The authors investigate how to improve real transmission control protocol (TCP) performance over optical burst switching (OBS) networks, providing larger available bandwidth and reliability. Considering the acknowledgement (ACK) packet losses, a simple modified analytic model is studied to give the real steady-state TCP throughput over OBS networks. The effect of ACK losses is analyzed based on the calculation and simulation results, and it is concluded that eliminating or mitigating ACK losses should improve the performance. Then, at the OBS routers, two novel ACK-based assembly and scheduling mechanisms, viz. a timer-based ACK assembly scheme and an ACK priority scheduling scheme, are proposed. Simulation results demonstrate that compared with conventional assembly and scheduling algorithms, the two new mechanisms can improve TCP performance over OBS networks to great extent, meeting the requirements of grid networks, such as lower network burst loss probability to provide reliable performance, and higher TCP throughput to increase the available bandwidth. It is also seen through comparison that the novel ACK priority scheduling mechanism has better performance than the usual one.