Classification of very large datasets is a challenging problem in data mining. It is desirable to have decision-tree classifiers that can handle large datasets, because a large dataset often increases the accuracy of the resulting classification model. Classification tree algorithms can benefit from parallelization because of large memory and computation requirements for handling large datasets. Clusters of shared-memory multiprocessors (SMPs), in which each shared-memory node has a small number of processors (e.g., 2--8 processors) and is connected to the other nodes via a high-speed inter-connect, have become a popular alternative to pure distributed-memory and shared-memory machines. A cluster of SMPs provides a two-tier architecture, in which a combination of shared-memory and distributed-memory paradigms can be employed. In this paper we investigate decision tree construction on a cluster of SMPs. We present an algorithm that employs a hybrid approach. The classification training dataset is partitioned across the SMP nodes so that each SMP node performs tree construction using a subset of the
records in the dataset. Within each SMP node, on the other hand, tasks associated with an attribute are dynamically scheduled to the light-weight threads running on the SMP node. We present experimental results on a Linux PC cluster with dual-processor SMP nodes. (Also cross-referenced as UMIACS-TR-2000-78)