When computing the speedup of plsiBackend, the timings provided by lsiBackend were used as the base serial performance. Since lsiBackend must load each term and document vector from disk as it is needed while each processor executing plsiBackend loads its set of term and document vectors into memory upon initialization (before timings on individual queries begin), the times produced by lsiBackend and plsiBackend are not directly comparable. However, the algorithm used by lsiBackend may be considered to be the best available serial algorithm since any single processor does not have enough local storage to load all the term and document vectors for a given collection into memory at once.