Parallelising a Large Functional Program or: Keeping LOLITA Busy

A parallel version of the LOLITA natural language engineering system is under construction. We believe that, at 47,000 lines of Haskell, LOLITA is the largest non-strict parallel functional program ever. In this paper we report on the ongoing parallelisation of LOLITA, which has the following interesting features common to real world applications of lazy languages: { the code was not speciically designed for parallelism; { laziness is essential for the eeciency in LOLITA; { LOLITA interfaces to data structures outside the Haskell heap, using a foreign language interface; { LOLITA was not written by those most closely involved in the paral-lelisation. Our expectations in parallelising the program were to achieve moderate speedups with small changes in the code. To date speedups of up to 2.4 have been achieved for LOLITA running under a realistic simulation of our 4 processor shared-memory target machine. We are currently tuning the program on the Sun SPARCserver target machine, where wall-clock speedup is limited by physical memory availability and a sequential foreign-language front end. In the process of parallelising LOLITA we made changes to both our parallel software engineering techniques and our implementation technology. The most intriguing aspect, however, is that the parallelism is achieved with a very small number of changes to, and without requiring an understanding of most of the application. We attribute the ease of parallelisation primarily to the top-down approach enabled by evaluation strategies, our new mechanism for introducing and controlling parallelism in non-strict languages.