Mathematica 9 is now available
Services & Resources / Wolfram Forums / MathGroup Archive
-----

MathGroup Archive 2011

[Date Index] [Thread Index] [Author Index]

Search the Archive

Memory leak in CUDALink

  • To: mathgroup at smc.vnet.net
  • Subject: [mg115616] Memory leak in CUDALink
  • From: Emil Hedevang <emilhedevang at gmail.com>
  • Date: Sun, 16 Jan 2011 05:54:56 -0500 (EST)
  • Reply-to: comp.soft-sys.math.mathematica at googlegroups.com

Hello,

I am playing around with CUDALink on my MacBook Pro and my Mac Pro. I am experiencing memory leaks in many of the built-in functions like CUDAFourier and CUDADot, see the code below.

First some information on the computer and the GPU.

$Version
Needs["CUDALink`"]
CUDAQ[]
CUDAInformation[1, "Name"]
$CUDADevice = 1;

Out[1]= "8.0 for Mac OS X x86 (64-bit) (November 13, 2010)"
Out[3]= True
Out[4]= "GeForce 9600M GT"

Then we make a large matrix, approximately 32MB in memory footprint.

n = 11;
a = RandomReal[{0, 1}, {2^n, 2^n}];
ByteCount[a]

Out[8]= 33554600

Finally we alternate between querying how large a percentage of the computer's memory the MathKernel is using and calculating the product of the matrix with itself. As you can see, the percentage increases steadily and if I continue in that fashion the computer will eventually begin to swap and become really slow.

RunThrough["ps -p " <> ToString[$ProcessID] <> " -o pmem", ""]
c = CUDADot[a, a];
RunThrough["ps -p " <> ToString[$ProcessID] <> " -o pmem", ""]
c = CUDADot[a, a];
RunThrough["ps -p " <> ToString[$ProcessID] <> " -o pmem", ""]
c = CUDADot[a, a];
RunThrough["ps -p " <> ToString[$ProcessID] <> " -o pmem", ""]

Out[9]= 1.7
Out[11]= 5.9
Out[13]= 8.8
Out[15]= 11.4

I should add that the same problem occurs on my Mac Pro with an Nvidia GTX 285 as GPU.

Do any of you expecient similar problems?

Best regards,
Emil Hedevang


  • Prev by Date: Washington DC Area Mathematica Special Interest Group
  • Next by Date: Re: Caution on BibTex references with Mathematica
  • Previous by thread: Washington DC Area Mathematica Special Interest Group
  • Next by thread: log expression not simplified