Difference between revisions of "Leonhard beta testing"

Revision as of 14:56, 24 January 2018

The Leonhard cluster is available for early-access beta testing.

Please read through the following to get started.

Current status

The Leonhard Open cluster is still in the beta testing phase.

The lo-a2-* and lo-gtx-* compute nodes are already connected to the Infiniband high-performance interconnect and can access the GPFS storage system (/cluster/scratch,/cluster/work). The new lo-s4-* nodes are set up, but they are not yet connected to the Infiniband network and can therefore only use the Ethernet network. These nodes can not access the GPFS storage system but they have access to the NetApp storage system (/cluster/project,/cluster/home).

The HPC team is working on connecting the lo-s4-* nodes to the Infiniband network and on setting them up such that they can access the GPFS file system.

Programs that are linked against the centrally installed OpenMPI 2.1.0 require to have access to certain Infiniband and GPFS libraries, i.e., OpenMPI 2.1.0 can currently not be used on the lo-s4-* nodes. We are currently testing a workaround that would allow users to run codes linked against OpenMPI 2.1.0 with the Ethernet interconnect instead of Infiniband.

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.

SSH

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:

ssh username@login.leonhard.ethz.ch

where username corresponds to your NETHZ username.

Storage

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:

/cluster/home/username
/cluster/scratch/username

Modules for preparing your environment

LMOD

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.

Hierarchical modules

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.

Batch system

Leonhard uses the same LSF batch system as the Euler cluster. You can find some general information about the batch system in the Using the batch system tutorial.

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

Until further notice most GPU nodes can not access the GPFS filesystem. This includes /cluster/scratch and /cluster/work. You are advised to stage files into your $HOME directory.

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.

TensorFlow example

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.

Third-party applications

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.

TensorFlow

On Leonhard, we provide several versions of TensorFlow. The following combinations are available:

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.