First of all, sorry if I make spelling mistakes, I'm not english.
I'm trying to use IBM Platform MPI v9.1.2 with CUDA 5.5 in Windows 7 to pass messages through GPUs, using CUDA-Aware MPI like this post says: http://devblogs.nvidia.com/parallelforall/introduction-cuda-aware-mpi/
My GPUs (Nvidia Tesla C2075) are compatible with UVA (Unified Virtual Addressing) technology, so it should work properly when I use MPI_Ssend and MPI_Recv passing the device pointers, but it doesn't and the program crashes.
I've installed only the IBM package but I couldn't find anything about any configuration needed.
Anyone know something about this and could help me?
Thank you so much.
You'll need to switch to linux.
IBM indicates PMPI GPU-Direct support is on linux only here and here
Hardware requirements
...
In addition, Platform MPI supports GPU-Direct 2.0 on Linux .
CUDA-Aware MPI depends on GPUDirect, as indicated in the blog you linked.
Related
I read on the CUDA toolkit documentation (11.3.0) that "Deployment and execution of CUDA applications on x86_32 is still supported, but is limited to use with GeForce GPUs."
This looks in conflict with the fact that I was able to run a 32-bit app on my Tesla T4. (I verified that the code was actually running on the GPU and the app was 32-bit).
Have I misinterpreted the documentation? Why am I able to run 32-bit apps on a Tesla GPU?
(I'm running Visual studio 2017 on Windows 10)
It's a question of what is supported.
Other things may work, or they may not.
I have a Cuda application that was built with Cuda Toolkit 9.0 and running fine on Jetson TX2 board.
I now have a Jetson Xavier board, flashed with Jetpack 4 that installs Cuda Toolkit 10.0 (only 10.0 is available).
What do I need to do if I want to run the same application on Xavier?
Nvidia documentation suggests that as long as I specify the correct target hardware when running nvcc, I should be able to run on future hardwares thanks to JIT compilation. But does this hold for different versions of Cuda toolkit (9 vs 10)?
In theory (and note I don't have access to a Xavier board to test anything), you should be able to run a cross compiled CUDA 9 application (and that might mean both ARM and GPU architecture settings) on a CUDA 10 host.
What you will need to make sure is that you either statically link or copy all the CUDA runtime API library components you require with your application on the Xavier board. Note that there is still an outside chance that those libraries might lack the necessary GPU and ARM features to run correctly on a Xavier system, or more subtle issues like libC incompatibility. That you will have to test for yourself.
I would like to know how the allocation of a memory space in CUDA is implemented under Ubuntu Linux. In other words, how cudaMalloc() works internally under Ubuntu Linux? What are the system calls used for this function?
CUDA is proprietary. It's likely that CUDA driver implementation is the same or similar to OpenCL.
But while OpenCL specification is open the implementation is not necessary and NVIDIA OpenCL driver isn’t open .
It's possible that the implementation is as simple as the driver submitting a malloc command completely handled on the hardware side with the kernel driver communicating with the system to achieve unified virtual addressing and to determine what memory resides in VRAM. Probably the technical part at the software side is to avoid the allocation or defer it.
Looking into pocl can give you some idea how things can look like. NVIDIA implementation can be very different though.
I have a Dell XPS L502 with the Nvidia 525M graphics card. I am only interested in using the gpgpu capabilities of the card for now.
I installed Ubuntu 12.04 as a dual boot with the Windows 7 that came with the machine and followed several installation procedures for installing the CUDA driver and developer kit from Nvidia ( many re-installs of Ubuntu ). In all cases the display drops to 640x480 resolution. Best I can determine this has something to do with Optimus technology and Linux. I tried Bumblebee to no avail.
I really don't care about using the NVidia card to drive the display. Is there any way that I can just install the NVidia drivers so that a program can use the CUDA capabilities of the graphics card and I still get the full resolution on the display?
I had a similar issue with my Alienware M11xR2, and posted the solution on the NVIDIA Forums. Unfortunately the forums are down at the moment but essentially the process is as follows:
Install the Nvidia Drivers, but when prompted to modify your X11 Config, select 'No'. This is because the Nvidia card cannot be used as a display device.
Install the CUDA SDK and run one of the samples as root. I found this to be a necessary step. After this you should be able to execute further CUDA programs as a normal user.
Hope that helps.
With the new release of CUDA 5 the, comes the installation guide, there you have just one file that installs drivers, toolkit and sdk (even nvidia nsight). And one thing that got my attention is that you also have optimus options in the installation process.
I also have and Alienware M14x, and i understand your problem, but i also wanted the drivers to work for me, so i didn't try too hard on that.
Maybe you could give that a try and comment with the rest of us.
Here you can look for the CUDA 5 release candidate: CUDA 5
and here is the installation guide (maybe give this a read first): CUDA 5 Starting Guide for Linux.
is there any way I can test the CUDA samples and codes from a computer with no NVIDIA graphic card?
I am using Windows and the latest version of CUDA.
There are several possibilities:
Use older version of CUDA, which has built-in emulator (2.3 has it for sure). Emulator is far from good, and you won't have features from latest CUDA releases.
Use OpenCL, it can run on CPUs (though not with nVidia SDK, you will have to install either AMD or Intel OpenCL implementation (AMD works fine on Intel CPUs, btw)). In my experience, OpenCL is usually slightly slower than CUDA.
There is windows branch of Ocelot emulator: http://code.google.com/p/gpuocelot/. I haven't tried it, though.
However, I would recommend buying some CUDA-capable card. 8xxx or 9xxx series is ok and really cheap. Emulation would allow you to get some basic skills of GPGPU programming, but is useless when you write some real-world application since it doesn't allow you to debug and tune performance.