[OpenAFS] OpenAFS in the ISP environment. A good idea...

Derrick J Brashear shadow@dementia.org
Thu, 6 Mar 2003 11:43:06 -0500 (EST)


On Thu, 6 Mar 2003, Leland J. Steinke wrote:

> > How much data will the volumes have, how big are the partitions, and how
> > much overcommitting do you want to do?
> > 
> > Lease space for the backup volume clones (a complete turnover could cause

"Leave" space. Sorry.

> > a 100% greater disk usage if the quota was used up and then entirely
> > turned over the contents, but that's not worth worrying about)
> 
> Here is the hard part.  Essentially, our customers have a 100M quota.  The 
> average mailbox size is <1M.  If I add in the average usage for personal 
> webspace, it comes to just 1M/user average.  I would not want to do a 100-to-1 
> oversell, but I think 25 or 50 to 1 would not be inappropriate.

Well, you can notice you're running out, and add more.

> I was wondering if the 3500 cell/partition was a software limit.  Your reply 
> seems to indicate that it is more of a "how many do you want fit in the 
> partition" issue.  I will put this another way:

If there's a limit, I don't know what it is. I don't think there is one.

> Assume 100K user volumes, each with 100M quota, and a 50-to-1 overcommit, so the 
> total storage requirement is 200G (this is way in the future).   Assume further 
> that there are two servers (initially) in the AFS cell, each with 100G of RAID5 
> available for AFS.  Could I do a single partition per machine with 50K volumes 
> on each, or should I put more partitions on the RAID array to get the number of 
> volumes/partition down to the 3-5K range?  

Could (...)? yes
Should (...)? debatable. If it's really just one RAID array, it may not
matter, but it might make migration easier depending if the underlying
setup of the RAID unit matches approximately the partitioning to a subset
of the number of disks for each. Your call on how you set it up.

> When does the number of volumes per 
> partition become problematic.  Is it a function (as you imply above) of 
> partition size, volume size (and distribution thereof), and number of volumes in 
> the partition?

It becomes problematic when you need to vacate the disk because it's
dying.

Also, I should note, you can do a parallel salvage on partitions. If said
does not become I/O bound, partitioning instead of "one big chunk" can be
a win. I don't know offhand if it becomes I/O bound.