[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Debian ISOs



[sry, I am coming late and I am breaking all references as I did not received this thread. Worse, I am answering to different mails at once ;-)]

I will start to introduce myself, François Pétillon, working for a small french ISP (Free) and managing one of the debian CD mirror (ftp.free.fr/ftp.proxad.net).

From: "Anthony L. Bryan" <albryan@comcast.net>
> Metalinks, a cross platform vendor neutral fortmat, are used by
> download managers & contain Mirror & p2p locations for segmented
> downloads, along with automatic checksum verification when the
> download completes. It spreads the download between multiple servers
> so its faster for users, more reliable, & less load on any one server.

No, that's bullshit. You are just telling that when x users download one DVD (around 5 GB), doing it with n connections will be more server-friendly (x*n connections to manage for all servers) than doing it with a single one (x connections).
A server load depends on several issues :
- CPU (the more you have to manage connections, the more CPU you waste)
- memory (each connection is using a few KB/MB)
- disk IOs

And, as far as I am concerned, disk IO is the greatest problem (disk capacity & bandwidth is increasing faster than seek time). Thus, we try to optimize disk IO to get max performance out of that kind of server. And optimization usually means trying to read the greatest amount of sectors for each disk seek.

But having to manage more connections means less memory per connection (thus reading less data per disk seek). And segmented downloads also mean you will not be able to optimize all disk access (I already have seen FTP clients requesting a DVD per 8KB segment, thus all disks request were around 8KB while the server would try to load a few MB at a time when feasible).

This for these reasons I prefer to see someone to download a full CD image rather than using jigdo if he has to download many packages one by one.

I do not have anything against Metalinks by itself, it is just that people _will_ make stupid things with it and they _will_ degrade servers performances.

From: Subredu Manuel <diablo@iasi.roedu.net>
> Let's see now. What does metalink does ? Permit the use of _all
> mirrors_ and better, permit the average Joe to download the images
> faster.

Wait a minute. Are you telling us your tool is creating bandwidth ? If someone is download at a slow rate, there must be a good reason. No ? The server or the network may be loaded. Fine, do your tool solve this problem ? No, it just steal others users bandwidth/server ressources. When everyone will have to use that kind of tool, then you'll be back to the start point but with degraded performances.

From: Josselin Mouette <joss@debian.org>
> Given that downloads like Debian ISOs are already putting a heavy
> bandwidth load on the servers and that they are already shared among
> many servers, I don't think it is a good idea to encourage users to
> load several servers at once with one download. We should instead push
> bittorrent as the main distribution media for ISOs.

Well, with a network traffic around 120-140 GB a day (less than 15 Mbps, cf ftp://ftp.free.fr/stats/debiancd.weekly.20060923.txt), I can hardly
consider this mirror as loaded... :-)
Do not forget that network is not free either and P2P may cost more for an ISP than hosting mirrors (for Free case, 30% of users are connected through FT IPADSL for which bandwidth cost is around 220 EU per Mbps per month, 10 times the transit cost, more than 300 times the hardware cost to output 1 Mbps/month on a FTP server).

François



Reply to: