|
Abstract : |
Earlier work with the Parallel Document Retrieval Engine was oriented toward parallel machines such as the AP1000, characterised by many nodes, few disks, small memory per node (by current standards), single-user operation and high communication performance, relative to node computational power. Present generation parallel machines are much more like clusters of workstations (COWs). There are typically fewer nodes but each is more powerful, runs a multi-user operating system, supports more memory and connects to at least one local disk. In general, COWs are characterised by poorer network performance. PADRE has been redesigned to operate in the COW environment. Indexing and retrieval algorithms and user-interface have been totally replaced, along with the PADRE model of parallelism. PADRE97 minimises communication and synchronisation in order to improve scalability on high-latency clusters. Results are presented to show that the new design has achieved its objectives. 1, |
