So, from what I understand, this is great but more of a proof of concept since fuse performance kills it. As far as putting it in production, there are a few unresolved questions which I haven't seen picked apart:
- Can dedup be integrated into the VFS layer, like unionfs is shooting for, or does it have be integrated with the underlying filesystem.
- Is online dedup possible, and does the answer change when running SSD.
- What's the best granularity (block-level? inode-level? block extent-level?) and how badly can it randomize the i/o. I imagine one would have to do a lot of real-world benchmarking to find this out.
- Are there possible privacy issues (i.e. finding through i/o patterns whether someone else has a given block or file stored) and how to deal with them
I was wondering - with the current amount of abstraction and similar (sometimes redundant) metadata on almost everything - what percent of duplicate blocks could be found on a standard desktop system?
I don't think it would be useful, I'm just interested in the level of "standard" data duplication.
"I was just toying around with a simple userspace app to see exactly how much I would save if I did dedup on my normal system, and with 107 gigabytes in use, I'd save 300
megabytes."
It's a relatively small amount. Then again - you're storing 300MB of exactly the same blocks of data... Unless they're manual backup files, this looks like a big waste to me.
Yup, that's about the same proportion I found when I recently tried copying my data across to a ZFS system with the dedup switched on.
I then decided to disable the dedup, because it comes at a cost - the checksum data (which would mostly be living on the SSD read cache I had attached) was occupying about 3 times the monetary worth of SSD storage space than the monetary worth of conventional disk space that the duplicate data was occupying.
I noticed that the opendedup site (linked from the article) claims a much lower volume of checksum data, relative to number of files; perhaps an order of magnitude less than I observed with ZFS, but they seem achieve that by using a fixed 128KB block size, which brings along its own waste. (ZFS uses variable block size.) I haven't actually done the numbers here but I wouldn't be at all surprised to find that for my data, the 128KB block size would be costing as much disk space as what dedup was saving me. (YMMV, of course.)
As it happens, in between first reading about ZFS dedup, and finally trying it out, I seem to have forgotten that the verify option existed. I just did "set dedup=on" - beyond that everything was whatever defaults you get on OpenIndiana build 148.
Were I to ponder that matter to any great depth, I suspect I'd find it rather difficult to get a handle on how concerned I ought to be about hash collision. Perhaps that's part of the answer to your puzzlement.
Because they believe in statistically low probability of that happening.
What they don't accept is that it might hit them in the worst time possible, that a backup of such information will also suffer the same problem, that after an upgrade an important part of the new kernel might have the same hash as the beginning of Morissette's "Ironic" mp3 you already store, etc.
I tested it and I don't recommend it. (It was like a year ago though)
It was really slow and some blog posts about the reliability of the data storage backend were a little bit scary.
I would recommend using zfs-fuse. You don't have the FUSE->File on a filesystem->Hard disk indirection (thus more speed). And additionaly you get all the cool ZFS features!
If you need even more speed there is a ZFS kernel module for linux and a dedup patch for btrfs. I don't think those are production ready though.
I tried ZFS dedup but there was something like a 20x slowdown to write files compared to ZFS without dedup, and this was on under ten gigabytes of files. I don't know if I somehow had the cache settings wrong or what the problem was, but I didn't manage to fix it, even trying both FUSE and kernel versions. (On ubuntu 11.4)
Yeah random acess on hard disks is awfully slow. And if you have dedup you can cause lots of random access...
If you have a little bit of data the hashtable used for dedup can also be to big to fit into memory. Then ZFS puts it onto the disk and it is even slower. Luckily there is a feature to use SSDs as a cache device in this case.
The tricky part seems to be 'too big to fit into memory'. From what I understood and calculated the dedup tables on my system should have been well under 100MB, and the amount of memory designated for metadata was over 350MB, yet the performance was terrible.
Based on my testing (not published anywhere, sorry) ZFS dedup works best when you enable compression. With compression, it's only slightly slower then without dedup.
ZFS is designed to have lots of horsepower and memory thrown at it.......big servers, available CPU power, lots of ECC ram. If there's going to be an SSD allocated as a cache disk, it's probably expected to be huge and enterprisey too.....
ZFS is awesome, but some features will be disappointing unless you are dealing with adequate resources.
Actually excepting DataDomain ultra expensive specialised hardware (and maybe a couple of similar enterprise solutions), all dedupe systems come with a huge performance hit. ZFS is no exception...
I don't understand the complication of using a database. The sensible approach would be something like BMDiff with [page] indexing on top for random access.
I remember a spate of academic articles a few (3-7?) years ago talking about how all filesystems were going to be replaced by single huge databases to hold all our "files", maybe this is partially a continuation of that research.
lessfs appears to do block level deduplication (like ZFS). This means that if I copy a huge file but add a few bytes at the start, I won't get any benefit from deduplication because the data doesn't align anymore with the original block boundaries.
* first post: http://blogs.intellique.com/tech/2010/12/22#dedupe
* detailed setup and benchmark results: http://blogs.intellique.com/tech/2011/01/03#dedupe-config
After more than 9 months running lessfs, I recommend it.