On 3/14/07, Emmanuel Lecharny <elecharny@gmail.com> wrote:
Alex Karasulu a écrit :

> This is wicked cool ... hurry up and commit the darn BigInteger -> Long
> replacement :).

Already committed in trunks, two days ago :)

> Thanks E!
> Alex

Wait, this is just nothing... I have played with the
AttributesSerializer, and the gain is huge : more than 25% speedup
against the server with Long. So it's around 30% faster than ADS 1.0 :)
You mean the one I wrote a while back?  I did some tests on it and it did not seem to
be making much difference.  Perhaps I was not measuring things well.  Anyways
this is great news regardless.  I cannot wait until we get to a phase where we can start
improving on performance again.

The tests I have done this morning with YJP showed that readObject was
responsible for 55% of all the time spent in the server. Improving the
serialization of data will result in a massive gain (estimated to a
speedup to 40%). As the best results we got last year with Ersin 4-cpu
computer and JRockit JVM went up to 950 request/s, we can expect
something like 1300 req/s. May be more :) And this will be 10% slower
than FDS or 15% slower than OpenLdap for this kind of test ... (those
numbers are *not* official !)
That's incredible.  Also there are lots of optimizations we could do besides these with
respect to heuristics used in the default partition design.  I think a Java LDAP server
might one day be just as fast as the C based servers.  At least SUN thinks so right :).