Re: more fun with proposals

Jarek Nieplocha (j_nieplocha@pnl.gov)
Thu, 11 Jul 1996 15:06:44 -0700

>If you have many non-overlapping regions, then by definition they can be spilt
>into seperate windows. If they do potentially overlap, then the application
>must have some way of guaranteeing mutual exclusion, which is where the locks
>and barriers come in.

Eric,

For example, splitting is not feasible if:

. you have a large region ( hundreds of megabytes) that you access in small
sections (kilobytes or less) possibly with variable size and shape

. you cannot decompose the region into multiple windows because
you don't know in advance how (might depend on dynamic load balancing)

At the same time, the application is able to guarantee that there
is no concurrent overlapping acceses, for example due to the physical nature
of the problem -- "if I work on molecule A, then there is no any other
process that could possibly update the array element potential(5, 9001, 6107)",
and is not willing to pay the cost of locking.

Jarek