Leonhard beta testing
Please read through the following to get started on the Leonhard cluster.
- 1 Current status
- 2 Accessing the cluster
- 3 Storage
- 4 Modules for preparing your environment
- 5 Batch system
- 6 Third-party applications
- 7 Troubleshooting
The Leonhard Open cluster is still in the beta testing phase.
The lo-a2-*, the lo-gtx-* and the lo-s4-* compute nodes are connected to the InfiniBand high-performance interconnect and can access the GPFS storage system (/cluster/scratch,/cluster/work) and the NetApp storage system (/cluster/project,/cluster/home).
Accessing the cluster
Who can access the cluster
Access is restricted to Leonhard shareholders and groups that want to test it before investing. Guest users cannot access the Leonhard cluster.
You can find general information about how to access our clusters via SSH in the Accessing the clusters tutorial.
The command to access the Leonhard cluster via SSH is:
where username corresponds to your NETHZ username.
You can find general information about the storage systems in the Storage systems tutorial.
Like on the Euler cluster, every user also has a home directory and a personal scratch directory:
Modules for preparing your environment
For the Leonhard cluster, we decided to switch from the environment modules that are used on the Euler cluster to Lmod modules, which provide some additional features. You should barely notice the transition from environment modules to Lmod modules as the commands are mostly the same. Therefore please refer to the Setting up your environment tutorial for a general documentation about the module commands.
[leonhard@lo-login-02 ~]$ module avail boost ----------------------------------------- /cluster/apps/lmodules/Compiler/gcc/4.8.5 ------------------------------------------ boost/1.63.0 Use "module spider" to find all possible modules. Use "module keyword key1 key2 ..." to search for all possible modules matching any of the "keys". [leonhard@lo-login-02 ~]$ module load boost/1.63.0 [leonhard@lo-login-02 ~]$ module list Currently Loaded Modules: 1) gcc/4.8.5 2) StdEnv 3) boost/1.63.0 [leonhard@lo-login-02 ~]$
Please note that this is work in progress and the module names might change. Currently, the number of software packages provided on Leonhard is not comparable to the software we provide on the Euler cluster, but it will grow over time.
LMOD allows to define a hierarchy of modules containing 3 layers (Core, Compiler, MPI). The core layer contains all module files which are not depending on any compiler/MPI. The compiler layer contains all modules which are depending on a particular compilers, but not on any MPI library. The MPI layer contains modules that are depending on a particular compiler/MPI combination.
When you login to the Leonhard cluster, the standard compiler gcc/4.8.5 is automatically loaded. Running the module avail command displays all modules that are available for gcc/4.8.5. If you would like to see the modules available for a different compiler, for instance gcc/6.3.0, then you would need to load the compiler module and run module avail again. For checking out the available modules for gcc/4.8.5 openmpi/2.1.0, you would load the corresponding compiler and MPI module and run again module avail'.
As a consequence of the module hierachy, you can never have two different versions of the same module loaded at the same time. This helps to avoid problems arising due to misconfiguration of the environment.
Unless otherwise specified, jobs requesting more than 36 cores will run on a single node. Regular nodes have 36 cores and 128 or 512 GB of RAM (of which about 90 and 460 GB, respectively, are usable).
Unlike Euler, requested memory is strictly enforced as a memory limit.
Submitting GPU jobs
All GPUs in Leonhard are configured in Exclusive Process mode. The GPU nodes have 20 cores, 8 GPUs, and 256 GB of RAM (of which only about 210 GB is usable). To run multi-node job, you will need to request span[ptile=20].
The LSF batch system has partial integrated support for GPUs. To use the GPUs for a job node you need to request the ngpus_excl_p resource. It refers to the number of GPUs per node. This is unlike other resources, which are requested per core.
For example, to run a serial job with one GPU,
bsub -R "rusage[ngpus_excl_p=1]" ./my_cuda_program
or on a full node with all eight GPUs and up to 90 GB of RAM,
bsub -n 20 -R "rusage[mem=4500,ngpus_excl_p=8]" ./my_cuda_program
or on two full nodes:
bsub -n 40 -R "rusage[mem=4500,ngpus_excl_p=8] span[ptile=20]" ./my_cuda_program
While your jobs will see all GPUs, LSF will set the CUDA_VISIBLE_DEVICES environment variable, which is honored by CUDA programs.
As an example for running a TensorFlow job on a GPU node, we are printing out the TensorFlow version, the string Hello TensorFlow! and the result of a simple matrix multiplication:
[leonhard@lo-login-01 ~]$ cd testrun/python [leonhard@lo-login-01 python]$ module load python_gpu/2.7.13 [leonhard@lo-login-01 python]$ cat tftest1.py #/usr/bin/env python from __future__ import print_function import tensorflow as tf vers = tf.__version__ print(vers) hello = tf.constant('Hello, TensorFlow!') matrix1 = tf.constant([[3., 3.]]) matrix2 = tf.constant([[2.],[2.]]) product = tf.matmul(matrix1, matrix2) sess = tf.Session() print(sess.run(hello)) print(sess.run(product)) sess.close() [leonhard@lo-login-01 python]$ bsub -n 1 -W 4:00 -R "rusage[mem=2048, ngpus_excl_p=1]" python tftest1.py Generic job. Job <10620> is submitted to queue <gpu.4h>. [leonhard@lo-login-01 python]$ bjobs JOBID USER STAT QUEUE FROM_HOST EXEC_HOST JOB_NAME SUBMIT_TIME 10620 leonhard PEND gpu.4h lo-login-01 *tftest.py Sep 28 08:02 [leonhard@lo-login-01 python]$ bjobs JOBID USER STAT QUEUE FROM_HOST EXEC_HOST JOB_NAME SUBMIT_TIME 10620 leonhard RUN gpu.4h lo-login-01 lo-gtx-001 *ftest1.py Sep 28 08:03 [leonhard@lo-login-01 python]$ bjobs No unfinished job found [leonhard@lo-login-01 python]$ grep -A3 "Creating TensorFlow device" lsf.o10620 2017-09-28 08:08:43.235886: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1080, pci bus id: 0000:04:00.0) 1.3.0 Hello, TensorFlow! 12. [leonhard@lo-login-01 python]$
Please note, that your job will crash if you are running the GPU version of TensorFlow on a CPU node, because TensorFlow is checking on start up if the compute node has a GPU driver.
Python on Leonhard
Because certain Python packages need different installations for their CPU and GPU versions, we decided to have separate Python installations with regards to using CPUs and GPUs.
|CPU version||GPU version|
|module load python_cpu/3.6.1||module load python_gpu/3.6.1|
On Leonhard, we provide several versions of TensorFlow. The following combinations are available:
|Module command||TensorFlow version|
|module load python_cpu/2.7.12||Python 2.7.12, TensorFlow 1.2.1|
|module load python_cpu/2.7.13||Python 2.7.13, TensorFlow 1.3|
|module load python_cpu/3.6.0||Python 3.6.0, TensorFlow 1.2.1|
|module load python_cpu/3.6.1||Python 3.6.1, TensorFlow 1.3|
|Module command||TensorFlow version|
|module load python_gpu/2.7.12||Python 2.7.12, TensorFlow 1.2.1, CUDA 8.0.61, cuDNN 5.1|
|module load python_gpu/2.7.13||Python 2.7.13, TensorFlow 1.3, CUDA 8.0.61, cuDNN 6.0|
|module load python_gpu/3.6.0||Python 3.6.0, TensorFlow 1.2.1, CUDA 8.0.61, cuDNN 5.1|
|module load python_gpu/3.6.1||Python 3.6.1, TensorFlow 1.3, CUDA 8.0.61, cuDNN 6.0|
If you would like to run a TensorFlow job on a CPU node, then you would need to load a CPU version of TensorFlow, whereas you would need to load a GPU version of TensorFlow for running a TensorFlow job on a GPU node.
I can no longer access the Leonhard Open cluster
Some users only got temporary access to the Leonhard Open cluster for a course or a project, which is limited in time. The HPC group does not manage membership of the shareholder groups on Leonhard. The shareholder groups are defined through a custom NETHZ ggroup. If you could access Leonhard Open for some time and can no longer access it, then you were most likely removed from the custom NETHZ group, which was used to define the share. If this is the case, then please contact the local IT support group (ISG) of your department for further information, as they are managing the custom NETHZ groups.
I get an error message, when running my software that uses a GPU
If you are getting error messages, about not finding a cuda library, like for instance:
ImportError: libcuda.so.1: cannot open shared object file: No such file or directory
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory
Then you are most likely running the software on a login node or a compute node without GPU's.
If you would like to run a software that requires access to the GPU driver, then you need to submit it as a batch job and request a GPU from the batch system.
Cluster is missing the h5py python package
The h5py Python package is linked against the HDF5 library, therefore you need to also load the HDF5 module, such that h5py can located the HDF5 libraries.
[leonhard@lo-s4-019 ~]$ module load python_gpu/3.6.1 hdf5/1.10.1 [leonhard@lo-s4-019 ~]$ python Python 3.6.1 (default, Sep 27 2017, 13:27:13) [GCC 4.8.5 20150623 (Red Hat 4.8.5-11)] on linux Type "help", "copyright", "credits" or "license" for more information. >>> import h5py >>> h5py.__version__ '2.7.1' >>>