GPU Direct 1 and 2

Accelerated Computing CUDA CUDA Programming and Performance
1

In NVIDIA GPU direct section:

It appears to me that GPU Direct 2.0 include GPU Direct 1.0 features ( it said “adds support” ).

Is this true ?

  • If yes, How could I test my MPI to see whether the feature of GPU Direct 1.0 is there ? I couldn’t figure out anything from the example code i.e mpi_pinned.c in the package.

  • If no, Is it possible to install the driver in the website together with CUDA 4.0 driver? Because the file name “nvidia-gpudirect-3.2-1.tar.gz” suggests driver 3.2.

Thank you

2

Could anyone enlighten me on this issue ? Thanks a lot

3

P2P has nothing to do with the old GPU direct patch ( needed only to use RDMA with the Infiniband driver on Linux).

With 4.0 final, the GPU direct patch will be obsolete, just wait a little bit.

4

you always spoil my surprises :(

5

Yeah that is twice he has done that :)

6

Yeah, Thanks guys! I will wait for the final release.

Hopefully It also supports “enhanced mode” of MPI i.e MPI_Send and MPI_Recv directly on cuda devices

Cheers !

7

CUDA 4.0 Final release is out however I couldn’t find anywhere mentioned GPU-Direct 1.0. Is the claim previously still correct ?

8

With the 4.0 driver, just set this variable:

export CUDA_NIC_INTEROP=1

This will enable an alternate path for cudaMallocHost that is compatible with the IB drivers.
No need to change your code to use cudaHostRegister or to install the GPU direct v1 patch.

9

Thank mfatica, It helps a lot.
Performance over MPI improves about 30%

10

Thank mfatica, It helps a lot.
Performance over MPI improves about 30%