Best practices for jobs

Do Not Run Production Jobs in Global Homes

As a general best practice,users should do production runs from $SCRATCH instead of $HOME.

$HOME is meant for permanent and relatively small storage. It is not tuned to perform well for parallel jobs. Home is perfect for storing files such as source codes and shell scripts, etc. Please note that while building software in /global/home is generally good, it is best to install dynamic libraries that are used on compute nodes in global common for best performance.

$SCRATCH is meant for large and temporary storage. It is optimized for read and write operations. $SCRATCH is perfect for staging data and performing parallel computations. Running in $SCRATCH also helps to improve the responsiveness of the global file systems (global homes and global project) in general.

Specify account

For users who are members of multiple NERSC repositories charges are made to the default account, as set in NIM, unless the #SBATCH --account=<NERSC repository> flag has been set. It is good practice to always set the account flag to ensure the appropriate allocation is charged.

Time Limits

Due to backfill scheduling short and variable length jobs generally start quickly resulting in much better job throughput.

#SBATCH --time-min=<lower_bound>
#SBATCH --time=<upper_bound>

Long Running Jobs

Simulations which must run for a long period of time achieve the best throughput when composed of main small jobs utilizing checkpoint/restart chained together.

Improve Efficiency by Preparing User Environment Before Running

When compute nodes are allocated for a batch job, all commands other than the srun command, such as: load modules, set up runtime environment variables, compile applications, and prepare input data, etc., will run on the head compute node (the first compute node in the allocated nodes pool). Running on a compute node is much more inefficient than running on a login node. It also creates a burden on the global home file system.

Using the Linux here document as in the example below will run those commands to prepare user environment for the batch job on the login node, to help improve job efficiency and save computing cost of the batch job. It can also help to alleviate the burden on the global home file system. This script also keeps the user environment needed for the batch job in a single file.


This is an example to prepare user environment on a login node and propagate this environment to a batch job, and submit the batch job. This can be accomplished in a single script.

You could do so by preparing a file named "" in the example below, and run it as "./" on a login node. This script:

  • Sets up the user environment for the batch job first on a login node, such as loading modules, set environment variables, or copy input files, etc.;
  • Creates a batch script named "";
  • Submits "", this job will inherit the user environment just set earlier in the script.
#!/bin/bash -l

# Submit this script as: "./" instead of "sbatch"

# Prepare user env needed for Slurm batch job
# such as module load, setup runtime environment variables, or copy input files, etc.
# Basically, these are the commands you usually run ahead of the srun command 

module load cray-netcdf

# Generate the Slurm batch script below with the here document, 
# then when sbatch the script later, the user env set up above will run on the login node
# instead of on a head compute node (if included in the Slurm batch script),
# and inherited into the batch job.

cat << EOF > 
#SBATCH -t 30:00
#SBATCH -q debug
#SBATCH -C haswell

srun -n 16 -c 32 --cpu_bind=cores ./myapp.exe 

# Other commands needed after srun, such as copy your output filies,
# should still be included in the Slurm script.
cp <my_output_file> <target_location>/.

# Now submit the batch job

I/O Performance

Cori has dedicated large, local, parallel scratch file systems. The scratch file systems are intended for temporary uses such as storage of checkpoints or application input and output. Data and I/O intensive applications should use the local scratch (or Burst Buffer) filesystems.

These systems should be referenced with the environment variable $SCRATCH.


On Cori the Burst Buffer offers the best I/O performance.


Scratch filesystems are not backed up and old files are subject to purging.

File System Licenses

A batch job will not start if the specified file system is unavailable due to maintenance or an outage or if a performance issue with filesystem is detected.

#SBATCH --license=SCRATCH,project


  • cscratch1 (or SCRATCH)
  • project
  • projecta
  • projectb
  • dna
  • seqfs
  • cvmfs

Running Large Jobs (over 1500 MPI tasks)

Large jobs may take a longer to start up, especially on KNL nodes. The srun option --bcast=<destination_path> is recommended for large jobs requesting over 1500 MPI tasks. By default Slurm loads the executable to the allocated compute nodes from the current working directory, this may take long time when the file system (where the executable resides) is slow. With the --bcast=/tmp/myjob, the executable will be copied to the /tmp/myjob directory. Since /tmp is part of the memory on the compute nodes, it can speed up the job startup time.

% sbcast --compress=lz4 ./mycode.exe /tmp/mycode.exe     # here -C is to compress first
% srun <srun options> /tmp/mycode.exe
# or in the case of when numactl is not needed: % srun --bcast=/tmp/mycode.exe --compress=lz4 <srun options> ./mycode.exe

Network locality

For jobs which are sensitive to interconnect (MPI) performance and utilize less than ~300 nodes it is possible to request that all nodes are in a single Aries dragonfly group.

Slurm has a concept of "switches" which on Cori are configured to map to Aries electrical groups. Since this places an additional constraint on the scheduler a maximum time to wait for the requested topology can be specified.


Wait up to 60 minutes

sbatch --switches=1@60

Additional details and information

Core Specialization

Core specialization is a feature designed to isolate system overhead (system interrupts, etc.) to designated cores on a compute node. It is generally helpful for running on KNL, especially if the application does not plan to use all physical cores on a 68-core compute node. Set aside 2 or 4 cores for core specialization is recommended.

The srun flag for core specialization is "-S" or "--core-spec". It only works in a batch script with sbatch. It can not be requested as a flag with salloc for interactive batch, since salloc is already a wrapper script for srun.

Process Placement

Several mechanisms exist to control process placement on NERSC's Cray systems. Application performance can depend strongly on placement depending on the communication pattern and other computational characteristics.

Examples are run on Cori.


user@nid01041:~> srun -n 8 -c 2|sort -nk 4
Hello from rank 0, on nid01041. (core affinity = 0-63)
Hello from rank 1, on nid01041. (core affinity = 0-63)
Hello from rank 2, on nid01111. (core affinity = 0-63)
Hello from rank 3, on nid01111. (core affinity = 0-63)
Hello from rank 4, on nid01118. (core affinity = 0-63)
Hello from rank 5, on nid01118. (core affinity = 0-63)
Hello from rank 6, on nid01282. (core affinity = 0-63)
Hello from rank 7, on nid01282. (core affinity = 0-63)


The MPICH_RANK_REORDER_METHOD environment variable is used to specify other types of MPI task placement. For example, setting it to 0 results in a round-robin placement:

user@nid01041:~> MPICH_RANK_REORDER_METHOD=0 srun -n 8 -c 2|sort -nk 4
Hello from rank 0, on nid01041. (core affinity = 0-63)
Hello from rank 1, on nid01111. (core affinity = 0-63)
Hello from rank 2, on nid01118. (core affinity = 0-63)
Hello from rank 3, on nid01282. (core affinity = 0-63)
Hello from rank 4, on nid01041. (core affinity = 0-63)
Hello from rank 5, on nid01111. (core affinity = 0-63)
Hello from rank 6, on nid01118. (core affinity = 0-63)
Hello from rank 7, on nid01282. (core affinity = 0-63)

There are other modes available with the MPICH_RANK_REORDER_METHOD environment variable, including one which lets the user provide a file called MPICH_RANK_ORDER which contains a list of each task's placement on each node. These options are described in detail in the intro_mpi man page.


For MPI applications which perform a large amount of nearest-neighbor communication, e.g., stencil-based applications on structured grids, Cray provides a tool in the perftools-base module called grid_order which can generate a MPICH_RANK_ORDER file automatically by taking as parameters the dimensions of the grid, core count, etc. For example, to place MPI tasks in row-major order on a Cartesian grid of size (4, 4, 4), using 32 tasks per node on Cori:

cori$ module load perftools-base
cori$ grid_order -R -c 32 -g 4,4,4
# grid_order -R -Z -c 32 -g 4,4,4
# Region 3: 0,0,1 (0..63)

One can then save this output to a file called MPICH_RANK_ORDER and then set MPICH_RANK_REORDER_METHOD=3 before running the job, which tells Cray MPI to read the MPICH_RANK_ORDER file to set the MPI task placement. For more information, please see the man page man grid_order (available when the perftools-base module is loaded) on Cori.


Huge pages are virtual memory pages which are bigger than the default page size of 4K bytes. Huge pages can improve memory performance for common access patterns on large data sets since it helps to reduce the number of virtual to physical address translations than compated with using the default 4K. Huge pages also increase the maximum size of data and text in a program accessible by the high speed network, and reduce the cost of accessing memory, such as in the case of many MPI_Alltoall operations. Using hugepages can help to reduce the application runtime variability.

To use hugepages for an application (with the 2M hugepages as an example):

module load craype-hugepages2M
cc -o mycode.exe mycode.c
And also load the same hugepages module at runtime.

The craype-hugepages2M module is loaded by deafult on Cori. Users could unload the craype-hugepages2M module explicitly to disable the hugepages usage.


The craype-hugepages2M module is loaded by default since the Cori CLE7 upgrade on July 30, 2019.

Due to the hugepages memory fragmentation issue, applications may get "Cannot allocate memory" warnings or errors when there are not enough hugepages on the compute node, such as:

libhugetlbfs [nid000xx:xxxxx]: WARNING: New heap segment map at 0x10000000 failed: Cannot allocate memory
The verbosity level of libhugetlbfs HUGETLB_VERBOSE is set to 0 on Cori by default to surpress debugging messages. Users can adjust this value to obtain more info.

When to Use Huge Pages

  • For MPI applications, map the static data and/or heap onto huge pages.
  • For an application which uses shared memory, which needs to be concurrently registered with the high speed network drivers for remote communication.
  • For SHMEM applications, map the static data and/or private heap onto huge pages.
  • For applications written in Unified Parallel C, Coarray Fortran, and other languages based on the PGAS programming model, map the static data and/or private heap onto huge pages.
  • For an application doing heavy I/O.
  • To improve memory performance for common access patterns on large data sets.

When to Avoid Huge Pages

  • Applications sometimes consist of many steering programs in addition to the core application. Applying huge page behavior to all processes would not provide any benefit and would consume huge pages that would otherwise benefit the core application. The runtime environment variable HUGETLB_RESTRICT_EXE can be used to specify the susbset of the programs to use hugepages.
  • For certain applications if using hugepages either causes issues or slowing down performances, users can explicitly unload the craype-hugepages2M module. One such example is that when an application forks more subprocesses (such as pthreads) and allocate memory, the newly allocated memory are the small 4K pages.

Serial Jobs

Users requiring large numbers of serial jobs have several options at NERSC.