MARVIN Cluster
This is a private cluster.
Hardware
- Head node: marvin.cac.cornell.edu.
- access modes: ssh
- Rocks 5.4.3 with CentOS 5.6
- 92 compute nodes with Dual 6-core X5670 CPUs @ 3 GHz, Hyperthreaded, 48 GB of RAM; 4 high memory nodes with 96 GB of RAM
- Cluster Status: Ganglia.
- "Why use a temporary directory?"
- Submit HELP requests: Help OR by sending email to: help@cac.cornell.edu
File Systems
Home Directories
- Path: ~
User home directories is located on a NFS export from the head node. Use your home directory (~) for archiving the data you wish to keep. Do NOT use this file system for computation as bandwidth to the compute nodes is very limited and will quickly be overwhelmed by file I/Os from large jobs.
Unless special arrangements are made, data in user home directories are NOT backed up.
Scratch File System
LUSTRE file system provided by Terascala and Dell Path: /scratch/<user name>
The scratch file system is a fast parallel file system. Use this file system for scratch space for your jobs. Copy the results you want to keep back to your home directory for safe keeping.
Scheduler/Queues
- Maui/Torque scheduler;
- Queues:
Name Description Time Limit viz 4 visualization Ensight Servers, each has 96GB RAM 24 hours default all nodes except for those in viz queue 24 hours long all nodes except for those in viz queue 72 hours all all nodes none
Software
Software Path Notes Intel Cluster Studio /opt/intel - 30-day trial license.
- Intel compilers in user default path.
- Use mpi-selector to select Intel MPI
gcc 4.6.2 /opt/gcc/4.6.2 Prepend /opt/gcc/4.6.2/bin to $PATH to use this gcc version openmpi 1.6.3 (gnu) /opt/openmpi/gnu/1.6.3 - Compiled by gcc 4.6.2
- To select this MPI implementation, use "mpi-selector --set openmpi-1.6.3-gcc-4.6.2" command. Log out and log back in. This will set gcc 4.6.2 as the default compiler as well.
openmpi 1.6.3 (Intel) /opt/openmpi/intel/1.6.3 - Compiled by Intel 12.1
- To select this MPI implementation, use "mpi-selector --set openmpi-1.6.3-intel" command. Log out and log back in.
openmpi 1.4.4 (gnu) /opt/openmpi/gnu/1.4.4 - Compiled by gcc 4.6.2
- To select this MPI implementation, use "mpi-selector --set openmpi-1.4.4-gcc-4.6.2" command. Log out and log back in. This will set gcc 4.6.2 as the default compiler as well.
openmpi 1.4.4 (Intel) /opt/openmpi/intel/1.4.4 - Compiled by Intel 12.1
- To select this MPI implementation, use "mpi-selector --set openmpi-1.4.4-intel" command. Log out and log back in.
mvapich 1.2 (gnu) /opt/mvapich/gnu/1.2 - Compiled by gcc 4.6.2
- To select this MPI implementation, use "mpi-selector --set mvapich-1.2-gcc-4.6.2" command. Log out and log back in. This will set gcc 4.6.2 as the default compiler as well.
mvapich 1.2 (Intel) /opt/mvapich/intel/1.2 - DO NOT USE -- So far, the Intel 12.1 compiler has failed to produce a working build of mvapich.
- (in the future) To select this MPI implementation, use "mpi-selector --set mvapich-1.2-intel" command. Log out and log back in.
Intel MPI /opt/intel/impi/3.1 - To select this MPI implementation, use "mpi-selector --set intel-4.0.3" command. Log out and log back in.
fftw 3.3 (gnu) /opt/fftw/gnu/3.3 - Compiled by gcc 4.6.2
- With Intel compilers, Use MKL (Intel Math Kernel Library) in /opt/intel/mkl.
lapack 3.4.0 (gnu) /opt/lapack/gnu/3.4.0 - Compiled by gcc 4.6.2
- With Intel compilers, Use MKL (Intel Math Kernel Library) in /opt/intel/mkl.
hypre 2.0.0 (gnu) /opt/hypre/gnu/2.0.0 Compiled by gcc 4.1.2 hypre 2.6.0b (gnu) /opt/hypre/gnu/2.6.0b Compiled by gcc 4.6.2 hypre 2.6.0b (Intel) /opt/hypre/intel/2.6.0b Compiled by Intel Compilers 12.1 ensight 9.2 /usr/local/CEI - Installed only on the head node and "viz" queue (compute-3-13 to compute-3-16)
ensight 10.0 /usr/local/CEI - Installed only on the head node and "viz" queue (compute-3-13 to compute-3-16)
VisIt 2.9.2 /opt/visit - Must default to use the following OpenMPI version for parallel visualization like this:
-bash-3.2$ mpi-selector --set openmpi-1.6.3-gcc-4.6.2 Defaults already exist; overwrite them? (y/N) y -bash-3.2$ mpi-selector --query default:openmpi-1.6.3-gcc-4.6.2 level:user
- Log out and log back in for the change to take effect.
Anaconda Python /opt/anaconda-python - Add the following line to your ~/.bashrc to use anaconda python:
export PATH="/opt/anaconda-python/bin:$PATH"
Quick Tutorial
The batch system treats each core of a node as a "virtual processor." That means the nodes keyword in batch scripts refers to the number of cores that are scheduled.
Select your default MPI
There are several versions of MPI on the Marvin cluster. Use the following commands to modify your default mpi.
- mpi-selector --query -> shows your default mpi
- mpi-selector --list -> shows all available mpi installations
- mpi-selector --set <mpi installation> -> sets your default mpi, note, you will have to exit and log back in for this to take effect.
Running an MPI Job on the Whole Cluster
- assuming /opt/openmpi/ is the default, the mpiexec options may change depending on your selected MPI.
- First use showq to see how many cores are available. It may be less than 1152 if a node is down.
#!/bin/sh
#PBS -l nodes=96:ppn=12 (note, this is PBS -l (small case L))
#PBS -N test
#PBS -j oe
#PBS -S /bin/bash
set -x
cd "$PBS_O_WORKDIR"
mpiexec --hostfile $PBS_NODEFILE <executable> (where executable is the program you wish to run)
Running an MPI Job using 12 Tasks Per Node
Because the nodes have 12 physical cores, you may want to limit jobs to 12 tasks per node. The node file lists each node 1 time, so make a copy with each node listed 12 times, and hand that version to MPI.
#!/bin/sh
#PBS -l nodes=48
#PBS -N test
#PBS -j oe
#PBS -S /bin/bash
set -x
cd "$PBS_O_WORKDIR"
# Construct a copy of the hostfile with only 12 entries per node.
# MPI can use this to run 12 tasks on each node.
uniq "$PBS_NODEFILE"|awk '{for(i=0;i<12;i+=1) print}'>nodefile.12way
# to Run 12-way on 4 nodes, we request 48 core to obtain 4 nodes
mpiexec --hostfile nodefile.12way ring -v
Running Many Copies of a Serial Job
In order to run 30 separate instances of the same program, use the scheduler's task array feature, through the "-t" option. The "nodes" parameter here refers to a core.
#!/bin/sh
#PBS -l nodes=1 (note, this is PBS -l (small case L))
#PBS -t 30
#PBS -N test
#PBS -j oe
#PBS -S /bin/bash
set -x
cd "$PBS_O_WORKDIR"
echo Run my job.
When you start jobs this way, separate jobs will pile one-per-core onto nodes like a box of hamsters.
Running on a specific node
To run on a specific node use the host= option
#!/bin/sh
#PBS -l host=compute-3-16 (note, this is PBS -l (small case L))
#PBS -N test
#PBS -j oe
#PBS -S /bin/bash
set -x
cd "$PBS_O_WORKDIR"
echo Run my job.
Running in the viz queue
To run in the viz queue use the -q option
#!/bin/sh
#PBS -l nodes=1 (note, this is PBS -l (small case L))
#PBS -N test
#PBS -j oe
#PBS -S /bin/bash
#PBS -q viz
set -x
cd "$PBS_O_WORKDIR"
echo Run my job.
Running an interactive job
from the command line: qsub -l nodes=1 -I