So what do you think about the proposal now, when benchmarks shows that keeping whole tilesets in one file is faster?
On 6/11/08, Jiri Klement <[EMAIL PROTECTED]> wrote: > Florian's benchmark already showed that reading from one big file is > faster than reading from lots of small files. My original idea was > keeping one z12 tileset in one file, not having one big file for many > tilesets as Florian tested, so I've made my own benchmark. > > The benchmark can be downloaded here: > http://jttt.110mb.com/tahbench.jar > > Tilesets have on average 500 files, one tileset is about 3.7MB. > > Results for XFS filesystem: > Creating 1000 tilesets: > flat 147 s > dir about 2 hours > > Reading 10000 random tiles > flat 151 s > dir 278 s > > Reading more than one tile from tileset using flat tileset > 1 tile - 151 s > 4 tiles - 100 s > 8 tiles - 81 s > > Reiserfs > Creating 10000 tilesets > flat - 468 s > dir - 4464 s > > Reading 100000 random tiles > flat - 1526 s > dir - 2604 s > > Reading more than one tile from tileset using flat tileset (10000 tiles) > 1 tile - 207 s > 4 tiles - 143 s > 8 tiles - 88 s > > Notes: > XFS is not faster than reiserfs, I've just used smaller number of tilesets. > Disc cache was droped before each test using echo 3 > > /proc/sys/vm/drop_caches > flat = one file per z12 tiles; dir = one file per tile > XFS and reiserfs were tested on different drives. XFS was tested on > laptop drive, reiserfs on old 20GB drive. > No caching is perfomed by application when reading more than one tile > from tileset. > _______________________________________________ Tilesathome mailing list [email protected] http://lists.openstreetmap.org/cgi-bin/mailman/listinfo/tilesathome
