The explosive growth of mobile data traffic has raised big penetration to cellular network. Fortunately, offloading part of traffic through Delay Tolerant Network(DTN) would be a promising option to relieve load pressure. However, the existing routing strategies designed for DTNs are inappropriate in offloading scenario, where i) large data item would take on the dominance of mobile traffic, ii) centralized control can be exploited with merge between DTN and cellular network. To improve the routing performance of large data transmission, we propose a novel optimization model by creating two-tier solution space based on fragmentation,maximizing the probability that the requested data item is successfully delivered before expiration, taking limited buffer space as constraint metric. Moreover, a specific routing scheme is instantiated from our model, which is supported by uniform fragmentation and fine-grained path selection. Extensive trace-driven simulations show that our scheme is more appropriate for offloading case with outstanding performance in terms of replication overhead and acceptable routing capability. The explosive growth of mobile data traffic has raised big penetration to cellular network. Fortunately, offloading part of traffic through Delay Tolerant Network (DTN) would be a promising option to relieve load pressure. However, the existing routing strategies designed for DTNs are inappropriate in offloading scenario, where i) large data item would take on the dominance of mobile traffic, ii) centralized control can be exploited with merge between DTN and cellular network. To improve the routing performance of large data transmission, we propose a novel optimization model by creating two-tier solution space based on fragmentation, maximizing the probability that the requested data item is successfully delivered before expiration, taking limited buffer space as constraint metric. Moreover, a specific routing scheme is instantiated from our model, which is supported by uniform fragmentation and fine-grained path selection. Extensive trace-driven simulations show that our scheme is more appropriate for offloading case with outstanding performance in terms of replication overhead and acceptable routing capability.