User:StefanoZacchiroli/Content deduplication

Some experiments on deduplicating contents at sub-file granularity.

Datasets

Linux kernel, Git repo

• origin: git.kernel.org, on 2018-01-06
• 1.653.941 content blobs, for a total of 19 GB (compressed)
• original size (uncompressed): 55.89 GB

Random sample 0.1%

• random 0.1% Bernoulli sample of Software Heritage contents, took on 2018-01-09
• 4.111.582 contents blobs, for a total of 155 GB (compressed)
• original size (uncompressed): 286.99 GB

Rabin fingerprint chunking

• Approach: use Rabin fingerprints as in LBFS
• Implementation: swh-dedup-blocks.py

test 1

Dataset: linux.git

Rabin fingerprint parameters:

• prime: 3
• window_size: 48
• chunk size (min/avg/max): 2 KB / 8 KB / 64 KB

Results:

• average chunk size (effective): 9.37 KB
• dedup chunk size (uncompressed): 19.87 GB (35.55%)

test 2

Dataset: linux.git

Rabin fingerprint parameters:

• prime: 3
• window_size: 48
• chunk size (min/avg/max): 512 B / 2 KB / 8 KB

Results:

• average chunk size (effective): 2.86 KB
• dedup chunk size (uncompressed): 9.09 GB (16.26%)

test 3

Dataset: linux.git

Rabin fingerprint parameters:

• prime: 3
• window_size: 48
• chunk size (min/avg/max): 512 B / 1 KB / 8 KB

Results:

• average chunk size (effective): 1.72 KB
• dedup chunk size (uncompressed): 6.49 GB (11.60%)

test 4

Dataset: random sample 0.1%

Rabin fingerprint parameters:

• prime: 3
• window_size: 48
• chunk size (min/avg/max): 512 B / 1 KB / 8 KB

Results:

• average chunk size (effective): 1.56 KB
• dedup chunk size (uncompressed): 244.09 GB (85.05%)

References

• LBFS
• Hailstorm
• DCT
• ScanCode tokenizer
• Pcompres chunking (linked from casync)