36
SUMMA Algorithm Analysis
The flood is more expensive than l time, but less that l(log
P) ... probably much less
[1..m,1..p] C := 0.0; -- Initialize C
for k := 1 to n do
[1..m,*] Col :=
>>[ ,k] A; -- Flood kth col of A
[*,1..p] Row :=
>>[k, ] B; -- Flood kth row of B
[1..m,1..p] C += Col*Row; -- Combine elements
end;
SUMMA does not require as much comm or data motion as Cannon’s, nor does it “touch” the array as much