Hi Arthur,
thanks a lot for investigating the performance of jackrabbit in such detail.
you seem to have quite a good understanding of the internals of jackrabbit. Can
I encourage you to propose patches for the hot spots you identified?
we will be happy to review them and incorporate the changes if performance can
be improved with them.
regards
marcel
Arthur Meyer wrote:
Hi,
these are the most important issues we have encountered so far:
Increasing the number of nodes in the repository also has a big negative
effect on the performance of queries.
When we multiplied our content by four, our queries took about twice as
much time to execute.
Hopefully, Lucene 2.0 may have improvements here aswell.
The item state cache is limited and can't be configured as far is I
know.
Since loading node items from the database is relatively expensive,
we would like to be able to cache as much as possible.
Node.isNodeType() calls cause monitor contention in a concurrent
environment.
For example, if you use the getUUID() method on nodes a lot, this will
become a serious bottleneck.
"http-18080-Processor8" daemon prio=1 tid=0x08571740 nid=0x79ca waiting
for monitor entry [0xa97f1000..0xa97f2eb0]
at
org.apache.jackrabbit.core.nodetype.NodeTypeRegistry.getEffectiveNodeTyp
e(NodeTypeRegistry.java:384)
- waiting to lock <0xb38677c8> (a
org.apache.jackrabbit.core.nodetype.NodeTypeRegistry)
at
org.apache.jackrabbit.core.NodeImpl.getEffectiveNodeType(NodeImpl.java:8
68)
at
org.apache.jackrabbit.core.NodeImpl.isNodeType(NodeImpl.java:1245)
at
org.apache.jackrabbit.core.NodeImpl.getUUID(NodeImpl.java:2802)
Further, the parsing of nodetype names seems to be a bottleneck in the
isNodeType() implementation
at
org.apache.jackrabbit.name.NameFormat.parseIgnoreCache(NameFormat.java:2
52)
at
org.apache.jackrabbit.name.NameFormat.parse(NameFormat.java:80)
at
org.apache.jackrabbit.core.NodeImpl.isNodeType(NodeImpl.java:2552)
Similar monitor contention issues occur in the CachingNamespaceResolver
class.
Using the util.concurrent classes could help solving these issues.
Finally, we're also very interested in using jackrabbit in a clustered
environment.
Arthur Meyer
