Big multidimensional vector initializaton C++

Question

I am trying to load simulation results from a file into a vector. The code works well for small and medium amount of data. When I try do load big files I get an exception. The array that makes my program crash is about 17000*7*3000 elements. I've tried splitting initialization into few steps but it has crashed as well. Could you please tell me if I can do something to make it work?

//////////////////////////////////////////////////////////
//Import of surface receivers file
    for(int freq=0;freq<7;freq++)
    {
        filePath=filePath_old;

        filePath.Replace(wxT("125 Hz"),wxString::Format(wxT("%i"), freqSTI[freq])+wxT(" Hz"));
        if(importer.ImportBIN(filePath,mainData))
        {
            if(timeTable.size()==0){
                for(int idstep=0;idstep<mainData.nbTimeStep;idstep++)
                {
                    timeTable.push_back(mainData.timeStep*(idstep+1)*1000);
                }
            }
            for(wxInt32 idrs=0;idrs<mainData.tabRsSize;idrs++)
            {
                for(wxInt32 idface=0;idface<mainData.tabRs[idrs].dataRec.quantFaces;idface++)
                {
                    if(tab_wj.size()<idrs+1){
                        tab_wj.push_back(std::vector<std::vector<std::vector<wxFloat32> > > (mainData.tabRs[idrs].dataRec.quantFaces,std::vector<std::vector<wxFloat32> >(7,std::vector<wxFloat32>(mainData.nbTimeStep,0.f))));
                    }
                    //Pour chaque enregistrement de cette face
                    int tmp=mainData.tabRs[idrs].dataFaces[idface].dataFace.nbRecords;
                    for(wxInt32 idenr=0;idenr<mainData.tabRs[idrs].dataFaces[idface].dataFace.nbRecords;idenr++)
                    {
                        t_faceValue* faceval=&mainData.tabRs[idrs].dataFaces[idface].tabTimeStep[idenr];
                        tab_wj[idrs][idface][freq][faceval->timeStep]=faceval->energy;
                    }
                }
            }
        }

    }

The place where an exception occures is tab_wj.push_back...

When I tried to split initialization into more than one step I wrote:

std::vector<wxFloat32> t1(mainData.nbTimeStep,0.f);
std::vector<std::vector<wxFloat32> > t2(7,t1);
std::vector<std::vector<std::vector<wxFloat32> > > t3(mainData.tabRs[idrs].dataRec.quantFaces,t2);

Then it crashed at last one of these lines.

Thanks for any help!

Answer 1

Looks like it is getting an exception while allocating memory. std::vector<> must reallocate its storage as you add new items to it. It doesn't do this for each push_back() because it allocates more than it needs. A vector<> of size 5 might actually have space for 8 items (size() vs. capacity()). When you insert the 9th item it must reallocate, and typically it will double its capacity (from 8 to 16 (*)).

This reallocation can fragment your heap. To improve you can preallocate the vector to the size it will eventually need, so it will never need to reallocate. vector::reserve() So, if you know you will eventually need 17000*7*3000, reserve this up front.

Also, if you have a lot of small vectors, allocating those at a small size can reduce your memory requirements. Usually vector<> has a default size which could be 8 or 16. If you only ever insert 3 items, then you are wasting a lot of storage. During allocation of the vector I believe you can specify its initial capacity.

Since you are using nested vectors, each of these will in turn have their own allocations. Depending on your version of C++, pushing a vector onto a vector can cause a number of allocations.

You might try restructuring the code to not require nested vectors.

(*): a naive vector implementation will double the allocation size when it needs more space, but a factor of 1.6 is better for a number of reasons (its just as big-O efficient and leads to much less memory fragmentation). See "The golden ratio": https://en.wikipedia.org/wiki/Golden_ratio