| layout | title | permalink | redirect_from | ||
|---|---|---|---|---|---|
post |
ARCHIVE |
/docs/archive |
|
Training neural networks on very large datasets is not easy (an understatement).
One of the many associated challenges is a so-called small-file problem - the problem that gets progressively worse given continuous random access to the entirety of an underlying dataset (that often also has a tendency to annually double in size).
One way to address the small-file problem involves providing some sort of serialization or sharding that allows to run unmodified clients and apps.
Sharding - is exactly the approach that we took in AIStore. Archiving or sharding, in the context, means utilizing TAR, for instance, to combine small files into .tar formatted shards.
While I/O performance was always the primary motivation, the fact that a sharded dataset is, effectively, a backup of the original one must be considered an important added bonus.
Today AIS equally supports formats: TAR, TGZ (TAR.GZ), TAR.LZ4, ZIP, where:
- TAR is a well-known format first introduced in Unix V7 circa 1979 with specific formatting flavors including USTAR, PAX, and GNU TAR (all three are equally supported);
- TGZ (aka TAR.GZ) and TAR.LZ4 provide, respectively, gzip and lz4 compression to tar files (aka tarballs);
- and ZIP is PKWARE ZIP first introduced in 1989.
AIS can natively read, write, append(**), and list archives.
All sharding formats are equally supported across the entire set of AIS APIs. For instance, list-objects API supports "opening" shards
ie., objects formatted as .tar, .tgz, etc. - see above
and including the corresponding pathnames into generated result sets. Clients can run concurrent multi-object (source bucket => destination bucket) transactions to en masse generate new archives from selected subsets of files.
APPEND to existing archives is also provided but limited to TAR only.
Maybe with exception of TAR, none of the listed sharding/archiving formats was ever designed to be append-able - that is, not if we are actually talking about appending and not some sort of extract-all-create-new type emulation (that will certainly break the performance in several well-documented ways).
See also: