Performance issue with Weblogic 9

Hi all, 

I encounter a performance issue with Weblogic 9 (both 9.1.0 and 9.2.0) when
running performance test on a web application.  Basically, the specific function
I tested on consisted of two web applications (both of them are deployed on the
same managed server).   

Here are the steps involved in executing the operation:

- The first web application (SOAP/xml) receives the request and does the common
validation, autehntication and authorization; 
- The first web application gets the URL of the second server from the database
and sends HTTP request to the second web application
- The second web application (servlet) does some data processing (that requires
no database access) and returns the results to the first application.
- The first application packages the results into response and tears down the
call. 


Test setting:
A 30-thread ramp-up test against the first application, with thread incrementing
every 60 seconds.  


Problems:
I observe a performance drawback happened at the ramp-up.  TPS climbs steadily
from 1 to 5 threads, then TPS drops rapidly from 5 to 10 threads; after that,
TPS starts to climb back up again all the way to 30 threads.  Note that the TPS
recorded at 5 threads is the highest.  Below is a sample of the metrics
measured. 

#Threads	TPS	Avg_Latency
1	80	12
2	150	13
3	216	14
4	262	15
5	292	17
6	304	20
7	265	27
8	203	40
9	136	68
10	127	81
11	118	96
12	117	105
13	121	111
14	124	116
15	126	122
16	133	124
17	135	129
18	144	128
19	149	131
20	156	132
21	161	133
22	156	145
23	172	137
24	178	139
25	180	143
26	183	146
27	186	149
28	184	157
29	187	159
30	183	163



Extra findings:
- the response time of the second web application (recorded inside the code of
second web application) is climbing steadily with no bumps.  

- the response time of the second web application (recorded inside the code of
the first web application) shows a sudden lift at the 5th thread.  

- Exact same applications and tests on Weblogic 8.1.6 show no performance issue
at all. 

- Enabled "use81-style-execute-queues" flag on WL9 and specified
default execute pool size to 50.  The same performance drawback is seen around 5
threads.  

- There is no error or warning found in logs. 


I greatly appreciate if you could give me some advices or suggestion on this
situation.  


Million thanks,