\\ **[[cluster:0|Back]]** ===== Cuda ===== Upgrading Cuda to latest drivers and tooltkit that supports our GeForce RTX 2080 SUPER (and Ti) gpu models (queues exx96 and amber128). Before we embark doing all nodes, we need to test backward compatibility and assess how troublesome the upgrade might be. First look at what is end-of-life these days ... middle of page go to table table **GPUs supported**, scroll down keeping an eye on **Turing** and **Lovelace** columns. Cuda version 10.2 (our RTX2080S) is still supported in version 12.4 so we can install latest toolkit and drivers. * https://en.wikipedia.org/wiki/CUDA And it looks like Tesla P4, P100 and V100 are next on the end of extended life list, consult table //"A. End-of-Software-Support Dates for GPUs Supported by NVIDIA vGPU Software"// * https://docs.nvidia.com/grid/news/vgpu-software-lifecycle-on-supported-gpus/index.html Find the drivers at this page (non-legacy gpu models as opposed to legacy models [[cluster:223|cuda toolkit]]), look for link **Linux x86_64/AMD64/EM64T** //Latest Production Branch Version:// __550.67__ * https://www.nvidia.com/en-us/drivers/unix/ Linux x86_64/AMD64/EM64T\\ Latest Production Branch Version: 550.67\\ Supported Products Tab \\ (make sure the models are listed -- they are) * https://www.nvidia.com/Download/driverResults.aspx/223426/en-us/ Agree&Download the NVIDIA-Linux-x86_64-550.67.run driver file. Now obtain the latest toolkit as our gpu models are supported as verified above * https://developer.nvidia.com/cuda-downloads * Select: Linux > x86_64 > CentOS > 7 > Runfile (local) * I see they have Rocky support too... and you will get instructions # centos 7 wget https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run sudo sh cuda_12.4.0_550.54.14_linux.run # rocky 8 (same file) wget https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run sudo sh cuda_12.4.0_550.54.14_linux.run The complete Linux Installer Guide is at * https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html ===== Installation ===== * On CentOS 7 hop on vlan52, **yum update** followed by reboot (EOL 06/30/2024) * skip; libnvidia-container repo has problems and may be crucial * On Rocky 8 * **yum update**; skip this step for now ... * Make both files executable (same files for both) * Install driver first (Nvidia-Linux-*.run, reboot) * Then install toolkit (cuda-12.4-*.run, reboot) # n78 first ... (no problem, tests success) # make sure /usr/src/kernels/$(uname -r) exists else # scp into place from n100 (centos8, possibly caused by warewulf...) # n79 next (no problem) # but during testing found this error which I had seen BEFORE cuda install reason: BUG: unable to handle kernel NULL pointer dereference at 0000000000000108 # follow up, nvidia driver "taints" the kernel by loading proprietary drive # this is a warning mostly by may interfere as does docker likely https://unix.stackexchange.com/questions/118116/what-is-a-tainted-linux-kernel # disabled docker on n79 for now 04/15/2024 9:06AM # also rotated the memory dimms some time later, seems to have fixed issue # started docker back up on n79 05/06/2024 9:56AM (has been up 17 days by now) # n89 next (no problem) # but upon reboot I encountered that error for the FIRST time on this node # need to research it is somewhat related to cuda install # n80 (same error upon reboot after driver install) # n81 (same error upon reboot after driver install) # n90 (same error upon reboot after toolkit install, not driver. weird) # n88 (failed toolkit install, ran /usr/bin/ndia-uninstall, reboot # re-installed driver, reboot, re-installed tookit, reboot, # no error occurs! ) sh ./NVIDIA-Linux-x86_64-550.67.run 32 bit compat? no dkms build? yes rebuild initramfs? yes xconfig? no error nvidia module can not be loaded reboot fixed that [root@n78 ~]# nvidia-smi Mon Apr 1 14:50:33 2024 +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 550.67 Driver Version: 550.67 CUDA Version: 12.4 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| | 0 NVIDIA GeForce GTX 1080 Ti Off | 00000000:02:00.0 Off | N/A | | 19% 38C P0 59W / 250W | 0MiB / 11264MiB | 0% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 1 NVIDIA GeForce GTX 1080 Ti Off | 00000000:03:00.0 Off | N/A | | 16% 25C P0 58W / 250W | 0MiB / 11264MiB | 0% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 2 NVIDIA GeForce GTX 1080 Ti Off | 00000000:81:00.0 Off | N/A | | 18% 28C P0 59W / 250W | 0MiB / 11264MiB | 0% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 3 NVIDIA GeForce GTX 1080 Ti Off | 00000000:82:00.0 Off | N/A | | 19% 29C P0 58W / 250W | 0MiB / 11264MiB | 0% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ +-----------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| | No running processes found | +-----------------------------------------------------------------------------------------+ rm -f /usr/local/cuda # soft link to current version /usr/bin/nvidia_uninstall # probably not needed, made no difference > Continue sh cuda_12.4.0_550.54.14_linux.run REBOOT and check date before launching slurm mv /var/spool/slurmd/cred_state /var/spool/slurmd/cred_state.bak =========== = Summary = =========== Driver: Installed Toolkit: Installed in /usr/local/cuda-12.4/ Please make sure that - PATH includes /usr/local/cuda-12.4/bin - LD_LIBRARY_PATH includes /usr/local/cuda-12.4/lib64, or, add /usr/local/cuda-12.4/lib64 to /etc/ld.so.conf and run ldconfig as root To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-12.4/bin To uninstall the NVIDIA Driver, run nvidia-uninstall Logfile is /var/log/cuda-installer.log # check /usr/local/cuda link, clean up others [root@n78 ~]# export PATH=/usr/local/cuda/bin:$PATH [root@n78 ~]# export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH [root@n78 ~]# which nvidia-smi /usr/bin/nvidia-smi [root@n78 ~]# which nvcc /usr/local/cuda/bin/nvcc [root@n78 ~]# nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2024 NVIDIA Corporation Built on Tue_Feb_27_16:19:38_PST_2024 Cuda compilation tools, release 12.4, V12.4.99 Build cuda_12.4.r12.4/compiler.33961263_0 # slurm finds gpus? yes... [2024-01-18T12:46:29.322] debug: gpu/generic: init: init: GPU Generic plugin loaded [2024-01-18T12:46:29.322] Gres Name=gpu Type=geforce_gtx_1080_ti Count=4 ===== Testing ===== ** Rocky 8 on n78 ** Previous software compiled using OpenHPC modules pulls in ''cuda/11.6'' module. So lets test if that works with new 550 driver and 11.2 toolkit. You can follow these steps to test any other software. # ssh to compute node and load module [hmeij@n78 ~]$ module load amber/22 # identify object of interest [hmeij@n78 ~]$ which pmemd.cuda /share/apps/CENTOS8/ohpc/software/amber/22/bin/pmemd.cuda # feed to ldd, look for missing libraries [hmeij@n78 ~]$ ldd `which pmemd.cuda` linux-vdso.so.1 (0x00007ffeb51ea000) libemil.so => /share/apps/CENTOS8/ohpc/software/amber/22//lib/libemil.so (0x00007f99ae5aa000) libnetcdff.so.6 => /share/apps/CENTOS8/ohpc/software/amber/22//lib/libnetcdff.so.6 (0x00007f99ae315000) libkmmd.so => /share/apps/CENTOS8/ohpc/software/amber/22//lib/libkmmd.so (0x00007f99ae10a000) libcufft.so.10 => /usr/local/cuda/lib64/libcufft.so.10 (0x00007f99a567d000) libdl.so.2 => /lib64/libdl.so.2 (0x00007f99a5479000) libcurand.so.10 => /usr/local/cuda/lib64/libcurand.so.10 (0x00007f999f033000) libcublas.so.11 => /usr/local/cuda/lib64/libcublas.so.11 (0x00007f99955bf000) libnetcdf.so.15 => /lib64/libnetcdf.so.15 (0x00007f9995273000) libarpack.so => /share/apps/CENTOS8/ohpc/software/amber/22//lib/libarpack.so (0x00007f9995056000) libopenblas.so.0 => /lib64/libopenblas.so.0 (0x00007f999328e000) libm.so.6 => /lib64/libm.so.6 (0x00007f9992f0c000) # does it work? [hmeij@n78 ~]$ pmemd.cuda --version pmemd.cuda: Version 22.0 So you'll notice libraries are pulled in from //.../amber/22/...// and //.../cuda/lib64/...//. The latter is technically wrong because that generic path points to 12.4 now not 11.6. The problem was the module **did not** specify the absolute path to cuda-11.6. Fixed that and provided soft links for individual libraries in cuda-12.4. Not the way to do it but it works. So toolkit 11.6 is compatible with 12.4 using the newly installed 550 driver. Amber 20/22 and Lammps 25Apr2023/7Feb2024 all work. See script ''~hmeij/slurm/run.rocky.2''. pmemd.cuda: error while loading shared libraries: libcufft.so.10 ...argh, Amber links in specific version of libraries # not the way to do things...amber 20&22 cd /usr/local/cuda-12.4/targets/x86_64-linux/lib/ ln -s /usr/local/cuda-11.6/lib64/libcufft.so.10 ln -s /usr/local/cuda-11.6/lib64/libcublas.so.11 ln -s /usr/local/cuda-11.6/lib64/libcublasLt.so.11 # not the way to do things...lammps 7Feb2024 and 25Apr2023 ln -s /usr/local/cuda-11.6/lib64/libcudart.so.11.0 ** CentOS 7 on n89 ** The steps above can also be done for the default cuda installation on exx96 where the soft link ''/usr/local/bin/cuda'' would have pointed to ''/usr/local/bin/cuda-10.2''. Do not follow the soft link and use the path with the toolkit version in it when setting your cuda environment. Next test is to see if older software runs compatible with newer drivers. We test that by running a gpu program against new 550 driver and cuda toolkit 9.2 and see if it works (~hmeij/slurm/run.centos7.2). # same steps as above using n89 # set up the environment the script uses [hmeij@n89 ~]$ which nvcc /usr/local/n37-cuda-9.2/bin/nvcc [hmeij@n89 ~]$ source /usr/local/amber20/amber.sh [hmeij@n89 ~]$ which pmemd.cuda /usr/local/amber20/bin/pmemd.cuda [hmeij@n89 ~]$ pmemd.cuda --version pmemd.cuda: Version 20.0 # going further back in time [hmeij@n89 ~]$ source /usr/local/amber16/amber.sh [hmeij@n89 ~]$ which pmemd.cuda /usr/local/amber16/bin/pmemd.cuda [hmeij@n89 ~]$ pmemd.cuda --version pmemd.cuda: Version 14.0 So it appears that all our cuda versions can use the new 550 driver that comes with cuda-12.4 toolkit. Two other cuda versions have not been tested but should function as well (cuda-11.2 in mwgpu and cuda-10.2 in exx96). But in these queues cuda-9.2 is present on local hard disk and software was compiled against that toolkit so which queue to use did not matter. (compilations against 10.2 did not run in 9.2, as expected). I was able to test and run lammps in 10.2 consult the file ''~hmeij/slurm/centos.2'' These compatibility results are way, way better than expected. Yea. Any new software compilations will use module ''cuda/12.4''. \\ **[[cluster:0|Back]]**