Memory required for struct array

In this answer https://www.mathworks.com/matlabcentral/answers/2010767-struct-function-in-matlab

Walter Roberson gives some number about memory used by struct

it seems to me the numbers derived by Walter are not correct.

In this thread I attempt to fix, mostly by reverse engineering. Note that this are not documented and possibly depend . on the MATLAB version. Currently it's R2023a.

So there is no question explicitly formulated here

Answer 1

[Bruno Luong]

The function structbase_bytesize_fun that computes the "base" memory of the struct s is as following:

fieldname_bytesize = namelengthmax() + 1; % == 64
pointer_bytesize = 8;
mxArray_byteSize = 96;

structbase_bytesize_fun = @(s) length(fieldnames(s)) * ...
                           (fieldname_bytesize + pointer_bytesize*numel(s));

When the struct array is allocated without specify meaningful field values, the field values are affect to [] and no more memory is allocated.

s=struct('f1',cell(1,10),'f2',cell(1,10))

s = 1×10 struct array with fields:
    f1
    f2

whos s

  Name      Size            Bytes  Class     Attributes

  s         1x10              288  struct              

structbase_bytesize_fun(s)

ans = 288

When a field of a structure is set by a rhs the byte size of the rhs must be added

rhs = (1:3);
rhsinfo = whos('rhs');

s(end).f1 = rhs;

whos s

  Name      Size            Bytes  Class     Attributes

  s         1x10              408  struct              

struct_bytesize = structbase_bytesize_fun(s) +  rhsinfo.bytes + mxArray_byteSize

struct_bytesize = 408

rhs2 = 1:4;
rhs2info = whos('rhs2');
[s(1:5).f2] = deal(rhs2);

whos s

  Name      Size            Bytes  Class     Attributes

  s         1x10             1048  struct              

struct_bytesize = structbase_bytesize_fun(s) + ...
    1*(rhsinfo.bytes + mxArray_byteSize) + ...  % s(end).f1
    5*(rhs2info.bytes + mxArray_byteSize)       % s(1:5).f2]

struct_bytesize = 1048

NOTE that on top of that memory seen by whos command, one need to count the memory taken header of s itself, which is hide by whos, meaning 96 bytes (mxArray_byteSize) needs to be added on top of that struct_bytesize.

It is not clear to me if shared data is correctly count (reduced memory), for example

[s(1:5).f2] = deal(rhs2); 

would s(1:5).f2 share one copy of the rhs somewhere?

[Bruno Luong]

Of course once you have the structure built, it's better to use whos to get the memory occupied.

The formula is useful to predict memory requirement without building the structure.

[Walter Roberson]

clear s
s(100).foo = [];
whos s

  Name      Size             Bytes  Class     Attributes

  s         1x100              960  struct              

clear s
s(99).foo  = [];
whos s

  Name      Size            Bytes  Class     Attributes

  s         1x99              952  struct              

960 - 952

ans = 8

so the number of extra bytes for each numel() must be 8, not 16.

[Bruno Luong]

OK actually the number of extra bytes is

8 * (numberoffields * numel(s))

The fields data pointers are 2D array.

I fix my code.

[Walter Roberson]

At the moment, I do not immediately see a difference between your size predictions and my previous enumeration?

My discussion was slightly sloppy in taking about sizes for additional entries, under the assumption that at least one entry was assigned to for each fieldname; it probably would have been better to rewrite them to a base of no entries written to. On the other hand, when no entries at all are allocated to a field, it looks like the 8-bytes-per-element pointer table is not allocated at all.