Interesting results. I cannot wait until we get this environment in place to really start banging against the server with slamd. So many ideas to test with respect to performance.
I did some small tests in order to check if storing DN into a BTree was
slower than using an Hash. Assuming that we have enough memory to cache
data, I found that using a Hash does not bring any kind of performance
boost compared to using a BTree. Worst, it slow down the wriate
operation (twice slower ...).
My small test was to store 500 000 DNs, using a 200 000 elements cache,
and searching for 500 000 random DNs from the created structure.
Some more insights :
- if the cache is not big enough, performances are just incredibly bad
(disk accesses, a hell lot ...)
- JDBM might be a little bit too light to support more than a few
hundred of thousands elements.