Fast recovery from losses is an important problem for reliable data communication protocols. The Internet standard TCP protocol (Transmission Control Protocol) implements two algorithms for that: an original one based on a special timer, and a later addition to TCP that is triggered when there are three packets received by the client after the missing one. Web communications are most frequently carrying rather small objects that are generally ineligible for anything other than timeout-based retransmission schemes. The present invention describes a system delivering performance improvements to TCP (and other reliable data communication protocols) that are compliant with present Internet standards. One improvement is to send duplicates of packets without waiting for a timeout to occur. Another is to split particular parts of a data stream into smaller packets. A third is to send packets out-of-order to force the remote host to acknowledge all packets that were successfully received.