- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -

Difference between revisions of "Batch System PBSPro (Hawk)"

From HLRS Platforms
(Improved python scripts to work with both python2 and python3)
m (Shall I use all the available cores?)
Line 39: Line 39:
 
#########################################################################
 
#########################################################################
 
# Usage:  ./distribute_by_fraction.py <numerator> <denominator>        #
 
# Usage:  ./distribute_by_fraction.py <numerator> <denominator>        #
# Example: ./distribute_by_fraction.py 1 8                              #
+
# Example: ./distribute_by_fraction.py 32 128                          #
 
#                                                                      #
 
#                                                                      #
 
# The script will then generate a list of <numerator>/<denominator>*128 #
 
# The script will then generate a list of <numerator>/<denominator>*128 #

Revision as of 08:51, 4 November 2019

The batch system on Hawk TDS is PBSPro 19.2.1. For general usage see the PBS User Guide (19.2.3)

At the moment the setup is basic and it works for the TDS only. More features, testing and productive like setup will be done in July.


Node types

There are two types of nodes installed in the TDS:

  • 16 x AMD EPYC Naples (2 x 32 cores each): select with #PBS -l node_type=naples
  • 1 x AMD EPYC Rome (2 x 64 cores each): select with #PBS -l node_type=rome


Core order

On Rome-based nodes, the core id corresponds to hyperthreads and sockets as follows:

core 0 - core 63: hyperthread 0 @ socket 0
core 64 - core 127: hyperthread 0 @ socket 1
core 128 - core 191: hyperthread 1 @ socket 0
core 192 - core 256: hyperthread 1 @ socket 1

Hence, cores 128 to 256 are using the same physical resources as cores 0 to 127! Only use them if you understand the concept of hyperthreads and actually like to use them! If you do not like to use them, start a maximum of 128 threads per node.

Pinning

We recommend to always (in hybrid as well as pure MPI jobs) use omplace to pin processes and threads to CPU cores (cf. below) in order to prevent expensive migration.


Shall I use all the available cores?

Due to limited memory bandwidth, it might be beneficial to not use all the available cores in a node. Unfortunately, you have to figure out your sweet spot by means of trial & error. While doing this, please have in mind the internal structure of the processor (cf. Tutorial Slides from HLRS Results & Review Workshop 2019) and try to uniformly distribute processes over architectural building blocks (i.e. CCXs, CCDs, NUMA nodes and sockets). In order to make things more easy, please use the block and stride features of omplace (cf. manpage) or use the scripts provided below to generate lists of core IDs to be passed to omplace via the -c flag if your intended placement is not possible by means of blocks & strides.

#!/usr/bin/python

#########################################################################
# Usage:   ./distribute_by_fraction.py <numerator> <denominator>        #
# Example: ./distribute_by_fraction.py 32 128                           #
#                                                                       #
# The script will then generate a list of <numerator>/<denominator>*128 #
# cores to be used, equally distributed among the available 128 cores.  #
#########################################################################

import sys

numerator   = int(sys.argv[1])
denominator = int(sys.argv[2])

core_list = ""
for offset in range(0, 127, denominator):
    for j in range(numerator):
        index = int(j*round(float(denominator)/float(numerator)))
        core_list = core_list + str(offset + index) + ","

sys.stdout.write(core_list[:-1] + "\n")
#!/usr/bin/python

#########################################################################
# Example usage: ./distribute_by_pattern.py 1 0 0 0                     #
#                                                                       #
# This will generate a list with core 0 being used, cores 1-3 not being #
# used and so on (i.e. pattern will be repeated until status of all 128 #
# cores is defined).                                                    #
#########################################################################

import sys

core_list = ""
for i in range(128):
    if sys.argv[i%(len(sys.argv) - 1) + 1] == "1":
        core_list = core_list + str(i) + ","

sys.stdout.write(core_list[:-1] + "\n")


Examples

See

man pbs_resources

regarding available resources (e.g. ncpus, mpiprocs, etc.) and how to specify resources in the job script.


pure MPI job using HPE MPI

Here is a simple pbs job script:

#!/bin/bash

#PBS -N Hi_Thomas
#PBS -l select=16:node_type=naples:mpiprocs=64
#PBS -l walltime=00:20:00
 
module load mpi/hpe/mpt/2.19
mpirun -np 1024 ./hi.hpe

To submit the job script execute

qsub Job.hi.hpe.pbs


pure MPI job using OpenMPI

Here is a simple pbs job script:

#!/bin/bash

#PBS -N Hi_Thomas
#PBS -l select=16:node_type=naples:mpiprocs=64
#PBS -l walltime=00:20:00
 
module load mpi/openmpi/4.0.1-gnu-9.1.0
mpirun -np 1024 --map-by core --bind-to core ./hi.hpe

hybrid MPI/OpenMP job using HPE MPI

To run a MPI application with 64 Processes and two OpenMP threads per process on two compute nodes, include the following in the pbs job script:

#!/bin/bash

#PBS -N Hi_MPI_OpenMP
#PBS -l select=2:node_type=naples:mpiprocs=32:ompthreads=2
#PBS -l walltime=00:20:00
 
module load mpi/hpe/mpt/2.19
export OMP_NUM_THREADS=2
mpirun -np 64 omplace  -nt 2 [-vv] ./hi.mpiomp

The omplace command helps with the placement of OpenMP threads within an MPI program. In the above example, the threads in a 64-process MPI program with two threads per process are placed as follows:

  • Rank 0, thread 0 on core 0 of socket 0 on compute node 0
  • Rank 0, thread 1 on core 1 of socket 0 on compute node 0
  • ...
  • Rank 15, thread 1 on core 31 of socket 0 on compute node 0
  • Rank 16, thread 0 on core 0 of socket 1 on compute node 0
  • ...
  • Rank 31, thread 1 on core 31 of socket 1 on compute node 0
  • Rank 32, thread 0 on core 0 of socket 0 on compute node 1
  • ...
  • Rank 63, thread 1 on core 31 of socket 1 on compute node 1

The optional -vv parameter print out the placement of the processes and threads to standard output.
Warning: Due to the limited scaling of the standard output, you should not use the optional parameter -vv for medium and large jobs.

hybrid MPI/OpenMP job using HPE MPI and hyperthreads

The job described before can be run on the same physical resources with twice the number of processes respectively threads by means of hyperthreads as follows:

#!/bin/bash

#PBS -N Hi_MPI_OpenMP_HT
#PBS -l select=2:node_type=naples:mpiprocs=64:ompthreads=2
#PBS -l walltime=00:20:00
 
module load mpi/hpe/mpt/2.19
export OMP_NUM_THREADS=2
mpirun -np 128 omplace  -nt 2 [-vv] ./hi.mpiomp

Ranks will be placed as follows:

  • Rank 0, thread 0 on logical core 0 of core 0 of socket 0 on compute node 0
  • Rank 0, thread 1 on logical core 0 of core 1 of socket 0 on compute node 0
  • ...
  • Rank 15, thread 1 on logical core 0 of core 31 of socket 0 on compute node 0
  • Rank 16, thread 0 on logical core 0 of core 0 of socket 1 on compute node 0
  • ...
  • Rank 31, thread 1 on logical core 0 of core 31 of socket 1 on compute node 0
  • Rank 32, thread 0 on logical core 1 of core 0 of socket 0 on compute node 0
  • ...
  • Rank 63, thread 1 on logical core 1 of core 31 of socket 1 on compute node 0
  • Rank 64, thread 0 on logical core 0 of core 0 of socket 0 on compute node 1
  • ...
  • Rank 127, thread 1 on logical core 1 of core 31 of socket 1 on compute node 1

pure MPI job with stride > 1

If you need to let cores unused, do as follows in order to anyway uniformly distribute processes over cores:

#!/bin/bash

#PBS -N Hi_Thomas
#PBS -l select=1:node_type=naples:mpiprocs=16
#PBS -l walltime=00:20:00
 
module load mpi/hpe/mpt/2.19
mpirun -np 16 omplace -c 0-63:st=4 ./hi.hpe

This will start processes on cores 0, 4, 8, etc., i.e. with a stride of 4 (which means having one process per CCX respectively L3 slice (cf. introduction slides)).