This comment comes under the following:
- Wishes for additional enhancements or new functionality to be taken into account after 3.0 is published.
I think there are some cases where persistent collective requests would make sense. Specifically as an opportunity to allow auto-tuning of the implementation based on the call parameters. The most compelling justification for this is to optimise AlltoAllv operations. Unlike the AlltoAll operation these cannot easily use tree based optimisations for small message sizes as each process only knows that part of the communication pattern they are directly involved with. However if the same communication pattern is to be re-used multiple times as a persistent request then the init call (which should itself be collective) could take the time to auto-tune a communication schedule best suited to the specified parameters. Stephen -- ====================================================================== |epcc| Dr Stephen P Booth Principal Architect |epcc| |epcc| [email protected] Phone 0131 650 5746 |epcc| ====================================================================== -- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.