Basically I'm using IDs on ClassNames now, so if you write an array of Objects, the className is only written once.
Besides that, I'm also using hashIds on field names. (composed by fieldType + name).
I was using also a pre-loaded DataContainer, and now I'm writing data direct to streaming when not doing deep copy. So, now JBossSerialization is also faster on regular ObjectOutputStream.
So far, I'm surprised by results on SingleThreadBenchmarkTestCase and MultiThreadJBossPerformanceTestCase. I wasn't expecting any alterations on performance after changing sizing, but it really made a difference.