The Mariposa distributed database management system

The Mariposa distributed database management system is based on an economic paradigm in which processing sites buy and sell resources such as CPU time, I/O capacity and network bandwidth (STO96). In a Mariposa economy, a site assigns a price to perform a service and may adjust its prices as it sees fit. In designing Mariposa, it was our hope that market forces could be used to achieve good performance while maintaining the independence of each processing site. In this paper, we address the problem of distributed query processing in Mariposa and compare its performance to a traditional cost-based distributed query optimizer. We demonstrate the ability of a Mariposa system to adapt to a dynamic workload by using simple economic concepts such as the effect of supply and demand on prices. We investigate the benefits of adaptive distributed query processing in Mariposa and its interaction with multi-user workloads, network latencies and query size. We present performance results which show that in multi-user situations, when response time is used as a metric, the Mariposa system outperforms a static optimizer by causing work to be distributed more evenly among the available sites and that the overhead introduced by Mariposa's bidding protocol is insignificant when used with large, expensive queries and is outweighed by the benefits of load balancing, even for relatively small queries. Our experiments demonstrate that the point at which our approach outperforms a static optimizer is affected by network latency and query size. Our performance comparisons are based on the TPC-D benchmark.