Transmission Control Protocol (TCP) performance over optical burst switching (OBS) is experimentally investigated on OBS testbed. The effect of burst losses on TCP performance over the OBS testbed is studied. The result shows that burst losses will lead to a network wide drop in TCP throughput and higher burst loss probabilities will result in lower available TCP bandwidth. Then the effects of delay penalty and correlation benefit are investigated in detail. Taking into consideration these effects, experimental results show that there exists an optimal assembly period to maximize the TCP bandwidth, and the optimal value is about 250 μs in the testbed, which is independent of the burst loss probabilities. To make the OBS testbed work steadily and improve the performance of TCP over the testbed to the greatest extent, the chosen burst assembly period should be equal to or slightly greater than the optimal value.