As an aside, in my performance testing of Lucene using JProfiler,
it seems
to me that the only way to improve Lucene's performance greatly can
come
from 2 areas
1. optimizing the JVM array/looping/JIT constructs/capabilities to
avoid
bounds checking/improve performance
2. improve function call overhead
Other than that, other changes will require a significant change in
the code
structure (manually unrolling loops), at the sacrifice of
readability/maintainability.
Just curious: are you more happy with JProfiler than with the JDK 1.5
profiler?
I haven't used JProfiler in quite a while but my impression back then
was that it's overheads tend to significantly perturb measurement
results. When I switched to the low-level JDK 1.5 profiler CPU tuning
efforts got a lot more targetted and meaningful.
So, in my experience, the least perturbing and most accurate profiler
is the one built into JDK 1.5. run java with
-server -agentlib:hprof=cpu=samples,depth=10' flags for long enough
to collect enough samples to be statistically meaningful, then study
the trace log and correlate its hotspot trailer with its call stack
headers (grep is your friend, a GUI isn't really needed). For a
background article on hprof see http://java.sun.com/developer/
technicalArticles/Programming/HPROF.html
Wolfgang.
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
