High Performance Computing 1
Parallel MG
•For Full Multigrid, assuming n^2 >= p, so that each processor has at least 1 grid point:
•Time = O( N/p + log(p)*log(N) )*f
•            + O( (log(N))^2 )*alpha
•            + O( sqrt(N/p) + log(p)*log(N) )*beta
•
•where N=n^2 is the number of unknowns. The speedup over the serial floating point work O( N ), is nearly perfect, O( N/p + log(p)*log(N) ), when N>>p, but reduces to log(N)^2 when p=N (the PRAM model).