BSP Skews A Synchronous Step in Time
•Consider
a series of steps
•
•
•
•
•Enough
operands must be shipped on step n to enable enough threads at
computation step n to cover the latency of the store of n-1 and the fetch of n+1
Load: n+1
Load: n
Load: n+2
Comp: n+1
Comp: n
Comp: n+2
Store: n+1
Store: n
Store: n+2
Store: n-1
Store: n-2
Comp: n-1
Notice that the strategy ignores
locality
A parallelogram represents many threads at the same step &
phase