How to choose the number of cores
Warning
Asking for too few can lead to underused nodes or longer run time
Asking for too too many can mean wasted CPU resources
Asking for too much can mean a lot longer queuing
Table of Contents
Why it matters
We request resources from the scheduler (queuing system). But the scheduler cannot tell how long the job will run and what resources it will really consume. Just the fact that I am asking the scheduler for 40 cores does not mean that the code will actually run in parallel and use all of them.
Just because a website says that code X can run in parallel or “scales well” does not mean that it will scale well for the particular feature/system/input at hand.
Therefore it is important to verify and calibrate this setting for your use case before computing very many similar jobs. Below we will show few strategies.
Note that you don’t have to go through this for every single run. This is just to calibrate a job type. If many of your jobs have similar resource demands, then the calibration will probably be meaningful for all of them.
Using Arm Performance Reports
(This currently does not work because of network routing issues. We are working on restoring this.)
Using top
While the job is running, find out on which node(s) it runs using squeue -u $USER,
then ssh into one of the listed compute nodes and run top -u $USER.
Some clusters also have htop available which produces similar output as top
but with colors and possibly clearer overview.
Timing a series of runs
Here is an example C code (example.c) which we can compile and test a bit:
#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
#include <time.h>
#include <math.h>
double compute_something(int n, int m)
{
double s = 0.0;
for(int i = 0; i < n; i++)
{
double f = rand();
for(int j = 0; j < m; j++)
{
f = sqrt(f);
}
s += f;
}
return s;
}
int main(int argc, char* argv[])
{
MPI_Init(&argc, &argv);
int size;
MPI_Comm_size(MPI_COMM_WORLD, &size);
int my_rank;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
const int k = 10000;
double my_values[k];
double buffer_recv[k];
for(int j = 0; j < 750; j++)
{
for(int i = 0; i < k; i++)
{
my_values[i] = compute_something(1000/size, 10);
}
for(int l = 0; l < 1000; l++)
{
MPI_Alltoall(&my_values, 10, MPI_DOUBLE, buffer_recv, 10, MPI_DOUBLE, MPI_COMM_WORLD);
}
}
MPI_Finalize();
return EXIT_SUCCESS;
}
It does not matter so much what the code does. Here we wish to find out how this code scales and what an optimum number of cores for it might be on our system.
We can build our example binary with this script (compile.sh):
#!/bin/bash
module purge
module load foss/2020b
mpicc example.c -O3 -o mybinary -lm
Now take the following example script
(adapt --account=nn____k; this is tested on Saga):
#!/bin/bash
#SBATCH --account=nn____k
#SBATCH --qos=devel
#SBATCH --job-name='scaling'
#SBATCH --time=0-00:07:00
#SBATCH --mem-per-cpu=1GB
#SBATCH --ntasks=8
#SBATCH -o 008.out
module purge
module load foss/2020b
srun ./mybinary
Run a series of calculations on 1, 2, 4, 8, 16, 32, 64, and 128 cores.
You might get the following timings:
Number of cores |
Time spent in mybinary |
|---|---|
1 |
00:05:38 |
2 |
00:02:51 |
4 |
00:01:28 |
8 |
00:00:47 |
16 |
00:00:29 |
32 |
00:00:22 |
64 |
00:00:28 |
128 |
00:01:22 |
Please try this. What can we conclude? And how can we explain it?
Using seff
seff JOBID is a nice tool which we can use on completed jobs.
Here we can compare the output from seff when using 4 cores and when using
8 cores.
Run with 4 cores:
Job ID: 5690604
Cluster: saga
User/Group: user/user
State: COMPLETED (exit code 0)
Nodes: 2
Cores per node: 2
CPU Utilized: 00:00:39
CPU Efficiency: 75.00% of 00:00:52 core-walltime
Job Wall-clock time: 00:00:13
Memory Utilized: 42.23 MB (estimated maximum)
Memory Efficiency: 1.03% of 4.00 GB (1.00 GB/core)
Run with 8 cores:
Job ID: 5690605
Cluster: saga
User/Group: user/user
State: COMPLETED (exit code 0)
Nodes: 4
Cores per node: 2
CPU Utilized: 00:00:45
CPU Efficiency: 56.25% of 00:01:20 core-walltime
Job Wall-clock time: 00:00:10
Memory Utilized: 250.72 MB (estimated maximum)
Memory Efficiency: 3.06% of 8.00 GB (1.00 GB/core)
If it does not scale, what can be possible reasons?
Here are typical problems:
At some point more time is spent communicating than computing
Memory-bound jobs saturate the memory bandwidth
At some point the non-parallelized code section dominates the compute time (Amdahl’s law)
Example for a memory-bound job
Here is an example Fortran code (example.f90) which we can compile and test a bit:
integer function random_integer(lower_bound, upper_bound)
implicit none
integer, intent(in) :: lower_bound
integer, intent(in) :: upper_bound
real(8) :: f
call random_number(f)
random_integer = lower_bound + floor((upper_bound + 1 - lower_bound)*f)
end function
program example
implicit none
! adjust these
integer, parameter :: size_mb = 1900
integer(8), parameter :: num_rounds = 500000000
integer :: size_mw
integer, external :: random_integer
real(8), allocatable :: a(:)
integer :: iround, iposition
size_mw = int(1000000*size_mb/7.8125d0)
allocate(a(size_mw))
a = 0.0d0
do iround = 1, num_rounds
iposition = random_integer(1, size_mw)
a(iposition) = a(iposition) + 1.0d0
end do
deallocate(a)
print *, 'ok all went fine'
end program
We can build our example binary with this script (compile.sh):
#!/bin/bash
module load foss/2020b
gfortran example.f90 -o mybinary
Now take the following example script
(adapt --account=nn____k; this is tested on Saga):
#!/bin/bash
#SBATCH --account=nn____k
#SBATCH --qos=devel
#SBATCH --job-name='mem-bandwidth'
#SBATCH --time=0-00:02:00
#SBATCH --mem-per-cpu=2000M
#SBATCH --ntasks=1
module purge
module load foss/2020b
module load Arm-Forge/21.1
perf-report ./mybinary
This will generate the following performance report:
...
Summary: mybinary is Compute-bound in this configuration
Compute: 100.0% |=========|
MPI: 0.0% |
I/O: 0.0% |
This application run was Compute-bound. A breakdown of this time and advice for
investigating further is in the CPU section below. As very little time is
spent in MPI calls, this code may also benefit from running at larger scales.
CPU:
A breakdown of the 100.0% CPU time:
Scalar numeric ops: 4.4% ||
Vector numeric ops: 0.0% |
Memory accesses: 91.6% |========|
The per-core performance is memory-bound. Use a profiler to identify
time-consuming loops and check their cache performance. No time is spent in
vectorized instructions. Check the compiler's vectorization advice to see why
key loops could not be vectorized.
...
How do you interpret this result? Will it help to buy faster processors? What limitations can you imagine when running this on many cores on the same node?
What is MPI and OpenMP and how can I tell?
These two parallelization schemes are very common (but there are other schemes):
Message passing interface: typically each task allocates its own memory, tasks communicate via messages, and it is no problem to go beyond one node.
OpenMP: threads share memory and communicate through memory, and we cannot go beyond one node.
How to tell if the code is using one of the two?
If you wrote the software: then you probably know
If it is written by somebody else:
It can be difficult to tell
Consult manual for the software or contact support (theirs or ours)
grep -i mpiandgrep -i "omp "the source codeExample: https://github.com/MRChemSoft/mrchem (uses both MPI and OpenMP)
Python/R/Matlab
Often not parallelized
But can use parallelization (e.g.
mpi4pyormultiprocessing)