Best practices for jobs¶
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.
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
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.
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
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 job.sh
Additional details and information
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.
Several mechanisms exsist 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 check-mpi.intel.cori|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)
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 check-mpi.intel.cori|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) 0,1,2,3,16,17,18,19,32,33,34,35,48,49,50,51,4,5,6,7,20,21,22,23,36,37,38,39,52,53,54,55 8,9,10,11,24,25,26,27,40,41,42,43,56,57,58,59,12,13,14,15,28,29,30,31,44,45,46,47,60,61,62,63
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 base 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
The craype-hugepages2M module will be loaded by deafult on Cori after the OS upgrade to CLE7 in July 2019, users do not need to explicitly load the hugepage modules at compile time and runtime after that. Users could unload the craype-hugepages2M module explicitly to disable the hugepages usage.
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
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.
Users requiring large numbers of serial jobs have several options at NERSC.