Compression for data archiving and backup revisited

Abstract

Data deduplication is a simple, dictionary based compression method that became very popular in storage archiving and backup. It has the advantage of direct, random access to any piece ("chunk") of a file in one table lookup; that's not the case with differential file compression, the other common storage archival method. The compression efficiency (chunk matching) of deduplication improves for smaller chunk sizes, however the sequence of hashes replacing the deduplicated object (file) increases significantly. Within the sequence of chunks that an object is decomposed, sub-sequences of adjacent chunks tend to repeat. We exploit this insight first in an online scheme used to reduce the amount of hash metadata generated. With each newly created entry in the chunk repository we add a "chronological" pointer linking this entry with the next new entry, in time order. When the hashes produced by the chunker follow the chronological pointers we encode them as a "sequence of hashes" by specifying the first hash in the sequence and the length of the sequence. The shrinkage is orders of magnitude smaller than what a customary compression algorithm (gzip) achieves and has a significant impact on the overall deduplication efficiency when relatively small chunks are used. A second scheme is also introduced that optimizes the chunk sizes by joining repeated sub-sequences of small chunks into new "super chunks" with the constraint to achieve practically the same matching performance. We employ suffix arrays to find these repeating subsequences and to determine a new encoding that covers the original sequence. As a result, fewer chunks are used to represent a file, reducing fragmentation i.e. the number of disk accesses needed to reconstruct the file, and requiring fewer entries in the chunk dictionary and fewer hashes to encode a file. As the experimental results show, this method provides over 10 time reduction in fragmentation and over 5 times reduction in the number of entries in repository while achieving similar or slightly better overall deduplication ratios. © 2009 SPIE.