Several people have proposed decreasing the default LMDB map size from 1TB down to something smaller (#3644, #3728). The large default causes problems on ARM and Windows, and makes Valgrind unhappy (#2404).
@hyc, a maintainer (the author?) of LMDB, suggested catching MDB_MAP_FULL errors and growing the map size as needed.
you can just grow the map whenever you hit a MDB_MAP_FULL error
NVIDIA/DIGITS#206 (comment)
This was implemented with only a few lines of code in DIGITS:
https://github.com/NVIDIA/DIGITS/blob/v3.0.0/tools/create_db.py#L636-L660
2016-02-25 12:48:41 [DEBUG] Processed 200/60000
2016-02-25 12:48:42 [DEBUG] Doubling LMDB map size to 20MB ...
2016-02-25 12:48:43 [DEBUG] Processed 500/60000
2016-02-25 12:48:44 [DEBUG] Doubling LMDB map size to 40MB ...
2016-02-25 12:48:45 [DEBUG] Processed 800/60000
2016-02-25 12:48:48 [DEBUG] Processed 1200/60000
2016-02-25 12:48:48 [DEBUG] Doubling LMDB map size to 80MB ...
2016-02-25 12:48:50 [DEBUG] Processed 1600/60000
2016-02-25 12:48:52 [DEBUG] Processed 1800/60000
2016-02-25 12:48:54 [DEBUG] Processed 2200/60000
2016-02-25 12:48:56 [DEBUG] Doubling LMDB map size to 160MB ...
2016-02-25 12:48:56 [DEBUG] Processed 2600/60000
...
I submit that the same should be done for Caffe rather than simply changing the default map size.
Several people have proposed decreasing the default LMDB map size from 1TB down to something smaller (#3644, #3728). The large default causes problems on ARM and Windows, and makes Valgrind unhappy (#2404).
@hyc, a maintainer (the author?) of LMDB, suggested catching MDB_MAP_FULL errors and growing the map size as needed.
This was implemented with only a few lines of code in DIGITS:
https://github.com/NVIDIA/DIGITS/blob/v3.0.0/tools/create_db.py#L636-L660
I submit that the same should be done for Caffe rather than simply changing the default map size.