Why is populating a square matrix faster than populating a long vector with the same number of elements?

[Joro Kolev]

Good afternoon,

It is a lot faster to populate a 100 X 100 matrix, than to populate a 10000 X 1 vectors. Why is that?

The code:

clear all

timer on 1

matrix A = J(10000,1,.)

forvalues i=1/10000 {
    matrix A[`i',1] = 1
}

timer off 1

timer on 2

matrix B = J(100,100,.)
forvalues i = 1/100 {
    forvalues j = 1/100 {
        matrix B[`i',`j'] = 1
    }
}

timer off 2

timer list

results in

. timer list
   1:      0.33 /        1 =       0.3340
   2:      0.03 /        1 =       0.0310

So populating the square matrix is 10 times faster.

[William Lisowski]

timer on 11
matrix A = J(10000,1,.)
timer off 11

timer on 12
forvalues i=1/10000 {
    matrix A[`i',1] = 1
}
timer off 12

timer on 21
matrix B = J(100,100,.)
timer off 21

timer on 22
forvalues i = 1/100 {
    forvalues j = 1/100 {
        matrix B[`i',`j'] = 1
    }
}
timer off 22

timer list

results in

. timer list
  11:      0.46 /        1 =       0.4580
  12:      0.19 /        1 =       0.1860
  21:      0.00 /        1 =       0.0000
  22:      0.17 /        1 =       0.1750

So creating the square matrix is infinitely faster, apparently.

[Mike Lacy]

I get similar results with your code using v. 15.1. I would hazard a guess that this is related to some issues of how matrices are allocated that came up in discussions several years ago about how a larger -set matsize- value can badly impair performance, at least in earlier Stata versions. See:

https://www.statalist.org/forums/for...mp2-vs-v-12-ic

https://www.statalist.org/forums/for...ng-set-matsize