Restart Strategies and Internet Congestion

In the emerging age of electronic commerce, it is of paramount importance to design mechanisms that ensure timely and reliable transactions in cyberspace. This is an important problem given the bursty nature of Internet congestion, which leads to large variability in the risk and cost of executing transactions in a distributed network environment. Earlier, we presented a methodology for quantitatively managing the risk and cost of executing transactions in such an environment. By associating the cost with the time it takes to complete the transaction and the risk with the variance in that time, we considered different methods that are analogous to asset diversification, and which yield mixed strategies that allow an efficient trade-off between the average and the variance in the time a transaction will take. Just as in the case of financial portfolios, we found that some of these mixed strategies can execute transactions faster on average and with a smaller variance in their speed. A potential problem with this portfolio methodology is that, if everybody uses it, the latency characteristics of the Internet might shift so as to render the method useless. In order to investigate this issue, we conducted a series of computer simulations of a group of agents deciding asynchronously whether to use the Internet or not. The agents base their decision on knowledge of the congestion statistics over a past window of time. We find that when every agent uses the portfolio strategy there is still a range of parameters such that (i) a portfolio exists and (ii) all agents are better off using it than not. Even when all agents do so, the optimum restart strategy remains preferable to the situation in which no one uses the restart strategy. Finally, use of the portfolio strategy increases the variance of Internet traffic in our model, thus making use of the portfolio strategy more attractive as more agents use it.