) |
Details
Hi!
I'm one of the mirror admins for ftp.acc.umu.se, and recently I revisited https://wiki.archlinux.org/title/DeveloperWiki:NewMirrors to find that not much seems to have changed since we set up our mirror. I don't see our bandwidth of 20 Gbps listed on https://archlinux.org/mirrors/acc.umu.se/ but perhaps the available bandwidth isn't tracked publicly?
I also looked at https://archlinux.org/mirrors/tier/1/ sorted by country, and there's a number of countries with a lot of tier1:s. I think this kind of defeats the whole purpose of tier1:s, which is to reduce the load on the master server to provide maximum performance (and minimal mirror distribution delays) to tier1:s, and by that minimize the total time to get all mirrors updated.
I would suggest you to increase/tighten the requirements for new Tier 1 mirrors. In the year of 2021, it would be prudent to require at least 10+ gigabits bandwidth, IPv4+IPv6, http/https/rsync support, and the use of a lastupdate-aware sync script to enable updates more often than hourly (say 4 times per hour). Having Tier 1:s update often kills the argument "I want to update from master because all tier 1:s are behind".
I would also like the requirement that master and tier 1:s have applied TCP tuning to allow for high speed long distance transfers. Due to the bandwidth-delay product, the default Linux 4 MiB TCP buffer size limit really limits transfer speeds between continents and 64 MiB is a more suitable limit for 10 gigabit class hosts. There is a lot of documentation on how to do this on the internet, https://wiki.neic.no/wiki/Operations_Tuning_Linux#OS_tcp.2Fnetwork_tuning is one example on how it's done within the Nordic research/HPC community.
Then I would suggest that you start demoting tier1:s in countries with multiple tier1:s that don't make the new requirements (or the old ones, there are currently tier1:s that does not provide rsync or isos). And stick to a limit of one tier1 per country. If tier1:s are good, there shouldn't be a problem for it to handle the tier2:s of that country.
For new Tier 2:s, I would suggest setting a bandwidth requirement of 1 gigabit in countries that already has gigabit-speed mirrors. Users generally aren't helped by more slow mirrors if there already are fast ones.
In general, there should be some wording to discourage servers on home/domestic networks. Even for those with 1 gigabit to the home (like me), the infrastructure of home/domestic internet isn't really built to provide persistent upload performance at that speed, and in the era of work-from-home such servers are likely targets of bandwidth-capping if they interfere with video conferencing and such.
That's it for my general gripe, a very apt description of this I think :-)