 |
OS version
If you're using many threads in your benchmark,
your benchmark will probably run better on HP-UX 11.0 than on 10.20.
You will need to adjust the tunable kernel parameters to accommodate
the number of threads in the benchmark.
first step
Run the benchmark with -verbosegc. Analyze the output looking at the amount
of space actually used by the application and the frequency of the
garbage collections.
tuning the JVM
Use a command line similar to:
java -ms<GCHeapSize> -mx<GCHeapSize> -alljit
-moncache=16x256 -montlscache=16 -allocCacheSize=4k <class>
where GCHeapSize is the size of the heap determined to result in the
maximum throughput for the machine on which the benchmark is running.
The size of the heap is dependent on the amount of memory on the machine.
A heap larger than the real memory size will force the machine to swap
pages during garbage collection. A heap too small will cause garbage
collections and compactions that are too frequent.
Additionally, using -noclassgc (not collecting classes), can reduce
garbage collection time. Using -verbosegc can help you determine the
tradeoff between the amount of space reclaimed by collecting classes
versus the amount of time required to do the garbage collection.
Often, we see the same class collected at each garbage collection,
which uses up space and time.
tuning the JVM process
On some benchmarks, using chatr for large pages will increase throughput:
cd /opt/java/bin/java/PA_RISC/native_threads
chmod +w java
chatr +pa 64M +pi 64M java
chmod 555 java
cd /opt/java/lib/java/PA_RISC/native_threads
chmod +w libjava.sl libnet.sl (+any others used)
chatr +pa 64M +pi 64M libjava.sl
chatr +pa 64M +pi 64M libnet.sl
chmod 555 libjava.sl libnet.sl (+any others used)
For some applications using a large negative renice of client and server
will boost performance:
renice -n -20 <pid>
for the BEA WebLogic benchmark
For 50,000 users, our BEA WebLogic expert has found that the following
is best for a large server (N-class):
ndd -set /dev/tcp tcp_conn_request_max 1024
java -ms768m -mx768m -moncache=16x128 -montlscache=16 -allocCacheSize=4k
-Dweblogic.system.gc.enabled=false <class>
The -Dweblogic.system.gc.enabled=false turns off the WebLogic
initiated GC. HP's JVM is better at scheduling garbage collections
when space is required rather than trying to have calls in the
application that do the garbage collection.
Other benchmarks will probably also benefit from turning off
program-generated garbage collections that occur when the
benchmark is running.
|
| |