🚀 go-pugleaf

ian@ilm.mech.unsw.edu.au (Ian Maclaine-cross) writes:

>>Hmmm .. bet you it costs a fair bit more to go beyond 4 x 3.5 Gb disks :)
>>(At least with scsi you can get up to 7 devices (excluding host) per bus and
>>have multiple busses)

>If you want a 30 GB mirror, consider two Linux Pentium boxes each with
>4 x 4 GB disks. Link them with NFS and the client will never know the
>difference. About $6000. Maybe a really large busy site like
>ftp.cdrom.com needs SCSI but if your link to the net is only 10 MHz
>why bother?

Ah wise move ... but .. not so clever in reality.
"Why?" you ask?
Well ... let me give you a view on some experience.
ftp.dstc.edu.au had/has a 10Mbit link to the world. Not bad, certainly
a good deal better than most.
It's 10Mbit link was choked, there were soo many packets being sent etc, that
effectively it was choking the rest of the connection for the org.

Now .. multiply this buy three for two machines connected via nfs.
(one for the packet going to the ftp'er and one for the nfs requst to the
other machine, one for the reply from the other machine.
This is not a problem on a light network .. but with the amount of traffic
we had .. it would have killed it out right. We had enough problems with
local disk.
My advice ... pay the extra grand, and get a crummy scsi card and scsi disks.
anyways .. this is all academic :)
Disk I/O is not the main factor (thou it does come into it a bit in some parts)

>>>We do not backup CTAN, it is a mirror not an archive.  Index and ls-lR
>>>files are locally generated after the ls-lR.Z file is downloaded from
>>>ftp.tex.ac.uk. We use a similar technique with our partial mirror of
>>>Debian. 
>>
>>Ah ... we didn't back ftp.dstc.edu.au for a long while ... and then we had
>>a disk failure(s). It took 3 months to get back upto date with the mirror.

>CTAN is about 2.3 MB and took a month to download from the UK. It
>downloaded the most recent files first. 

>We need multiple *accurate* mirrors in Australia of important overseas
>sites. If one crashes it may be only days to bring it up again by
>mirroring a local site.

This is the way it "should be" ... in practice, ftp.dstc.edu.au was mirrored
by a lot of uther linux mirrors. Mainly because .. it was so hard getting
stuff (and a connection) from sunsite.unc.edu, It is expensive to get stuff
from the USA (Uni's don't have much choice on ISP's :), it is one heck of
a lot faster to mirror from someone who already has a full(ish) copy
local.

Mirroring works really well when you have a fast(ish) link to the "real" copy.
What I am trying to point out is, it was hell mirroring to sunsite.unc.edu
(the problem was the 2Mb link it had to the outside world).
I don't know if it has changed .. as I have not cared about it for a while
and once the mirror is close to up-to-date then it is easy.

>>I strongly suggest to anyone they take atleast weekly full backups of there
>>linux mirror. It is way too slow re-get the entire linux mirror.

>If you are running an archive (or source site) backups are
>essential. The labour and cost of tape backups is prohibitive for a
>mirror. 

>Much better to share the load by having a nearby secondary mirror paid
>for and run by someone else. If your site crashes, you backload from
>them first.

Hmmm ... tape is cheap.

>The idea of mirroring is to save communication costs. It is done
>when the net is idle so its marginal cost is nothing. If you are
>paying through the nose for this you need a new service providor.

Hmmm .. it is easy to say .. its marginal when it is not your pocket it
is coming from :)
--
 _--_|\  David Conran,                       Voice: +61-7-3365-3254
/      * Systems Administrator               Fax:   +61-7-3365-1477
\_.--._/ Department of Mathematics,          Email: lucifer@maths.uq.edu.au
      v  The University of Queensland, 4072 "Reach out and grep someone"