Running BarraCUDA on Amazon AWS GPU

[ajminich]

Has anyone run BarraCUDA on an Amazon AWS HVM with dual NVIDIA Tesla M2050 GPUs and gotten it to work?

The main problem I'm experiencing at the moment is the maximum memory: BarraCUDA expects at least 675 MB for buffering, but the 2050 only has enough for a 390 MB buffer. The README.FIRST document in the BarraCUDA 0.6 beta release mentions a work-around by changing the SEQUENCE_TABLE_SIZE_EXPONENTIAL constant from 23 to 21 and recompiling, but the result is still an error message claiming insufficient memory. Anyone have a better workaround?

Thanks ahead of time,
AJ

[ajminich]

Workaround Fix

It turns out that the following block (starting at line 2731 in bwtaln.cu) is the issue:

if(mem_available>>20 < (CUDA_TESLA >> 1)) 
{
    fprintf(stderr,"[aln_core] Not enough memory to perform alignment (min: %d).\n", CUDA_TESLA >> 1);
    return;
}

Commenting out this block and recompiling seems to work.

[brianlamx]

Hi, it might be useful to disable ECC so that there is more space for the buffer too.

And if you use the G37 reference from 1K genome project you might be able to get away from doing the 23 to 21 trick. We will test this early next week and let you know what the outcome is.

Brian

[ajminich]

Thanks Brian. I will experiment with disabling the GPU's ECC and see what happens.

[brianlamx]

Have fun aligning

I can now confirm that you can use a C2050 or even a Geforce GTX580 (3GB) to align reads to the Human Genome without altering the parameter in the code. The only requirement is that the ECC has to be disabled and also the reference genome is the 3.1GB version of the V37 assembly. This can be downloaded from the 1000 genomes project from http://www.1000genomes.org/analysis

Brian