Java: Heap crunch

Profile-code-deploy-load was all that I did yesterday and yes it did improve the system performance and technically it did survived the brutal four hour load test without causing OOOE (out of memory exception). The primary hypothesis under which I'm working is that their the certain long living references have retained a sizable of memory. But, I'm sure I'm not exactly able to find the root cause for the system degradation under load. The important observations are;

• We are loading the system using a pattern where we ramp up the load in x minutes and maintain the peak load for y hours and hen ramp down to zero in x minutes.
• The heap usage percentage follows the similar pattern during ramp up, but although the load has plateaued the heap usage percentages continues go higher, while ramp down it eases a bit but never follows the decrease in load and never actually reaches zero or minimum even though there is no activity in the application.
  
• CPU times in GC increases almost linearly albeit with a low gradient and decreases to zero after load completes.

I plan to study GC logs today to understand the system characteristics and fine tune the JVM fit for our purpose and I am sill not sure if I have a memory leak... more later.