On 31/03/11 15:00, Frank Budinsky wrote:
Hi Brian,
On my Xeon server the number of graphs makes a big difference in the
test, since at first (e.g., the first 5k) the performance of direct and
memory mapped modes is basically the same, but as it continues to run,
direct mode stays constant, but the memory mapped mode keeps getting
slower and slower.
Btw, I tried various JVMs and the times vary somewhat, but do not change
the really slow time I see with memory mapped mode. It really looks like
memory mapped mode just doesn't work properly on my particular hardware??
Brian made a interesting discovery about OS disk caching settings.
http://mail-archives.apache.org/mod_mbox/incubator-jena-users/201103.mbox/%[email protected]%3E
Frank.
I'm using ext4 on Ubuntu.
(there is various discussions about issues with ext3 and out-of-order
writes and disk corruption).
As previously mentioned: in the development version of TDB, less sync()
operations are done - only explicit calls to sync cause the DB to sync.
There is a snapshot build with those changes in.
Andy
