I hope that programmers is the correct stack exchange for this, as it is quite a concept based question.

I'm working on a data structure in C++ that is a represents data in 3D space. The x-y plane is large (say 0 - 10 000) while the z-axis has a very small range (0 - 10). All the data is located on integer (x,y,z) points. However there can be more than one than one data point on an (x,y,z) coordinate.

I have thought of three ways of approaching this subject, and don't have the necessary experience to choose between them.

  1. Using a 2D linked list inside vectors (dynamic arrays) vector<vector<structureA*>> Data where the first list item for that (x,y) point can be accessed using Data[x][y]. structureA contains an its z value, its actual data, as well as a pointer to the next structure in the list (that may, or may not have a different z coordinate, but is located on the same line defined by x and y). Accessing elements at the middle and end of this list is slow.

  2. Using a three dimensional array/vector: vector<vector<vector<structureB>>> Data. In this case the first two vectors define the x and y coordinates (as before) and the third vector gives us an array of all of the elements on that line defined by x and y. We can quickly determine how many elements are on this line (by querying the third vector) and can iterate from the back and the front. structureB contains, the point's z value and its data. This seems like a fairly good and safe (regarding memory allocation) solution to me.

  3. Using a three dimensional array/vector and a linked list. In this case the data could be accessed as such: Data[x][y][z] and would point to the first element of a linked list of data that resides at this coordinate. The structure pointed to would not contain its z value here as this is already dealt with in its address in the array. However, many (x,y,z) points have no data, whereas most of those that do have data have 10+ data structures. (All lines defined by a specific (x,y) contain at least one structure). This seems the worst solution to me as in many cases I would have to iterate through empty z coordinates, before hitting a linked list to iterate through.

The emphasis in this program is speed, my data set is less than 100MB on a modern computer, so even an inefficient method of storing it in RAM is fine. Option 2 seems the best to me, although in the past I've normally seen people more inclined to go with option 1. I'm interested to hear your opinions on the best way to handle this along with your reasoning. Many Thanks.


Thanks for the responses, I stand to learn quite a lot from your comments as I am a hobby programmer, having taken only one short course on the fundamentals of C a few years back.

@MorphingDragon The array is not jagged, all z "slices" have the same range of x and y. The values also don't seem sparse enough for a hash table (which IIRC slow down significantly on hash collision - which is guaranteed as multiple elements have the same (x,y,z)). The dimensions of the set are known on file load (they are included in the file header along with an md5 hash of the file for error checking and a few bit relevant to the analysis).

@J Trana / Florian F The data set is pre-built and is infrequently modified by the program - a piece of data may need to be moved or deleted. For adjusting data particularly just adjusting its z-value (or removing it), which are the most frequent operations. The linked list will be faster as these operations are O(1) as opposed to O(n) in a vector. However, I believe these changes are infrequent enough that having O(n) for changes may be worthwhile if the lookup speed is even marginally faster (as this is the more frequent operation).

The majority of the time the data will be accessed iteratively, taking an n*p slice from the x-y plane. Pseudocode example:

for each y between startY and (startY + p)
  for each x between startX and (startX + n)
    read all nodes in Data[x][y] //gets nodes for each z on an (x,y) line
    analyse(&nodes) //possibly modify the nodes, possibly just compute something based on their values

Data access is always related to the (x,y) line an element is on, but we don't always need all the nodes on that line (dependant on their z). So we either have a structure that doesn't require us to retrieve all of them (just the ones we want) or we perform a secondary sort. The data on file represents a sorted list in ascending z for each (x,y).

I will also look into those other data structures. I've never met them, as with very little training, my method is to use STL vectors or lists for everything. I'd like to spend the time to do this properly this time though.

I spent a fair amount of time thinking about this last night after posting the question and seeing the first responses. I realised that "if one piece of data from a point (x,y,z) is required, then all elements from that point are required at the same time." This has led me to think that option 3, my least preferred when I started this yesterday afternoon, is actually the best of the options I've enumerated. If I use a three-dimensional dynamic array of pointers to linked list, I don't exactly lose much speed or heap space setting a pointer to NULL rather than attaching a struct. This also allows full access to any 3d, which while not currently required may reduce some operations from O(n) to O(1).

Having had a brief look over octaves/r-trees, I feel safe saying that my data set is not complex enough to require such an engineering heavy solution. When it comes down to it, i am fairly sure that any of the three options I outlined would work, even if not the fastest, but I would like to learn a good and relatively efficient method, rather than just relying on the awesome power of modern computers to brute force a problem like I normally do. Thanks for the replies and suggestions so far. :)

  • Are the dimensions jagged? Is the size of each array dimension known and fixed before doing operations over it? The fastest option is a single contiguous piece of memory as using a vector has quite a bit of abstraction. Alternatively if the [x,y,z] values are going to be quite sparse, you could use a hash map?
    – BlamKiwi
    Aug 28, 2014 at 22:49
  • @Goobley - You've started to answer how you'll use this "...in many cases I would have to iterate through empty z coordinates...", "The emphasis in this program is speed...". etc. but I think we're still missing some details. Let's go down the list: space/speed - you choose speed. Is the data set pre-built or will you be updating it often? How are you going to access it: primarily random access by x,y,z; iteratively; or something else? Also, have you looked at octrees, r-trees, or any other spatial data structures yet? ...this sounds like a fun problem!
    – J Trana
    Aug 29, 2014 at 0:02
  • 2
    How do you access the data? Do you sometimes need to iterate over the items for all x, with a given y and z?
    – Florian F
    Aug 29, 2014 at 8:30
  • Florian is correct how you iterate through the data is important.. The second question is how empty is you data set? The data structure you will want to build is very different if it is dense aka mostly or completely full of data versus mostly empty less than 50% of the points have data. A third question is how often will the data set change in size? If it does not change after creation use static native structures as they will reduce your overhead. And you can put all the fancy logic yourself if you need it.
    – Rob
    Aug 29, 2014 at 12:35
  • As for you RAM concerns, if you use memory map files you can access very large data sets while keeping you RAM use low by only putting in memory what you are currently using. Basically it allows you to write your own memory management using a file a back end. (like the swap file for the OS).
    – Rob
    Aug 29, 2014 at 12:37

2 Answers 2


I hope it's ok to answer my own question. I believe I have found the optimal (without overcomplicating the problem) data structure for my problem. There was at least minor idiocy on my part for not recognising this earlier. The data doesn't need to be accessed by (x,y,z) but instead by (x, y, range of z (say 0 - 3)). This give a C++ struct as follows:

struct node {
  struct node *next;
  int zGroup;
  int z;
  50 bytes of misc data };

I can then address this through a 3D dynamic array (vectors):

vector< vector < vector < node* > > > Data;

Any given Data[x][y][zGroup] points to the first element of a linked list, the entirety of which is needed every time one element of it is needed. No value of this array is NULL, every one contains a linked-list of at least one element.

The third dimension of the array - the zGroup has jagged dimensions, but with dynamic arrays this isn't an issue. Given the data and computations being performed on it, I know that the max x and y values are set when the file is read and do not change, neither does the number of z groups on any given (x,y) line, the actual z-values of nodes may change, but they will remain inside the same z-groups, giving a constant-sized, fully populated array.

With the way that the file is structured it is also easy enough to page it in and out of memory if I am brought to do this with much larger data sets.


I don't fully follow what's going from your answer, but a couple of comments: 1. You should use contiguous memory. At the very least, at the outer level (for the x-y coordinates) you should have a single vector, not a vector of vectors. A vector is basically a pointer to a contiguous block of memory. A vector of vectors is a contiguous block of pointers, to contiguous blocks of memory, i.e. the data itself is not contiguous. Contiguity is incredibly important to speed due to factors like cache missing and pre-fetching. 2. Linked lists are slow. The working of modern processors tends to give vectors a very large practical advantage. Linked lists are especially terrible for small types, for several reasons. First off, if the types are small the extra space costs of the pointers is huge. Second, when the types are small and in contiguous storage, the processor can read them very efficiently because processors generally read from memory an entire cache line at a time, which is 64 bytes. If your data type is just a double, that means that 8 doubles are read simultaneously by the processor. When you use linked lists you give this up.

The bottom line: you want a loop over your data to read the data precisely sequentially from memory in one contiguous chunk. This is the model that will give the best performance.

Edit: Ok, I reread your requirements more carefully. Here's what I think is best, however let me preface this: Alexandrescu has a great article on performance where he notes that we have made our processors better and more intelligent, and the cost of a simple mental model of how it works. So it's very hard nowadays to predict the speed of code compared to before. So profiling is more important than ever. So if you really care about speed, profile! (btw, link is https://www.facebook.com/notes/facebook-engineering/three-optimization-tips-for-c/10151361643253920).

Given that your x and y locations will never change, only z, you can create a struct as you suggested that stores x, y, z and misc data. Take all the initial data points, and put them in a single std::vector of these structs, call this variable DataPoints. Sort them by x, then y, the z sorting is optional. If you are doing more operations that change z, maintaining sort order will be expensive and its easier just to loop through all z's and do an if check. If you usually just scan the data, then keep sort order by z, and you will have to pay O(N) to keep it sorted when modifying the z coordinate, where N is the number of points with the same x and y coords.

Now, we are going to build a lookup table. We create another array, a single contiguous piece of memory, that you do lookups on given an x and y coordinate. This table will return a pair of integers, telling you the first and last index in DataPoints that have that exact x and y. Now, there's just one more clever bit. Since we have this lookup table, we don't actually need the x and y coordinates stored in DataPoints at all. So we create a new DataPointsReduced; it's the same data in the same order, but with the x data stripped. We could strip the y data too, but that may or may not pay off as you'll see.

We're now basically done. When you do the loop over x and y, you can extract all the indices you need from the lookup table. Sadly because you do a range of x and range of y, the resulting data of interest can't be fully contiguous. But the scan should be contiguous over a fixed value of x. The outer loop is x, for that fixed x, since the secondary sorting is y, the range of y's of interest all contain data that is contiguous in memory. So there are only two loops, not three. The inner loop goes over a block of memory containing all z's for a range of y's, for a fixed x. Inside this inner loop you do whatever operations you want. Note that if we strip y too and we need it, we will need to return to three loops, or do some kind of lookup to determine y at each stage, which can be slower. If you don't need y at all once the range of y's is specified, then strip y from your data as well.

That's basically it. If your range of x's and y's gets very large, this lookup table will eventually use a huge amount of space (but note: no more space than your solution, this lookup table could probably store 2 32 bit unsigned integers for every x-y pair, whereas your table would store at least one 64 bit pointer for every x-y pair). You can avoid the lookup table at small cost. Since we've sorted by x and y, you can have a small prestep where you do some clever binary searching to quickly find the indices you need, and then proceed as before. The cost of this prestep is less than linear in the total amount of data being scanned, so asymptotically it doesn't affect running time. However in this case you obviously can't strip the x or y data. This shouldn't make a huge difference though, if the number of points in the contiguous y range is reasonable prefetching should nullify this difference.

This setup (with no lookup table) is pretty damn good if the ranges of x and y are huge, and your data is progressively more sparse. Note that this solution does not have any space requirement proportional to the ranges of x and y. Whereas even dynamically allocated arrays or linked lists at every x-y pair implies at least one pointer per x-y pair, which is space proportional to the ranges of x and y.

Final thought. This setup has potentially more if branching, because I basically recommend just looping through a contiguous block of memory and doing whatever checks you need to on z, instead of looping over the exact known range of z. It depends exactly what your code is doing, but for very simple if operations (e.g. 1-2 lines of code, no side effects) often the cost of branching is virtually nil: the processor can compute the if condition and both branches simultaneously on 3 different pipelines, and then move the appropriate result back.

Final final thought: profile! What I wrote here is based on moderate experience, but honestly it is fully possible that my answer will not work out the way I think it will for any number of reasons.

Good luck.

  • Ok, thanks, I've opened the question up again. If I should keep my data in a single vector, then would a sensible method be to define a structure say struct Point {int x; int y; int z}; and then create a vector based on this tuple? That should keep the memory contiguous.
    – Goobley
    Aug 31, 2014 at 19:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.