[Tiling] Tilestore Benchmarks

[Arne Kepp]

These results are great, doesn't look like SQLite should not be a 
bottleneck. I have basically the same question as Steve, is SQLite 
better at getting the attention of the filesystem cache? (What does the 
memory allocation look like towards the end of those runs ?)

Things you can hopefully add, to make it easier to compare the results:
What filesystem did you use during these tests?
What options was the filesystem mounted with, such as noatime or 
writeback caching? (The latter is key if you compare with async sqlite.)
What is the block size of the filesystem?

-Arne


On 5/2/11 3:35 PM, Oliver Tonnhofer wrote:
> Hi there,
>
> I promised to post my benchmarks of an SQLite based tile store. I had to rerun my tests with different options and different cache sizes, but here are finally the results:
>
> http://dev.mapproxy.org/misc/tilestorebench/
>
> As you'll notice, SQLite is faster in any benchmark and some of the numbers are quite impressive.
>
> With a hot filecache, the SQLite store peeked at 26k tiles/s while the file system peeked only at 2.5k (the graphs are smoothed with a running average of 10). Even for tile sets that do not fit in memory SQLite is much faster (~300 vs. ~150t/s with meta requests, ~85 vs ~40t/s for complete random requests).
>
>
> Regards,
> Oliver
>