Practical experience with TCP over GPRS

We present the results of a series of experiments used to characterise the performance of a GPRS (General Packet Radio Service) wireless data network, hence highlighting issues that software architects should consider when designing applications to run over this soon to be widely-deployed service. In summary, we show that packet round trip time (RTTs) are large (>1000 ms) and can be highly variable, packet losses are relatively rare, and that available bandwidth can be quite variable. These network characteristics do not interact well with current TCP implementations. We show how it takes many seconds before a new TCP connection can expand its congestion window to make use of the full bandwidth available, leading to very poor performance of protocols like HTTP. Beyond the point of full bandwidth utilisation, TCP continues to expand the window needlessly, resulting in excessive queueing at the GPRS router. This leads to greatly inflated RTTs (10s of seconds) and hence poor interactive response and slow recovery should loss occur. We show how a simple transparent proxy interposed between the fixed and GPRS networks can be used to significantly improve TCP connection performance, particularly for activities like Web browsing.