if you will never need any kind of row-wise operation, then probably vector of vectors is best, as this will make your broadcasting life easier rather than having to use eachcol or mapslices constantly.

Yes, the (inner) vectors are of the same length. I do use a lot of broadcasting so I find the vector-of-vectors structure easier. I tried matrices but then found myself having to work with Slices and got scared.

A noteworthy difference between matrix and vector of vectors is memory locality: in a matrix, all of your data will be stored contiguously in memory, which may speed up access

On the flip side, if youâ€™re adding, removing or swapping columns, then you probably need a vector of vectors. It depends on your use case, and locality of your access pattern. E.g. if you need coulmns i and N-i simultaneously, it might make sense to use a vector of vectors.

as this is the natural datastructure for your problem description. Why create extra entities unnecessarily?
If you have other requirements, like keeping all â€śvectorsâ€ť close in memory, you can later switch to a matrix or whatever.

I realized that in trying to present a simplified version of the problem I might have lied a little. I am working with two types, one stores data in vectors-- the v1, v2 I mentioned originally. The other stores them in 2-column matrices (the data comes in pairs). These I can not change. So I should probably go with the vector-of-vectors. (I could use matrices for the original problem I posed if it really makes sense to do that.)

My inner vectors/matrices will not need to be re-written once created and will almost always be read in sequence (v1 then v2, or v2 then v1). So it makes sense to have them in roughly the same place in memory. I didnâ€™t realize vector-of-vectors would not store things around the same place in memory. Maybe it depends on how they were created. Two questions:

How can I check whether the inner vectors in my vector-of-vectors are contiguous in memory?

Is there a way to ensure that they are? Would a deepcopy of my vector-of-vectors do the trick (even if it is a hack)?
Iâ€™m not concerned about the cost of creating the vector-of-vectors. Iâ€™m more concerned about the speed/ease of reading from it once created.

I am more of a researcher than a computer scientist. It might not turn out to make much difference in the end-- weâ€™re talking about reading vectors or 2-column-matrices of length no more than 100,000 per second. But Iâ€™ve also not worked on something like this before. Iâ€™m trying to think more about design early on.

I donâ€™t know of any easier way than to use a matrix, precisely because in Julia they are stored in column-major order. So pretty much by definition, a matrix is a â€śvector of vectorsâ€ť stored contiguously.

But once you have your matrix M, be sure to use @view M[:, i] instead of M[:, i] to access the i-th column, otherwise the data will be copied.