Functional Multicasting

Multicasting, where a message is sent from one processor to number of other processors, can optimise the communication for data-parallel functional computing. This paper describes the introduction of two multicast primitives in the functional language Concurrent Clean. The usage and performance are showed, with special emphasis on the usage in combination with distributed arrays. In a data-parallel programming model, many instructions are sent to all processors. An eecient way to do this is broadcasting, which sends a message simultaneously to all processors available. Some interconnection schemes, like buses and trees, can support broadcasting directly, because the sender can have a direct connection to all other processors. Distributed architectures based on other interconnection schemes often have special hardware to support broadcasting. Sometimes a variant of broadcasting, multicasting is supported, where a message is sent only to a limited subset of processors. Broadcasting is a real optimisation technique: it does not introduce new functionality to a system, because the same eeect, albeit with lower performance, can be achieved by sending the same message multiple times. There has been a lot of work on parallel functional programming, but broadcasting or multicasting has not received much attention. This is partly because many ways of introducing parallelism in functional programs are not suited for it. Annotations PvE93] annotate a function application which will be send to one processor only. The par primitive of GpH PH93] evaluates one expression at one processor. Strategies TLHP96] could make use of multicasting, but no work on that has been mentioned yet. Other ways for introducing parallelism, like skeletons Col89] are rather high-level and an issue like multicasting is considered to be part of the implementation of the runtime system and not part of the language. In this paper we will introduce two multicast primitives which use the broadcast or multicast hardware to achieve better performance. No new functionality is added, as we can describe the semantics in standard Clean. The main use for multicasting is optimising the implementation of distributed arrays Ser97].