Catch std::bad_alloc of a vector constructed inside a member initializer list

Hi!
I am dealing with a problem in which I have to allocate a huge distance matrix (~600 GB on a cluster node). In order to save some memory, since the matrix is symmetric, I defined an upper triangular class as follows:

Code:

---

class utriag_matrix {
    uint32_t dim;
    std::vector<double> buffer;
    public:
    utriag_matrix(uint32_t d): dim(d), buffer( ( (d*(d-1))/2 ) ,0 ){}
   
    uint32_t size() {return dim;}

    double& at(uint32_t i, uint32_t j)
    {
        if(i==j){ throw std::invalid_argument( "at(i,j) with i == j is not a valid element of utriag_matrix." );}
        if (i>j){std::swap(i,j);}
        return buffer[dim*i - (i*(i+1))/2 - i + j - 1 ];
    }
};

---

Note that the internal buffer is an std::vector.

The problem I think I am having is that the system is not able to allocate the std::vector even though I have more than enough RAM (the node has 1.5 TB of RAM so I am not even using half of it).

Is there a way of catching a std::bad_alloc of a vector constructed inside a member initializer list?

Is it a problem to allocate an std::vector of hundreds of GBs if my RAM is big enough (this is the first time I work with this amount of RAM)?

Thanks in advance!

Quote:

---

Originally Posted by Fede

Is there a way of catching a std::bad_alloc of a vector constructed inside a member initializer list?

---

Yes, you could use a function try block. However, in the case of a constructor, the catch block must always throw an exception, so you cannot really recover from an exception propagated by the member initialiser list: it is always a constructor failure; the function try block is more for stuff like logging or translating the exception.

Quote:

---

Originally Posted by Fede

I have more than enough RAM (the node has 1.5 TB of RAM so I am not even using half of it).

---

That is an impressive amount of RAM.

Quote:

---

Originally Posted by Fede

Is it a problem to allocate an std::vector of hundreds of GBs if my RAM is big enough (this is the first time I work with this amount of RAM)?

---

One possibility is that you may have more than enough memory available in total, but you might not have enough memory available contiguously.

Substitute your value of d.

Code:

---

#include <iostream>
#include <vector>

int main() {
  size_t d = 1000;
  std::vector<double> buffer( ( (d*(d-1))/2 ), 0 );
  std::cout << "Size=" << buffer.size() << std::endl;
}

---

Does it work?

Thanks for the replies.

Quote:

---

Yes, you could use a function try block. However, in the case of a constructor, the catch block must always throw an exception, so you cannot really recover from an exception propagated by the member initialiser list: it is always a constructor failure; the function try block is more for stuff like logging or translating the exception.

---

I don't care about not recovering from the exception but I would like to have warnings printed to std::cerr. The implementation would be to include the try block inside the member initialization list?

Regarding Salem's reply, while checking the code I did notice that uint32_t for the dimension could be the problem. I had already launched a set of benchmarks which includes this suggested modification among others.
I'll know the results in a few hours although now that I read your reply I am almost certain that this was the problem. I'll update the thread once I know the result of the tests.

Once again thanks for the help!

> I am dealing with a problem in which I have to allocate a huge distance matrix (~600 GB on a cluster node)
Yeah, uint32 is nowhere near large enough for doing calculations up to 600GB.

Effectively the problem was the uint32 was overflowing when calculating the buffer size: (d*(d-1))/2 ) with d~500000
A mistake I do constantly :D
Thanks for the replies!