Garbage collection tuning questions

Question

I apologize if these questions have been answered elsewhere.
We are using java 1.4.2_08
SunOS pswebp10 5.9 Generic_118558-06 sun4u sparc SUNW,Sun-Fire-V440
2 CPUs
weblogic 8.1 SP4
Java options:
-server
-Xms2304m
-Xmx2304m
-XX:+UseAdaptiveSizePolicy
-XX:PermSize=192m
-XX:MaxPermSize=192m
-Xloggc:/www/tmsproddomain/logs//gc.log
-XX:+PrintGCDetails
-XX:+PrintGCTimeStamps
-XX:+DisableExplicitGC
-XX:+PrintTenuringDistribution
-XX:+UseParallelGC
-Djava.awt.headless=true
The JVM spaces get created as:
Eden:768
S0: 85
S1: 85
Old: 1.5
Perm: 192
The survivor spaces increase in size over time, so
the S[01] sizes reported above are obviously larger than
at JVM start.
*) We are using the UseParallelGC garbage collector. Our
system has 2 CPUs and were wondering if adding 2 more CPUs
would improved GC times at all. If so, any ballpark
on how much?
*) We once attempted to use the ParNewGC collector as well
as UseConcMarkSweepGC with
a large allocation as in the above specs, and found that it
made strange decisions on the sizes of the various memory
spaces that led to very frequent GCs. This was with an older
1.4.2 JVM though. It was strange enough that we chose to go
with the Parallel collector. Is there information on how
the ParNewGC collector sizes the memory spaces?
*) Is there any documentation on these two collectors,
i.e how objects promoted from eden to old, etc. The docs I have
seen are all about the other, older collectors.
*) Does the UseAdaptiveSizePolicy work with these 2 collectors?
*) If the AggressiveHeap option is used, does this
have any issues with either of these 2 collectors?
*) the -Xloggc option seems to return less information
than the verbosegc option - is this documented any where?
Can the two options be used together to have all the output of
verbosegc, but in the designated log file?
Testing the ParNewGC and ConcSweep,
I get:
Eden Space: 7.8M
S0: 64K
S1: 64K
Old: 1.45G
Perm 192
And this collector seems to use the tenuring system,
with 31 slots.
It seems like to use this collector, one might want to
speed up the migration of objects from eden to old
by tuning the tenuring?
Any advice here would be greatly appreciated.

Answer 1

(1) The basic stuff:
http://java.sun.com/docs/hotspot/
(2) The presentations from JavaOne conference might be of use:
http://developers.sun.com/learning/javaoneonline/
(3) Something I've stumbled across at some point:
http://ua.sun.com/java/Performance_SRitter.pdf
I think that you should leave the sizing and ratios up to the GC; with 1.5 it seems to totally ignore your input, and there the adaptive sizing is on by default (might be for 1.4 as well).
Also, when using CMS, there is some GC done in parallel, so having more GC events output doesn't necessarily mean that your app is running slower; the ParallelGC is to maximize throughput, while CMS is when you need lower GC pause times, and don't mind sharing the CPU with the GC threads.