Next Meeting: Thursday April 13th, 2000 at 11AM (CIT will call usual number).
Software Status:
Masha Barnes begain by reporting that she had finised all the mpi send code to send out input data to the slaves with the exception of support for the history structure. This will be added soon. She has not yet worked on receiving data back from the slaves. The UWM group spent the week working on issues of templates within templateFilters().
Questions & Answers:
UWM had some questions on the modification of input data by the functions in the dynamically allocated library. Kent remarked that the functions should not modify the original input data, but that the structure allowed for appending new data. Another question dealt with the using of the static structures; Specifically the memory management of these. Kent remarked that the memory for these could be freely created by the dll and that hte freeFilter() would be called once at the end to de-allocate this memory. No pointers would be passed into freeFilters(), it would use internal variables to keep track of the memory allocated. This was an acceptable approach to everyone present. Another question was how do the slaves know which template to process. Kent explained this as being handled by the index built by indexFilters(). The master would just step through these sequentially asking each slave to process the next N where N is decided by a command line argument.
Second Inner Loop:
UWM is still thinking about how best to use this loop. It was decided that this should also be a while loop though. UWM will send out some pseudo-code illustrating a model for its use.
Load Balancing:
The topic of load balancing came up. There is a possibility that some algorithms running on a single node will trigger a secondary computation which will be much more demanding than the trigger algorithm. To handle load balancing in this enviroment, several things were discussed. One being to have the node send a signal to the master informing it that it was going to be busy. This would pose an extra listen loop on the master which may or may not be used in all dll's. It was also proposed that the node that was triggered into this more demanding computation perform this in a parallel manner using either existing nodes or some reserve of nodes, all of which had the data pre-positioned by the wrapperAPI.
Schedules:
The subject of schedules will be revisited next meeting once everyone has had a chance to settle down into a code writing mode and can make more accurate projections on time for various functions.
Organizational Details:
Kent took minutes.
Next weeks meeting time will return to the 11AM PDT time slot.
Agenda for next Week: