Running in parallel

• First at all, to know how many processors/cores you have available in your computer,
type in the terminal:
• $> lscpu
• The output for this particular workstation is the following:

Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Total number of cores available after
CPU(s): 24
hyper threading (virtual cores)
On-line CPU(s) list: 0-23
Thread(s) per core: 2
Core(s) per socket: 6 Number of threads per core (hyper threading)
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel Number of cores per socket or physical
CPU family: 6 processor
Model: 44
Model name: Intel(R) Xeon(R) CPU X5670 @ 2.93GHz
Stepping: 2 Number of sockets (physical processors)
CPU MHz: 1600.000
CPU max MHz: 2934.0000
CPU min MHz: 1600.0000
BogoMIPS: 5851.91
Total number of physical cores
Virtualization: VT-x
=
L1d cache: 32K
Number of cores per socket X Number of sockets
L1i cache: 32K
L2 cache: 256K
L3 cache: 12288K Total number of physical cores = 6 X 2 = 12 cores
NUMA node0 CPU(s): 0-5,12-17
NUMA node1 CPU(s): 6-11,18-23

This is what makes a processor expensive

 Running in parallel
• OpenFOAM® does not take advantage of hyper threading technology (HT).
• HT is basically used by the OS to improve multitasking performance.
• This is what we have in the workstation of the previous example:
• 24 virtual cores (hyper threaded)
• 12 physical cores

• To take full advantage of the hardware, we use the maximum number of physical
cores (12 physical cores in this case) when running in parallel.
• If you use the maximum number of virtual cores, OpenFOAM® will run but it will be
slower in comparison to running with the maximum number of physical cores (or even
less cores).
• Same rule applies when running in clusters/super computers, so always read the
hardware specifications to know the limitations.
 Running in parallel
Why use parallel computing?
• Solve larger and more complex problems (scale-up):
Thanks to parallel computing we can solve bigger problems (scalability). A single computer has limited
physical memory, many computers interconnected have access to more memory (distributed memory).

• Provide concurrency (scale-out):

A single computer or processor can only do one thing at a time. Multiple processors or computing
resources can do many things simultaneously.

• Save time (speed-up):

Run faster (speed-up) and increase your productivity, with the potential of saving money in the design
process.

• Save money:
In theory, throwing more resources at a task will shorten its time to completion, with potential cost
savings. Parallel computers can be built from cheap, commodity components.

• Limits to serial computing:

Both physical and practical reasons pose significant constraints to simply building ever faster serial
computers (e.g, transmission speed, CPU clock rate, limits to miniaturization, hardware cooling).
 Running in parallel
Speed-up and scalability example

• In the context of high performance computing (HPC), there are two common metrics that measure the scalability
of the application:
• Strong scaling (Amdahl’s law): which is defined as how the solution time varies with the number of
processors for a fixed problem size (number of cells in CFD)
• Weak scaling (Gustafson’s law): which is defined as how the solution time varies with the number of
processors for a fixed problem size per processor (or increasing the problem size with a fix number of
processors).
• In this example, when we reach 12 cores inter-processor communication slow-downs the computation. But if we
increase the problem size for a fix number of processors, we will increase the speed-up.
• The parallel case with 1 processor runs slower than the serial case due to the extra overhead when calling the
MPI library.
 Running in parallel
• The method of parallel computing used by OpenFOAM® is known as domain
decomposition, in which the geometry and associated fields are broken into pieces
and distributed among different processors.

Shared memory architectures – Workstations and portable computers

Distributed memory architectures – Clusters and super computers

 Running in parallel
Some facts about running OpenFOAM® in parallel:
• Applications generally do not require parallel-specific coding. The
parallel programming implementation is hidden from the user.
• In order to run in parallel you will need an MPI library installation in your
system.
• Most of the applications and utilities run in parallel.
• If you write a new solver, it will be in parallel (most of the times).
• We have been able to run in parallel up to 14000 processors.
• We have been able to run OpenFOAM® using single GPU and multiple
GPUs.
• Do not ask about scalability, that is problem/hardware specific.
• If you want to learn more about MPI and GPU programming, do not look
in my direction.
• And of course, to run in parallel you need the hardware.
 Running in parallel
To run OpenFOAM® in parallel you will need to:

• Decompose the domain.

To do so we use the decomposePar utility. You also need the dictionary
decomposeParDict which is located in the system directory.

• Distribute the jobs among the processors or computing nodes.

To do so, OpenFOAM® uses the standard message passing interface (MPI).
By using MPI, each processor runs a copy of the solver on a separate part
of the decomposed domain.

• Additionally, you might want to reconstruct (put back together) the

decomposed domain.
This is done by using the reconstrucPar utility. You do not need a
dictionary to use this utility.
 Running in parallel
Domain Decomposition in OpenFOAM®
• The mesh and fields are decomposed using the decomposePar utility.
• They are broken up according to a set of parameters specified in a dictionary named
decomposeParDict that is located in the system directory of the case.
• In the decomposeParDict dictionary the user must set the number of domains in
which the case should be decomposed (using the keyword numberOfSubdomains).
The value used should correspond to the number of physical cores available.

numberOfSubdomains 128; Number of subdomains

method scotch; Decomposition method

• In this example, we are subdividing the domain in 128 subdomains, therefore we

should have 128 physical cores available.
• The main goal of domain decomposition is to minimize the inter-processors
communication and the processor workload.
 Running in parallel
Domain Decomposition Methods
• These are the decomposition methods available in OpenFOAM® 5.x
• hierarchical
• manual
• metis
• multiLevel
• none We highly recommend you to use this method.
The only input that requires from the user is
• scotch the number of subdomains/cores. This method
attempts to minimize the number of processor
• simple boundaries.

• structured

• If you want more information about each decomposition method, just read
the source code:
• $WM_PROJECT_DIR/src/parallel/decompose/
 Running in parallel
Running in parallel – Gathering all together

The information inside the

directories polyMesh/ and
0/ is decomposed using the
utility decomposePar

decomposePar

processor0 processor1 processor2 processor3

• Inside each processorN directory you will have the mesh information, boundary conditions,
initial conditions, and the solution for that processor.
 Running in parallel
Running in parallel – Gathering all together
• After decomposing the mesh, we can run in parallel using MPI.

$> mpirun –np <NPROCS> <application/utility> –parallel

• The number of processors to use or <NPROCS>, needs to be the same as the

number of partitions (numberOfSubdomains).
• Do not forget to use the flag –parallel.
 Running in parallel
Running in parallel – Gathering all together

• In the decomposed case, you will find the mesh

information, boundary conditions, initial
conditions, and the solution for every processor.
• The information is inside the directory
processorN (where N is the processor number).

reconstructPar

• When you reconstruct the case, you glue together

all the information contained in the decomposed
case.
• All the information (mesh, boundary conditions,
initial conditions, and the solution), is transfer to
the original case folder (polyMesh and time
solution directories).
 Running in parallel
Running in parallel – Gathering all together

• Summarizing, to run in parallel we proceed in the following way:

1. $> decomposePar
2. $> mpirun –np <NPROCS> <application/utility> –parallel
3. $> reconstructPar

• You can do the post-processing and visualization on the decomposed case

or reconstructed case. We are going to address this later on.

• If you are doing remeshing or using AMR you will need to use
reconstructParMesh before reconstrucPar.
 Running in parallel
Kelvin Helmholtz instability in a coarse mesh

Mesh size
Processors Clock time (seconds)
in x, y, and z directions

1 955 800 X 160 X 1

2 564 800 X 160 X 1

4 333 800 X 160 X 1

8 234 800 X 160 X 1

12 244 800 X 160 X 1

Volume fraction
www.wolfdynamics.com/wiki/kelvin_helmholtz/ani1.gif

You will find this case in the directory: $PTOFC/parallel/kelvin_helmholtz

 Running in parallel
Comparison of the Kelvin Helmholtz instability in a coarse and fine mesh

Cells 200 X 40 X 1 Cells 3200 X 640 X 1

 Running in parallel
Visualization of a parallel case

• The traditional way is to first reconstruct the case and then do the post-
processing and visualization on the reconstructed case.

• To do so, we type in the terminal:

1. $> reconstructPar
2. $> paraFoam

• Step 1 reconstruct the case. Remember, you can choose to reconstruct all
the time steps, the last time step or a range of time steps.

• In step 2, we use paraFoam to visualize the reconstructed case.

 Running in parallel
Visualization of a parallel case

• An alternative way to visualize the solution, is by proceeding in the following

way
• $> paraFoam –builtin

• The option –builtin let us post-process the decomposed case directly.

• Remember, you will need to select on the object inspector the Decomposed
Case option.
 Running in parallel
Visualization of a parallel case

• Both of the previous methods are valid.

• When we use the option –builtin with paraFoam, we have the option to
work on the decomposed case directly.
• That is to say, we do not need to reconstruct the case.
• But wait, there is a third option.
• The third option consist in post-processing each decomposed domain
individually.
• To load all processor directories, you will need to manually create the file
processorN.OpenFOAM (where N is the processor number) in each
processor folder.
• After creating all processorN.OpenFOAM files, you can launch paraFoam
and load each file (the processorN.OpenFOAM files).
• As you can see, this option requires more input from the user.
 Running in parallel
Decomposing big meshes
• One final word, the utility decomposePar does not run in parallel. So, it is
not possible to distribute the mesh among different computing nodes to do
the partitioning in parallel.

• If you need to partition big meshes, you will need a computing node with
enough memory to handle the mesh. We have been able to decompose
meshes with up to 500.000.000 elements, but we used a computing node
with 512 gigs of memory.

• For example, in a computing node with 16 gigs of memory, it is not possible

to decompose a mesh with 30.000.000. You will need to use a computing
node with at least 32 gigs of memory.

• Same applies for the utility reconstructPar.

 Running in parallel
Do all utilities run in parallel?
• At this point, you might be wondering if all solvers/utilities run in parallel.
• To know what solvers/utilities do not run in parallel, in the terminal type:
• $> find $WM_PROJECT_DIR -type f | xargs grep –sl ‘noParallel’
• Paradoxically, the utilities used to decompose the domain and reconstruct the
domain do not run in parallel.
• Another important utility that does not run in parallel is blockMesh. So to generate
big meshes with blockMesh you need to use a big fat computing node.
• Another important utility that does not run in parallel by default is paraFoam.
• To compile paraFoam with MPI support, in the file makeParaView4 (located in the
directory $WM_THIRD_PARTY_DIR), set the option withMPI to true,
• withMPI = true
• While you are working with the file makeParaView4, you might consider enabling
Python support,
• withPYTHON = true
 Running in parallel
Running in a cluster using a job scheduler
• Running OpenFOAM® in a cluster is similar to running in a normal
workstation with shared memory.
• The only difference is that you will need to launch your job using a job
scheduler.
• Common job schedulers are:
• Terascale Open-Source Resource and Queue Manager (TORQUE).
• Simple Linux Utility for Resource Management (SLURM).
• Portable Batch System (PBS).
• Sun Grid Engine (SGE).
• Maui Cluster Scheduler.
• BlueGene LoadLeveler (LL).
• Ask your system administrator the job scheduler installed in your system.
Hereafter we will assume that you are using PBS.
 Running in parallel
• To launch a job in a cluster with PBS, you will need to write a small shell script where
you tell to the job scheduler the resources you want to use and what you want to do.
• Ask the system administrator about how to write these scripts.

#!/bin/bash
#
# Simple PBS batch script that reserves 16 nodes and runs a
# MPI program on 128 processors (8 processor on each node)
# The walltime is 24 hours !
#
#PBS -N openfoam_simulation //name of the job
#PBS -l nodes=16,walltime=24:00:00 //max resources and execution time
#PBS -m abe -M joel.guerrero@unige.it //send an email as soon as the job
//is launch or terminated

cd PATH_TO_DIRECTORY //go to this directory

decomposePar //decompose the case

mpirun –np 128 pimpleFoam -parallel > log //run parallel solver

The green lines are not PBS comments, they are comments inserted in this slide. PBS comments use the number
sign (#).
 Running in parallel
• To launch your job you need to use the qsub command (part of the PBS job
scheduler). The command qsub will send your job to queue.

• $> qsub script_name

• Remember, running in a cluster is no different from running in your

workstation or portable computer. The only difference is that you need to
schedule your jobs.

• Depending on the system current demand of resources, the resources you

request and your job priority, sometimes you can be in queue for hours, even
days, so be patient and wait for your turn.

• Remember to always double check your scripts.

 Running in parallel
• Finally, remember to always plan how you will use the resources available.
• For example, if each computing node has 8 gigs of memory available and 8 cores.
You will need to distribute the work load in order not to exceed the maximum
resources available per computing node.
• So if you are running a simulation that requires 32 gigs of memory, the following
options are valid:
• Use 4 computing nodes and ask for 32 cores. Each node will use 8 gigs of
memory and 8 cores.
• Use 8 computing nodes and ask for 32 cores. Each node will use 4 gigs of
memory and 4 cores.
• Use 8 computing nodes and ask for 64 cores. Each node will use 4 gigs of
memory and 8 cores.
• But the following options are not valid:
• Use 2 computing nodes. Each node will need 16 gigs of memory.
• Use 16 computing nodes and ask for 256 cores. The maximum number of cores
for this job is 128.