While profiling Shadesmar a couple of weeks ago, I noticed that for large binary unserialized messages (>512kB) most of the execution time is spent doing copying the message (using memcpy) between process memory to shared memory and back.