Online dedup for Btrfs
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the memcmp() dance to make sure that the extent we are deduping really matches the extent.
So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.git dedup
Thanks!
Josef
Quoted 33 lines · Josef Bacik
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the
memcmp()dance to make sure that the extent we are deduping really matches the extent.So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.gitdedup
Quoted 7 lines
Thanks!
Josef
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Hey Josef,
that's really cool! Can this be used together with lzo compression for example? How high (roughly) is the impact of something like force-compress=lzo compared to the 15% hit from this dedup?
Thanks!
Harald
Apart from the dates, this sounds highly plausible :-)
If the hashing is done before the compression and the compression is done for isolated blocks, then this could even work!
Any takers? ;-)
For a performance enhancement, keep a hash tree in memory for the "n" most recently used/seen blocks?...
A good writeup! Thanks for a good giggle. :-)
Regards,
Martin
Quoted 50 lines · Harald Glatt
Quoted 33 lines · Josef Bacik
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the
memcmp()dance to make sure that the extent we are deduping really matches the extent.So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.gitdedupQuoted 7 lines
Thanks!
Josef
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.htmlHey Josef,
that's really cool! Can this be used together with lzo compression for example? How high (roughly) is the impact of something like force-compress=lzo compared to the 15% hit from this dedup?
Thanks!
Harald
Quoted 26 lines · Josef Bacik
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the
memcmp()dance to make sure that the extent we are deduping really matches the extent.So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
Quoted 9 lines
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.gitdedupThanks!
It's been pointed out to me that this is probably too serious, so just FYI it's April 1st where I am. Thanks,
Josef
Oh man :D It was so elaborate that I really believed it.... :P
Quoted 45 lines · Josef Bacik
Quoted 26 lines · Josef Bacik
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the
memcmp()dance to make sure that the extent we are deduping really matches the extent.So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
Quoted 9 lines
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.gitdedupThanks!
It's been pointed out to me that this is probably too serious, so just FYI it's April 1st where I am. Thanks,
Josef
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 32 lines · Josef Bacik
Quoted 31 lines · Josef Bacik
Hello,
I was bored this weekend so I hacked up online dedup for Btrfs. It's working quite well so I think it can be more widely tested. There are two ways to use it
1) Compatible mode - this is a bit slower but will handle being used by older kernels. We use the csum tree to find duplicate blocks. Since it is relatively easy to have crc32c collisions this also involves reading the block from disk and doing a memcmp with the block we want to write to verify it has the same data. This is way slow but hey, no incompat flag!
2) Incompatible mode - so this is the way you probably want to use it if you don't care about being able to go back to older kernels. You select your hashing function (at the momement I only support sha1 but there is room in the format to have different functions). This creates a btree indexed by the hash and the bytenr. Then we lookup the hash and just link the extent in if it matches the hash. You can use -o paranoid-dedup if you are paranoid about hash collisions and this will force it to do the
memcmp()dance to make sure that the extent we are deduping really matches the extent.So performance wise obviously the compat mode sucks. It's about 50% slower on disk and about 20% slower on my Fusion card. We get pretty good space savings, about 10% in my horrible test (just copy a git tree onto the fs), but IMHO not worth the performance hit.
The incompat mode is a bit better, only 15% drop on disk and about 10% on my fusion card. Closer to the crc numbers if we have -o paranoid-dedup. The space savings is better since it uses the original extent sizes, we get about 15% space savings. Please feel free to pull and try it, you can get it here
Quoted 6 lines
Quoted 4 lines
git://git.kernel.org/pub/scm/linux/kernel/git/josef/btrfs-next.gitdedupThanks!
It's been pointed out to me that this is probably too serious, so just FYI it's April 1st where I am. Thanks,
Well I believed it too, and was writing an email with questions etc. I almost sent it, but then I saw git was downloading hundreds and hundreds of MB of data :)
Well done anyway!
Quoted 6 lines
Josef
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html