[OpenAFS] Migrating existing data onto vice partition on the fly

Atro Tossavainen openafs@atrotossavainen.fi
Tue, 30 Dec 2014 19:39:43 +0200


> Same server because I need the transfer to be quick. Once I start - I
> need to do the migration in max 1-2 hrs... tops. I know... I know...

Let's say you have exactly one terabyte of data, 1,000,000 megabytes of
it.  You can stretch the migration time to your indicated maximum, 2
hours, 7200 seconds.

Do the math.

Your disk system would have to be able to read at a constant speed of
140 megabytes/second AND write at the same 140 at the same time.  And
that's not taking into account the network traffic and CPU load that
the AFS processes cause you, just the hardware.

You can easily test whether you can meet your own requirements.

# time dd if=/dev/WHATEVERITIS of=/THATPARTITION/BOGUSFILE bs=1048576 count=16384 

Keep iostat running in another window to get an idea.

That's the simplest it gets; absolutely sequential reads since you are
reading block by block from the physical disk, and writing to one file
only, on the same filesystem.

I could meet your goal of 140 MB/s both ways (and exceed it by a safe
margin; it gave me 237) on an old (circa 2008 I'd say?) Supermicro
storage system running CentOS 6 that has an Areca ARC-1261 controller
with fourteen 1TB disks in a RAID6 configuration.  What's your storage
system like?

The question is merely posed so that you can evaluate it for yourself
whether your requirements for data transfer are realistic given the
hardware you have and the type of operation you intend to perform.

If you have been given 2 hours to transfer 1 terabyte of data out of
and back into a system that does 50 MB/s tops, by the time you need to
go back online you've only completed one third of your estimated
workload.  So it looks like you need to think of something else...

Atro Tossavainen, Chairman of the Board
Infinite Mho Oy, Helsinki, Finland
tel. +358-44-5000 600, http://www.infinitemho.fi/