Accelerating access of struct members

Dear all,

I noticed that accessing data within a struct seems to be very slow compared to an array. To illustrate this, I put together a small script that randomly accesses data within an array and a struct respectively:

M = 20;
N = 1e8;
data = rand(M, 1);
data_struct = struct('a', num2cell(data));
for n = 1:N
  i = round(M * rand() + 1 / 2);
  c = data_struct(i).a;
  e = data(i);
end

Profiling this script reveals that the assignment of c consumes about 20x the time of assigning e. Is there any way to accelerate the access of the struct, i.e. by somehow specifying the type of data it contains?

Answer 1

[Sujit Muduli]

Hi Torsten,

Profiling might not give you the exact result but you may use tic-toc to measure the exact time taken. Please note, using structure arrays is about 2 times slower than the simple matrix in MATLAB. I also did a small experiment to profile the time taken in the assignments in my MATLAB R2017a. Profiler usually gives you relative time values, not the exact this helps you find the bottleneck in your program. In this case, it is doing its job well by hinting the bottleneck is in the access of the struct values.

Thanks

Sujit

[Torsten Knüppel]

Thanks for the reply - it really helped interpreting the profiler's results.

Answer 2

[Jan]

Matlab's JIT acceleration can improve te execution speed of e.g. loops massively. During debugging and profiling the JIT is disabled, because it can reorder the commands to increase the speed, but this would confuse the examination of the code. Therefore the power of the profiler is limited, unfortunately.

Of course accessing a field of a struct takes time, because there is a substantial work to do in each iteration. Storing data in structs help to write clear code, but it reduces the execution speed compared to arrays.

[Torsten Knüppel]

Good to know what causes the differences between the profiler's results and what tic-toc yields.