I think by far, MIRIAD needs to have a re-examination of its IO infrastructure (it appears that most (all?) tasks are heavily IO bound). The current infrastructure is doing things that are best handled by the OS, and any improvements to how the IO is done is hard/improbable to engineer at the moment.
This also has implications for how memory management is accomplished in the future. A mixture of in memory combined with some understanding of available real memory vs leaving data on disk when unneeded is probably the direction it should take.
Colby 15:29, 19 February 2008 (EST)