At ALS beamlines 8.3.1 and 12.3.1 we use a combination of DVD-R and
LTO-4 tapes for long-term backup, and have the entire data collection
history of each beamline backed up on DVD-R disks. This is at about 50
TB for 8.3.1 (built in 2001) and 30 TB for 12.3.1 (built in 2004). We
also make a DVD of the user's data automatically and near-real-time
using a ~$4k robot that inkjet prints the user's name and dataset
summary onto each disk. Portable hard disk drives for "sneakernet" are
also popular, but so is transferring the data over the internet, which
can also be done in near-real time.
I started using LTO-4 tapes recently for two reasons: 1) the price per
TB became competitive with DVD-R and the tape drive is only ~$4k. 2) I
used to keep two copies of each DVD, but found this was not really
"redundant" because if you write two DVD's one after the other on the
same day with the same writer using media from the same batch, then if
you can't read one of these disks 4 years later, the chances of not
being able to read the other disk are pretty high.
So, a lesson I learned is to store data on two very different media
types so you get "orthogonal" failure modes.
I can also tell you that it is a good idea to erase your LTO tapes 2-3
times before writing any data to them. I think this is because the
primary source of error on these tapes is the roughness of the edge of
the tape itself (which is used for alignment) and running it back and
forth a few times probably wears/folds down any big bumps. Sounds
strange, but I had some tapes I initially thought had "bad spots" on
them, but upon erasing and re-writing the data to them again, the "bad
spots" are now gone, and have remained gone each time I have checked
those tapes over the last year. Subsequent tapes that I have erased 3x
before use have never had "bad spots". Also, you need to write data to
them at a minimum of 80 MB/s, or you can actually have problems reading
back the tape. I do my writes in 2 GB chunks from the system RAM.
ALWAYS test reading back the tape. Preferably more than once.
DVD-R media should also be verified and preferably in a low-quality DVD
drive. This is because writers tend to have much higher quality than
average drive mechanisms and I have seen many DVDs that read back just
fine in the drive that wrote them, but throw all kinds of media errors
when you take them home to a dusty old DVD reader.
As for getting the PDB to do image backup for us, I don't think that
will be easy.
The average data collection rate at 8.3.1 is 2 GB/hour or ~10 TB/year.
So I imagine storing all of the data from the ~100 MX beamlines around
the world would be a ~1 PetaB/year proposition. Since an average of 25
to 50 data sets are collected for every one that is published, the
storage demand on the PDB would be ~30 TB/year. Why only 1 in 50 you
ask? That is a very good question, and it will probably never be
answered unless the 49 of 50 unsolved data sets can be made available to
methods developers.
I just now Froogled for media prices and got this:
$33/TB LTO-4
$60/TB DVD-R
$100/TB hard disks
$400/TB Blue-Ray
$3000/TB Solid-state drives (such as USB thumbdrives)
$3M/TB clay tablets
So PDB will only need to find an "extra" ~$1k/year to buy the media for
1 dataset/structure, or $30k/year for all of the data. Unfortunately,
the media is not nearly as expensive as access to it. An LTO tape
library with ~50 TB storage capacity is ~$20k on eBay, but this is
EMPTY! You have to fill it with tapes, and then write software to make
the data sets available on the web. Tape librarys in the multiple PetaB
range are available, but not their prices.
Clearly this represents a non-trivial investment in resources and effort
for the PDB. The central problem is that the per-GB prices of storage
do not scale well to PetaB-class systems. However, there is now
Stimulus Package money available in the US for large equipment
investments like this. Perhaps someone at Rutgers could submit one? I,
for one, am very willing to write them a letter of support.
Another approach is to try and spread the storage out across the world
and create a central registry for finding it. The TARDIS initiative in
Australia (Androulakis et al. Acta D 2008) seems to be an important step
in that direction, but I haven't been able to test it since I don't have
a Fedora Repository Server. I do, however, have a web server, and I
think a repository of URLs is probably better than nothing.
-James Holton
MAD Scientist
David Aragao wrote:
> Dear All,
>
> I wonder how people currently do their long term backups. I see
> DATs/DLTs being slowly dropped off at the beamlines and most people
> brings their data home in external HDs.
>
> Anyone using blue-ray or double layer DVDs for long term backups? If
> so what kind of hardware? Do you use HDs for long term storage? If so,
> do you do a second copy and how do you store them?
>
> I will try to compile the answers and relay back to the list a resume.
>
> Thanks,
> David
>
|