Skip to content
/ Silo Public

SILO: fast saving, loading and operations on vectors of objects in MATLAB

License

Notifications You must be signed in to change notification settings

sg-s/Silo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SILO: fast saving, loading and operations on vectors of objects

View Silo on File Exchange

Problem description

Often, we want to write a class to enforce a certain data structure and to associate bound functions (methods) to that object. For example, here is a simple class that defines some object:

classdef ScalarClass
properties

	X (1,1) double = 0
	Y (1,1) double = 0
	Z (1,1) double = 0

	Type (:,1) string = "A"

end % props

end % classdef

Now that we have our nice object, we often want to create vectors of objects, because, that's just nice. We can do something like this:

N = 5e3;
A = repmat(ScalarClass,N,1);

This is great. However, there is a big problem: loading this vector of objects from disk is extremely slow, because MATLAB has to type-check every element and every property on load:

tic; save('as_class.mat','A');
load('as_class.mat','A'); toc

Elapsed time is 2.237674 seconds.

which is awful.

In fact, even saving and loading vectors of structs is pretty slow:

for i = length(A):-1:1
	B(i) = struct(A(i));
end
B = B(:);
tic
save('as_struct.mat','B');
load('as_struct.mat','B'); 
toc

Elapsed time is 1.302606 seconds.

1.3 s to just save and load a structure array that contains the data in those objects! And we're not even counting the time taken to reconstitute those objects. Is there a better way?

A single object with vectors

Why yes there is. If we instead rewrite our class to look like this:

classdef VectorClass
properties

	X (:,1) double = 0
	Y (:,1) double = 0
	Z (:,1) double = 0
	Type (:,1) string = "A"

end % props
end % classdef

we see we explicitly allow for arrays of variables. However, this creates a new set of problems:

  • there is nothing forcing X to be the same length as Y, etc.
  • We've lost the ability to index into the array: we could do A(1), but we can't really do that in this new architecture.

Oh well, let's soldier on. How performant is this architecture for saving and reading from disk?

V = VectorClass;
V.A = zeros(N,1);
V.B = zeros(N,1);
V.C = zeros(N,1);
V.Type = repmat("A",N,1);

tic; save('V.mat','V');
load('V.mat','V'); toc

Elapsed time is 0.011277 seconds.

That's a 200X speedup! We can't throw that away...

Let's instead focus on trying to get this new architecture to behave like the first one.

The solution: Silo

One solution is the Silo abstract class, which allows us to write a new class to encapsulate our data that looks like:

classdef SiloClass < Silo 
properties

	A (:,1) double = []
	B (:,1) double = []
	C (:,1) double = []
	Type (:,1) string  {mustBeMember(Type, ["A","B"])} 

end % props
end % classdef

So it's identical to the previous class, except it inherits from Silo.

Because it's identical to the previous class, it's still extremely fast at writing to disk and loading objects from disk:

S = SiloClass;
S = S.add(struct(V));

tic; save('S.mat','S');
load('S.mat','S'); toc

Elapsed time is 0.011494 seconds.

But it solves the problems with the previous solution. Specfically:

Indexing

S contains 5000 records, as we can see:

S

S = 

SiloClass with 5000 elements and with properties:
    'A'
    'B'
    'C'
    'Type'

but we can index into it as though it were a vector of objects:

S(1)

ans = 

  SiloClass with properties:

       A: 0
       B: 0
       C: 0
    Type: "A"

Data structure

All fields of classes inheriting from Silo are forced to be the same length. If you modify or create Silos using the add method, it is impossible to end up in a situation where one field is longer than another.

About

SILO: fast saving, loading and operations on vectors of objects in MATLAB

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages