Memory allocation : strange behavior

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

Yesterday I posted a topic about a memory allocation problem.
The problem is the following : I do 4 cudaMalloc. The sum of the memory allocated is function of 3 variables N, M, and D. What I do is to use my function with different values of M (actually from 256 to 3072). N=38400 and D=96.

Before, when I used M=2560, the total memory allocated was 410MB, and among the 4 malloc, one tryied to allocate 394MB. In this case, this malloc failed in spite of that the free memory is 750MB.

Now, I use the value returned by cudaMalloc to manage errors during memory allocation. Moreover, I do the biggest memory allocation in the beginning. With these small modifications, the memory allocations seem to work well.

I don’t understand why memory allocation failed.

Do you think that the order of cudaMalloc can change anything?
Do anybody have this kind of problem?

Thanks
Vince

2

Having free memory and Having continously free memory are two different things. There should be a better interface than “cuGetMemInfo” or whatever… to find out whats the maximum free-chunk size available. Probably there is one… I dont know.

ANd, certain things could depend on the memory allocator’s algorithm itself. All depends on the NVIDIA driver’s memalloc strategy.

3

Hum… I think you are right. 400MB is more than half memory size. It seems to be logic that it is hard to have this amount of free linear memory after 2 cudaMalloc (respectively of 15MB and 1MB).
Thanls for your helpfull answer!!!
Vince

4

Thanks for your words. But I still think, if you have 700MB free – I would expect 400MB of it being linear… Hmm… The answer really depends on the memory-allocator’s algorithm.

5

Actually, I do this

   // Allocation CUDA memory

    cudaMallocPitch( (void **) &ref_dev    , &ref_pitch    , ref_width*sof   , ref_height   );

    cudaMallocPitch( (void **) &query_dev  , &query_pitch  , query_width*sof , query_height );

    cudaMallocPitch( (void **) &dist_dev   , &dist_pitch   , query_width*sof , ref_width    );

    cudaMallocPitch( (void **) &output_dev , &output_pitch , query_width*sof , 1            );

These 4 malloc are the first things done after the variable initialisation.

As you can see, it’s very simple… The cudaMalloc which cause the error is the 3rd one (using dist_dev).