Unified Memory On TX1

Autonomous Machines Jetson & Embedded Systems Jetson TX1
1

I’m using ‘Unified Memory’ on tx1 to get a better IO performance. Normally it works fine and also I can get a good performance on memory operations. But once I requested too much memory using cudaMallocManaged, I got some errors. When changed the memory allocation method with cudaMalloc for some memory, the error disappeared.

The error sometime like: ‘Unlunched kernel error or segment error’

  1. How many Unified memory can I get through cudaMallocManaged API on Jetson TX1? Could I have some tools to monitor the memory or cuda related resource limitation and its usage ?

  2. How to get more detailed error information in such case on tx1? (I tried cuda-gdb but it can not work, conflict with other application)

  3. Do I need cudadevicesync to guaranteed cache consistency when using Unified memory? (I lunch the cuda kernel with stream and there is cudaStreamSynchronize after the cuda kernel)

2

Hi,

1. Guess that you meet a known issue which is already fixed in JetPack3.1.
Here is the detail and solution for your reference:
https://devtalk.nvidia.com/default/topic/1013464/jetson-tx2/gpu-out-of-memory-when-the-total-ram-usage-is-2-8g/post/5172688/#5172688

2. Please run your application with cuda-memcheck.

cuda-memcheck [app]

3. Unified memory on Jetson requires exclusive access.
Please remember to call cudaDeviceSynchronize() to make sure the memory is available for CPU.
http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-gpu-exclusive

Thanks.

3

Hi, AastaLLL

Thanks for your answers.

I’ll try to upgrade our system and recheck.

Good tools !

In such case, if I changed the output memory with cudaMallockHost and I don’t touch any memory alloced by cudaMallocManaged, do I still need a cudaDeviceSynchronize()? and how about the IO performance now?

Thanks.

4

Hi,

cudaDeviceSynchronize() is required when touching unified memory with CPU after kernel execution.
The coherence is automatically handled by GPU driver. Usually user can ignore the IO transfer.

Here is a tutorial of unified memory for your reference:
[url]https://devblogs.nvidia.com/unified-memory-cuda-beginners/[/url]

Thanks