writing to an existing variable 9% slower than clearing + creating it again? Generating new one 10 times as fast!?

Hi,

I know preallocation is primarily meant to save overhead and fragmentation which occurs when resizing a variable but I've stumbled upon something strange regarding large variables that i would call an inverse preallocation behavior.

take this example:

clear; 
tic;  b = zeros(35000); t1 = toc;
tic;  b = zeros(35000); t2 = toc;

t1 = 0.0097
t2 = 0.1018

the same happens to gpuArrays. So generating a new variable is significantly faster than writing to an existing one. It applies to all functions or operations I can come up with. Clearing the variable usually takes less time than actually writing to it, however the combined time of clearing and generating a new one on average seems to be exactly how long it takes to write to a preallocated one.

I was wondering if this behavior was normal and/or if there was anything one could to speed things up. Again, it's a factor of 10. Ideally there was a way to always have it the faster way.

For large loops, if I can afford the memory penalty, I now actually go and open several variables instead of overwriting one.

no kidding, this is sometimes significantly faster than re-using the same variable. The code looks awful then, sometimes manually iterating within loops (opening new variables) is faster than just looping over all of it. Same applies to parfor loops and gpuArrays...

In the end one can gain 10% or so by doing the opposite of preallocation, actually clearing the variables one uses in each loop. Example:

clear;
t1 = 0; 
t2 = 0;

a = single(rand(16000));
a = gpuArray(a);

for i = 1:200
    tic; 
    b = sinc(a+rand(1)).^2;
    clear b;
    t1 = t1  + toc;
end
t1

Either for CPU or for GPU the time is always about 7-10% faster if b gets cleared in the loop. Only for small variables, the opposite is true. Please tell me I'm missing something, I don't wanna code like this but saving significant time out of week long simulations ... sure, I'm inclined to do it :-/ I would prefer though that matlab under the hood offered me optimal performance.

my system spec is: Intel 9900k running Windows 10 (all updates) and r2018b, 64GB RAM and an nvidia Titan RTX

regards, Arnold

[G A]

What is the size of free RAM on your system? May be you are overloading it by big arrays.

[James Tursa]

b is over 9GB. So your RAM availability will be critical here. Having said that, smaller variables seem to have this behavior as well.

[arnold]

64GB which should be plenty, always more than half empty.

[arnold]

no more ideas? Anyone from mathworks?

[Walter Roberson]

Note that with gpu in play you should be using gputimeit()