San Fermín: Aggregating Large Data Sets Using a Binomial Swap Forest

San Fermı́n is a system for aggregating large amounts of data from the nodes of large-scale distributed systems. Each San Fermı́n node individually computes the aggregated result by swapping data with other nodes to dynamically create its own binomial tree. Nodes that fall behind abort their trees, thereby reducing overhead. Having each node create its own binomial tree makes San Fermı́n highly resilient to failures and ensures that the internal nodes of the tree have high capacity, thereby reducing completion time. Compared to existing solutions, San Fermı́n handles large aggregations better, has higher completeness when nodes fail, computes the result faster, and has better scalability. We analyze the completion time, completeness, and overhead of San Fermı́n versus existing solutions using analytical models, simulation, and experimentation with a prototype built on peer-to-peer system deployed on PlanetLab. Our evaluation shows that San Fermı́n is scalable both in the number of nodes and in the aggregated data size. San Fermı́n aggregates large amounts of data significantly faster than existing solutions: compared to SDIMS, an existing aggregation system, San Fermı́n computes a 1MB result from 100 PlanetLab nodes in 61–76% of the time and from 2-6 times as many nodes. Even if 10% of the nodes fail during aggregation, San Fermı́n still includes the data from 97% of the nodes in the result and does so faster than the underlying peer-to-peer system recovers from failures.