Since plsiBackend has very little communication (a few broadcasts before computation begins and a global sort after all the cosines have been computed) and the data is distributed evenly between processors, its scalability depends more on the size of the document collection being searched than the computation required to search the collection. As shown by the speedup graphs, smaller document collections limit the number of processors that can be applied to the problem while still decreasing the time to search the collection. For example, neither the PVM nor the PLMTP document collections benefit from more than 6-8 processors. However, the CCE collection appears to use up to 16 processors efficiently while the USENET collection might be able to use more than 24 processors without difficulty.
Given enough term and document vectors, plsiBackend is scalable to a relatively large number of processors. Since plsiBackend requires very little communication, and the data is distributed evenly across the nodes, plsiBackend should scale to almost any number of processors.