Frontier User Guide

System Overview

Frontier is a HPE Cray EX supercomputer located at the Oak Ridge Leadership Computing Facility. With a theoretical peak double-precision performance of approximately 2 exaflops (2 quintillion calculations per second), it is the fastest system in the world for a wide range of traditional computational science applications. The system has 74 Olympus rack HPE cabinets, each with 128 AMD compute nodes, and a total of 9,408 AMD compute nodes.

Frontier Compute Nodes

Each Frontier compute node consists of [1x] 64-core AMD “Optimized 3rd Gen EPYC” CPU (with 2 hardware threads per physical core) with access to 512 GB of DDR4 memory. Each node also contains [4x] AMD MI250X, each with 2 Graphics Compute Dies (GCDs) for a total of 8 GCDs per node. The programmer can think of the 8 GCDs as 8 separate GPUs, each having 64 GB of high-bandwidth memory (HBM2E). The CPU is connected to each GCD via Infinity Fabric CPU-GPU, allowing a peak host-to-device (H2D) and device-to-host (D2H) bandwidth of 36+36 GB/s. The 2 GCDs on the same MI250X are connected with Infinity Fabric GPU-GPU with a peak bandwidth of 200 GB/s. The GCDs on different MI250X are connected with Infinity Fabric GPU-GPU in the arrangement shown in the Frontier Node Diagram below, where the peak bandwidth ranges from 50-100 GB/s based on the number of Infinity Fabric connections between individual GCDs.

Note

TERMINOLOGY:

The 8 GCDs contained in the 4 MI250X will show as 8 separate GPUs according to Slurm, ROCR_VISIBLE_DEVICES, and the ROCr runtime, so from this point forward in the quick-start guide, we will simply refer to the GCDs as GPUs.

Note

There are [4x] NUMA domains per node and [2x] L3 cache regions per NUMA for a total of [8x] L3 cache regions. The 8 GPUs are each associated with one of the L3 regions as follows:

NUMA 0:

• hardware threads 000-007, 064-071 | GPU 4
• hardware threads 008-015, 072-079 | GPU 5

NUMA 1:

• hardware threads 016-023, 080-087 | GPU 2
• hardware threads 024-031, 088-095 | GPU 3

NUMA 2:

• hardware threads 032-039, 096-103 | GPU 6
• hardware threads 040-047, 104-111 | GPU 7

NUMA 3:

• hardware threads 048-055, 112-119 | GPU 0
• hardware threads 056-063, 120-127 | GPU 1

Node Types

On Frontier, there are three major types of nodes you will encounter: Login, Launch, and Compute. While all of these are similar in terms of hardware (see: Frontier Compute Nodes), they differ considerably in their intended use.

Node Type	Description
Login	When you connect to Frontier, you’re placed on a login node. This is the place to write/edit/compile your code, manage data, submit jobs, etc. You should never launch parallel jobs from a login node nor should you run threaded jobs on a login node. Login nodes are shared resources that are in use by many users simultaneously.
Compute	Most of the nodes on Frontier are compute nodes. These are where your parallel job executes. They’re accessed via the srun command.

System Interconnect

The Frontier nodes are connected with [4x] HPE Slingshot 200 Gbps (25 GB/s) NICs providing a node-injection bandwidth of 800 Gbps (100 GB/s).

File Systems

Frontier is connected to the center-wide IBM Spectra filesystem called Alpine, which provides 250 PB of storage capacity with a peak write speed of 2.5 TB/s. Frontier also has access to the center-wide NFS-based filesystem (which provides user and project home areas). Each compute node has two 1.92TB Non-Volatile Memory storage devices. See Data and Storage for more information.

Frontier connects to the center’s High Performance Storage System (HPSS) - for user and project archival storage - users can log in to the Data Transfer Nodes (DTNs) to move data to/from HPSS.

Operating System

Frontier is running SUSE Linux Enterprise Server (SLES) version 15.3

GPUs

Each Frontier compute node contains 4 AMD MI250X. The AMD MI250X has a peak performance of 53 TFLOPS in double-precision for modeling and simulation. Each MI250X contains 2 GPUs, where each GPU has a peak performance of 26.5 TFLOPS (double-precision), 110 compute units, and 64 GB of high-bandwidth memory (HBM2) which can be accessed at a peak of 1.6 TB/s. The 2 GPUs on an MI250X are connected with Infinity Fabric with a bandwidth of 200 GB/s (in each direction simultaneously).

Connecting

To connect to Frontier, ssh to frontier.olcf.ornl.gov. For example:

$ ssh <username>@frontier.olcf.ornl.gov

For more information on connecting to OLCF resources, see Connecting for the first time.

---

Data and Storage

For more detailed information about center-wide file systems and data archiving available on Frontier, please refer to the pages on Data Storage and Transfers. The subsections below give a quick overview of NFS, GPFS, and HPSS storage spaces as well as the on node NVMe “Burst Buffers” (SSDs).

NFS

Area	Path	Type	Permissions	Quota	Backups	Purged	Retention	On Compute Nodes
User Home	/ccs/home/[userid]	NFS	User set	50 GB	Yes	No	90 days	Read-only
Project Home	/ccs/proj/[projid]	NFS	770	50 GB	Yes	No	90 days	Read-only

GPFS

Area	Path	Type	Permissions	Quota	Backups	Purged	Retention	On Compute Nodes
Member Work	/gpfs/alpine/[projid]/scratch/[userid]	Spectrum Scale	700	50 TB	No	90 days	N/A	Yes
Project Work	/gpfs/alpine/[projid]/proj-shared	Spectrum Scale	770	50 TB	No	90 days	N/A	Yes
World Work	/gpfs/alpine/[projid]/world-shared	Spectrum Scale	775	50 TB	No	90 days	N/A	Yes

Note

Frontier will be migrating from the Alpine Spectrum Scale filesystem to the Orion Lustre filesystem in the coming months. Details will be added to this section as they become available.

HPSS Archival Storage

Please note that the HPSS is not mounted directly onto Frontier nodes. There are two main methods for accessing and moving data to/from the HPSS. The first is to use the command line utilities hsi and htar. The second is to use the Globus data transfer service. See HPSS Data Archival System for more information on both of these methods.

Area	Path	Type	Permissions	Quota	Backups	Purged	Retention	On Compute Nodes
Member Archive	/hpss/prod/[projid]/users/$USER	HPSS	700	100 TB	No	No	90 days	No
Project Archive	/hpss/prod/[projid]/proj-shared	HPSS	770	100 TB	No	No	90 days	No
World Archive	/hpss/prod/[projid]/world-shared	HPSS	775	100 TB	No	No	90 days	No

NVMe

Each compute node on Frontier has [2x] 1.92TB Non-Volatile Memory (NVMe) storage devices (SSDs), colloquially known as a “Burst Buffer” with a peak sequential performance of 5500 MB/s (read) and 2000 MB/s (write). The purpose of the Burst Buffer system is to bring improved I/O performance to appropriate workloads. Users are not required to use the NVMes. Data can also be written directly to the parallel filesystem.

The NVMes on Frontier are local to each node.

NVMe Usage

To use the NVMe, users must request access during job allocation using the -C nvme option to sbatch, salloc, or srun. Once the devices have been granted to a job, users can access them at /mnt/bb/<userid>. Users are responsible for moving data to/from the NVMe before/after their jobs. Here is a simple example script:

#!/bin/bash
#SBATCH -A <projid>
#SBATCH -J nvme_test
#SBATCH -o %x-%j.out
#SBATCH -t 00:05:00
#SBATCH -p batch
#SBATCH -N 1
#SBATCH -C nvme

date

# Change directory to user scratch space (GPFS)
cd /gpfs/alpine/<projid>/scratch/<userid>

echo " "
echo "*****ORIGINAL FILE*****"
cat test.txt
echo "***********************"

# Move file from GPFS to SSD
mv test.txt /mnt/bb/<userid>

# Edit file from compute node
srun -n1 hostname >> /mnt/bb/<userid>/test.txt

# Move file from SSD back to GPFS
mv /mnt/bb/<userid>/test.txt .

echo " "
echo "*****UPDATED FILE******"
cat test.txt
echo "***********************"

And here is the output from the script:

$ cat nvme_test-<jobid>.out

*****ORIGINAL FILE*****
This is my file. There are many like it but this one is mine.
***********************

*****UPDATED FILE******
This is my file. There are many like it but this one is mine.
frontier0123
***********************

AMD GPUs

The AMD Instinct MI200 is built on advanced packaging technologies enabling two Graphic Compute Dies (GCDs) to be integrated into a single package in the Open Compute Project (OCP) Accelerator Module (OAM) in the MI250 and MI250X products. Each GCD is build on the AMD CDNA 2 architecture. A single Frontier node contains 4 MI250X OAMs for the total of 8 GCDs.

Note

The Slurm workload manager and the ROCr runtime treat each GCD as a separate GPU and visibility can be controlled using the ROCR_VISIBLE_DEVICES environment variable. Therefore, from this point on, the Frontier guide simply refers to a GCD as a GPU.

Each GPU contains 110 Compute Units (CUs) grouped in 4 Compute Engines (CEs). Physically, each GPU contains 112 CUs, but two are disabled. A command processor in each GPU receives API commands and transforms them into compute tasks. Compute tasks are managed by the 4 compute engines, which dispatch wavefronts to compute units. All wavefronts from a single workgroup are assigned to the same CU. In CUDA terminology, workgroups are “blocks”, wavefronts are “warps”, and work-items are “threads”. The terms are often used interchangeably.

The 110 CUs in each GPU deliver peak performance of 26.5 TFLOPS in double precision. Also, each GPU contains 64 GB of high-bandwidth memory (HBM2) accessible at a peak bandwidth of 1.6 TB/s. The 2 GPUs in an MI250X are connected with [4x] GPU-to-GPU Infinity Fabric links providing 200+200 GB/s of bandwidth. (Consult the diagram in the Frontier Compute Nodes section for information on how the accelerators are connected to each other, to the CPU, and to the network.

Note

The X+X GB/s notation describes bidirectional bandwidth, meaning X GB/s in each directions.

AMD vs NVIDIA Terminology

AMD	NVIDIA
Work-items or Threads	Threads
Workgroup	Block
Wavefront	Warp
Grid	Grid

We will be using these terms interchangeably as they refer to the same concepts in GPU programming, with the exception that we will only be using “wavefront” (which refers to a unit of 64 threads) instead of “warp” (which refers to a unit of 32 threads) as they mean different things.

Blocks (workgroups), Threads (work items), Grids, Wavefronts

When kernels are launched on a GPU, a “grid” of thread blocks are created, where the number of thread blocks in the grid and the number of threads within each block are defined by the programmer. The number of blocks in the grid (grid size) and the number of threads within each block (block size) can be specified in one, two, or three dimensions during the kernel launch. Each thread can be identified with a unique id within the kernel, indexed along the X, Y, and Z dimensions.

• Number of blocks that can be specified along each dimension in a grid: (2147483647, 2147483647, 2147483647)
• Max number of threads that can be specified along each dimension in a block: (1024, 1024, 1024)
  • However, the total of number of threads in a block has an upper limit of 1024 [i.e. (size of x dimension * size of y dimension * size of z dimension) cannot exceed 1024].

Each block (or workgroup) of threads is assigned to a single Compute Unit i.e. a single block won’t be split across multiple CUs. The threads in a block are scheduled in units of 64 threads called wavefronts (similar to warps in CUDA, but warps only have 32 threads instead of 64). When launching a kernel, up to 64KB of block level shared memory called the Local Data Store (LDS) can be statically or dynamically allocated. This shared memory between the threads in a block allows the threads to access block local data with much lower latency compared to using the HBM since the data is in the compute unit itself.

The Compute Unit

Each CU has 4 Matrix Core Units (the equivalent of NVIDIA’s Tensor core units) and 4 16-wide SIMD units. For a vector instruction that uses the SIMD units, each wavefront (which has 64 threads) is assigned to a single 16-wide SIMD unit such that the wavefront as a whole executes the instruction over 4 cycles, 16 threads per cycle. Since other wavefronts occupy the other three SIMD units at the same time, the total throughput still remains 1 instruction per cycle. Each CU maintains an instructions buffer for 10 wavefronts and also maintains 256 registers where each register is 64 4-byte wide entries.

HIP

The Heterogeneous Interface for Portability (HIP) is AMD’s dedicated GPU programming environment for designing high performance kernels on GPU hardware. HIP is a C++ runtime API and programming language that allows developers to create portable applications on different platforms. This means that developers can write their GPU applications and with very minimal changes be able to run their code in any environment. The API is very similar to CUDA, so if you’re already familiar with CUDA there is almost no additional work to learn HIP.

Programming Environment

Frontier users are provided with many pre-installed software packages and scientific libraries. To facilitate this, environment management tools are used to handle necessary changes to the shell.

Environment Modules (Lmod)

Environment modules are provided through Lmod, a Lua-based module system for dynamically altering shell environments. By managing changes to the shell’s environment variables (such as PATH, LD_LIBRARY_PATH, and PKG_CONFIG_PATH), Lmod allows you to alter the software available in your shell environment without the risk of creating package and version combinations that cannot coexist in a single environment.

General Usage

The interface to Lmod is provided by the module command:

Command	Description
module -t list	Shows a terse list of the currently loaded modules
module avail	Shows a table of the currently available modules
module help <modulename>	Shows help information about <modulename>
module show <modulename>	Shows the environment changes made by the <modulename> modulefile
module spider <string>	Searches all possible modules according to <string>
module load <modulename> [...]	Loads the given <modulename>(s) into the current environment
module use <path>	Adds <path> to the modulefile search cache and MODULESPATH
module unuse <path>	Removes <path> from the modulefile search cache and MODULESPATH
module purge	Unloads all modules
module reset	Resets loaded modules to system defaults
module update	Reloads all currently loaded modules

Searching for Modules

Modules with dependencies are only available when the underlying dependencies, such as compiler families, are loaded. Thus, module avail will only display modules that are compatible with the current state of the environment. To search the entire hierarchy across all possible dependencies, the spider sub-command can be used as summarized in the following table.

Command	Description
module spider	Shows the entire possible graph of modules
module spider <modulename>	Searches for modules named <modulename> in the graph of possible modules
module spider <modulename>/<version>	Searches for a specific version of <modulename> in the graph of possible modules
module spider <string>	Searches for modulefiles containing <string>

Compilers

Cray, AMD, and GCC compilers are provided through modules on Frontier. The Cray and AMD compilers are both based on LLVM/Clang. There is also a system/OS versions of GCC available in /usr/bin. The table below lists details about each of the module-provided compilers. Please see the following Compiling section for more detailed inforation on how to compile using these modules.

Cray Programming Environment and Compiler Wrappers

Cray provides PrgEnv-<compiler> modules (e.g., PrgEnv-cray) that load compatible components of a specific compiler toolchain. The components include the specified compiler as well as MPI, LibSci, and other libraries. Loading the PrgEnv-<compiler> modules also defines a set of compiler wrappers for that compiler toolchain that automatically add include paths and link in libraries for Cray software. Compiler wrappers are provided for C (cc), C++ (CC), and Fortran (ftn).

Note

Use the -craype-verbose flag to display the full include and link information used by the Cray compiler wrappers. This must be called on a file to see the full output (e.g., CC -craype-verbose test.cpp).

MPI

The MPI implementation available on Frontier is Cray’s MPICH, which is “GPU-aware” so GPU buffers can be passed directly to MPI calls.

---

Compiling

Compilers

Cray, AMD, and GCC compilers are provided through modules on Frontier. The Cray and AMD compilers are both based on LLVM/Clang. There is also a system/OS versions of GCC available in /usr/bin. The table below lists details about each of the module-provided compilers.

Note

It is highly recommended to use the Cray compiler wrappers (cc, CC, and ftn) whenever possible. See the next section for more details.

Vendor	Programming Environment	Compiler Module	Language	Compiler Wrapper	Compiler
Cray	PrgEnv-cray	cce	C	cc	craycc
			C++	CC	craycxx or crayCC
			Fortran	ftn	crayftn
AMD	PrgEnv-amd	rocm	C	cc	amdclang
			C++	CC	amdclang++
			Fortran	ftn	amdflang
GCC	PrgEnv-gnu	gcc	C	cc	${GCC_PATH}/bin/gcc
			C++	CC	${GCC_PATH}/bin/g++
			Fortran	ftn	${GCC_PATH}/bin/gfortran

Cray Programming Environment and Compiler Wrappers

Cray provides PrgEnv-<compiler> modules (e.g., PrgEnv-cray) that load compatible components of a specific compiler toolchain. The components include the specified compiler as well as MPI, LibSci, and other libraries. Loading the PrgEnv-<compiler> modules also defines a set of compiler wrappers for that compiler toolchain that automatically add include paths and link in libraries for Cray software. Compiler wrappers are provided for C (cc), C++ (CC), and Fortran (ftn).

Note

Use the -craype-verbose flag to display the full include and link information used by the Cray compiler wrappers. This must be called on a file to see the full output (e.g., CC -craype-verbose test.cpp).

MPI

The MPI implementation available on Frontier is Cray’s MPICH, which is “GPU-aware” so GPU buffers can be passed directly to MPI calls.

Implementation	Module	Compiler	Header Files & Linking
Cray MPICH	cray-mpich	cc, CC, ftn (Cray compiler wrappers)	MPI header files and linking is built into the Cray compiler wrappers
		hipcc	-L${MPICH_DIR}/lib -lmpi -I${MPICH_DIR}/include

GPU-Aware MPI

To use GPU-aware Cray MPICH, users must set the following modules and environment variables:

module load craype-accel-amd-gfx90a
module load rocm

export MPICH_GPU_SUPPORT_ENABLED=1

Note

There are extra steps needed to enable GPU-aware MPI on Frontier, which depend on the compiler that is used (see 1. and 2. below).

1. Compiling with the Cray compiler wrappers, cc or CC

To use GPU-aware Cray MPICH with the Cray compiler wrappers, the following environment variables must be set before compiling. These variables are automatically set by the cray-mpich modulefile:

## These must be set before compiling so the executable picks up GTL
PE_MPICH_GTL_DIR_amd_gfx90a="-L${CRAY_MPICH_ROOTDIR}/gtl/lib"
PE_MPICH_GTL_LIBS_amd_gfx90a="-lmpi_gtl_hsa"

In addition, the following header files and libraries must be included:

-I${ROCM_PATH}/include
-L${ROCM_PATH}/lib -lamdhip64

where the include path implies that #include <hip/hip_runtime.h> is included in the source file.

2. Compiling with hipcc

To use GPU-aware Cray MPICH with hipcc, users must include appropriate headers, libraries, and flags:

-I${MPICH_DIR}/include
-L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa

HIPFLAGS = --amdgpu-target=gfx90a

Determining the Compatibility of Cray MPICH and ROCm

Releases of cray-mpich are each built with a specific version of ROCm, and compatibility across multiple versions is not guaranteed. OLCF will maintain compatible default modules when possible. If using non-default modules, you can determine compatibility by reviewing the Product and OS Dependencies section in the cray-mpich release notes. This can be displayed by running module show cray-mpich/<version>. If the notes indicate compatibility with AMD ROCM X.Y or later, only use rocm/X.Y.Z modules. If using a non-default version of cray-mpich, you must add ${CRAY_MPICH_ROOTDIR}/gtl/lib to either your LD_LIBRARY_PATH at run time or your executable’s rpath at build time.

The compatibility table below was determined by linker testing with all current combinations of cray-mpich and rocm modules on Frontier.

cray-mpich	ROCm
8.1.12	4.5.2, 4.5.0
8.1.14	4.5.2, 4.5.0
8.1.15	5.1.0, 5.0.2, 5.0.0
8.1.16	5.1.0, 5.0.2, 5.0.0
8.1.17	5.1.0, 5.0.2, 5.0.0

OpenMP

This section shows how to compile with OpenMP using the different compilers covered above.

Vendor	Module	Language	Compiler	OpenMP flag (CPU thread)
Cray	cce	C, C++	cc (wraps craycc) CC (wraps crayCC)	-fopenmp
		Fortran	ftn (wraps crayftn)	-homp -fopenmp (alias)
AMD	rocm	C C++ Fortran	cc (wraps amdclang) CC (wraps amdclang++) ftn (wraps amdflang)	-fopenmp
GCC	gcc	C C++ Fortran	cc (wraps $GCC_PATH/bin/gcc) CC (wraps $GCC_PATH/bin/g++) ftn (wraps $GCC_PATH/bin/gfortran)	-fopenmp

OpenMP GPU Offload

This section shows how to compile with OpenMP Offload using the different compilers covered above.

Note

Make sure the craype-accel-amd-gfx90a module is loaded when using OpenMP offload.

Vendor	Module	Language	Compiler	OpenMP flag (GPU)
Cray	cce	C C++	cc (wraps craycc) CC (wraps crayCC)	-fopenmp
		Fortran	ftn (wraps crayftn)	-homp -fopenmp (alias)
AMD	rocm	C C++ Fortran	cc (wraps amdclang) CC (wraps amdclang++) ftn (wraps amdflang) hipcc (requires flags below)	-fopenmp

Note

If invoking amdclang, amdclang++, or amdflang directly for openmp offload, or using hipcc you will need to add:

-fopenmp -target x86_64-pc-linux-gnu -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a.

HIP

This section shows how to compile HIP codes using the Cray compiler wrappers and hipcc compiler driver.

Note

Make sure the craype-accel-amd-gfx90a module is loaded when compiling HIP with the Cray compiler wrappers.

Compiler	Compile/Link Flags, Header Files, and Libraries
CC Only with PrgEnv-cray PrgEnv-amd	CFLAGS = -std=c++11 -D__HIP_ROCclr__ -D__HIP_ARCH_GFX90A__=1 --rocm-path=${ROCM_PATH} --offload-arch=gfx90a -x hip LFLAGS = --rocm-path=${ROCM_PATH} -L${ROCM_PATH}/lib -lamdhip64
hipcc	Can be used directly to compile HIP source files. To see what is being invoked within this compiler driver, issue the command, hipcc --verbose To explicitly target AMD MI250X, use --amdgpu-target=gfx90a

HIP + OpenMP CPU Threading

This section shows how to compile HIP + OpenMP CPU threading hybrid codes.

Note

Make sure the craype-accel-amd-gfx90a module is loaded when compiling HIP with the Cray compiler wrappers.

Vendor	Compiler	Compile/Link Flags, Header Files, and Libraries
AMD/Cray	CC	CFLAGS = -std=c++11 -D__HIP_ROCclr__ -D__HIP_ARCH_GFX90A__=1 --rocm-path=${ROCM_PATH} --offload-arch=gfx90a -x hip -fopenmp LFLAGS = --rocm-path=${ROCM_PATH} -fopenmp -L${ROCM_PATH}/lib -lamdhip64
	hipcc	Can be used to directly compile HIP source files, add -fopenmp flag to enable OpenMP threading To explicitly target AMD MI250X, use --amdgpu-target=gfx90a
GNU	CC	The GNU compilers cannot be used to compile HIP code, so all HIP kernels must be separated from CPU code. During compilation, all non-HIP files must be compiled with CC while HIP kernels must be compiled with hipcc. Then linking must be performed with the CC wrapper. NOTE: When using cmake, HIP code must currently be compiled using amdclang++ instead of hipcc.

---

Running Jobs

Computational work on Frontier is performed by jobs. Jobs typically consist of several componenets:

• A batch submission script
• A binary executable
• A set of input files for the executable
• A set of output files created by the executable

In general, the process for running a job is to:

1. Prepare executables and input files.
2. Write a batch script.
3. Submit the batch script to the batch scheduler.
4. Optionally monitor the job before and during execution.

The following sections describe in detail how to create, submit, and manage jobs for execution on Frontier. Frontier uses SchedMD’s Slurm Workload Manager as the batch scheduling system.

Login vs Compute Nodes

Recall from the System Overview that Frontier contains two node types: Login and Compute. When you connect to the system, you are placed on a login node. Login nodes are used for tasks such as code editing, compiling, etc. They are shared among all users of the system, so it is not appropriate to run tasks that are long/computationally intensive on login nodes. Users should also limit the number of simultaneous tasks on login nodes (e.g. concurrent tar commands, parallel make

Compute nodes are the appropriate place for long-running, computationally-intensive tasks. When you start a batch job, your batch script (or interactive shell for batch-interactive jobs) runs on one of your allocated compute nodes.

Warning

Compute-intensive, memory-intensive, or other disruptive processes running on login nodes may be killed without warning.

Note

Unlike Summit and Titan, there are no launch/batch nodes on Frontier. This means your batch script runs on a node allocated to you rather than a shared node. You still must use the job launcher (srun) to run parallel jobs across all of your nodes, but serial tasks need not be launched with srun.

Simplified Node Layout

To easily visualize job examples (see Process and Thread Mapping Examples further below), the compute node diagram has been simplified to the picture shown below.

In the diagram, each physical core on a Frontier compute node is composed of two logical cores that are represented by a pair of blue and grey boxes. For a given physical core, the blue box represents the logical core of the first hardware thread, where the grey box represents the logical core of the second hardware thread.

Slurm

Frontier uses SchedMD’s Slurm workload manager for scheduling jobs. Proir systems used different schedulers (Summit used IBM’s LSF, many others used Moab). While most of these systems provide similar functionality, but with different commands/options. A comparison of a few important commands is below, but SchedMD provides a more complete reference in their Rosetta Stone of Workload Managers.

Slurm documentation for each command is available on the system via the man utility as well as on the web at https://slurm.schedmd.com/man_index.html. Additional documentation is available at https://slurm.schedmd.com/documentation.html.

Some common Slurm commands are summarized in the table below. More complete examples are given in the Monitoring and Modifying Batch Jobs section of this guide.

Command	Action/Task	LSF Equivalent
squeue	Show the current queue	bjobs
sbatch	Submit a batch script	bsub
salloc	Submit an interactive job	bsub -Is $SHELL
srun	Launch a parallel job	jsrun
sinfo	Show node/partition info	bqueues or bhosts
sacct	View accounting information for jobs/job steps	bacct
scancel	Cancel a job or job step	bkill
scontrol	View or modify job configuration.	bstop, bresume, bmod

Batch Scripts

The most common way to interact with the batch system is via batch scripts. A batch script is simply a shell script with added directives to request various resoruces from or provide certain information to the scheduling system. Aside from these directives, the batch script is simply the series of commands needed to set up and run your job.

To submit a batch script, use the command sbatch myjob.sl

Consider the following batch script:

#!/bin/bash
#SBATCH -A ABC123
#SBATCH -J RunSim123
#SBATCH -o %x-%j.out
#SBATCH -t 1:00:00
#SBATCH -p batch
#SBATCH -N 1024

cd $MEMBERWORK/abc123/Run.456
cp $PROJWORK/abc123/RunData/Input.456 ./Input.456
srun ...
cp my_output_file $PROJWORK/abc123/RunData/Output.456

In the script, Slurm directives are preceded by #SBATCH, making them appear as comments to the shell. Slurm looks for these directives through the first non-comment, non-whitespace line. Options after that will be ignored by Slurm (and the shell).

Line	Description
1	Shell interpreter line
2	OLCF project to charge
3	Job name
4	Job standard output file (%x will be replaced with the job name and %j with the Job ID)
5	Walltime requested (in HH:MM:SS format). See the table below for other formats.
6	Partition (queue) to use
7	Number of compute nodes requested
8	Blank line
9	Change into the run directory
10	Copy the input file into place
11	Run the job ( add layout details )
12	Copy the output file to an appropriate location.

Interactive Jobs

Most users will find batch jobs an easy way to use the system, as they allow you to “hand off” a job to the scheduler, allowing them to focus on other tasks while their job waits in the queue and eventually runs. Occasionally, it is necessary to run interactively, especially when developing, testing, modifying or debugging a code.

Since all compute resources are managed and scheduled by Slurm, it is not possible to simply log into the system and immediately begin running parallel codes interactively. Rather, you must request the appropriate resources from Slurm and, if necessary, wait for them to become available. This is done through an “interactive batch” job. Interactive batch jobs are submitted with the salloc command. Resources are requested via the same options that are passed via #SBATCH in a regular batch script (but without the #SBATCH prefix). For example, to request an interactive batch job with the same resources that the batch script above requests, you would use salloc -A ABC123 -J RunSim123 -t 1:00:00 -p batch -N 1024. Note there is no option for an output file…you are running interactively, so standard output and standard error will be displayed to the terminal.

Common Slurm Options

The table below summarizes options for submitted jobs. Unless otherwise noted, they can be used for either batch scripts or interactive batch jobs. For scripts, they can be added on the sbatch command line or as a #BSUB directive in the batch script. (If they’re specified in both places, the command line takes precedence.) This is only a subset of all available options. Check the Slurm Man Pages for a more complete list.

Option	Example Usage	Description
-A	#SBATCH -A ABC123	Specifies the project to which the job should be charged
-N	#SBATCH -N 1024	Request 1024 nodes for the job
-t	#SBATCH -t 4:00:00	Request a walltime of 4 hours. Walltime requests can be specified as minutes, hours:minutes, hours:minuts:seconds days-hours, days-hours:minutes, or days-hours:minutes:seconds
--threads-per-core	#SBATCH --threads-per-core=2	Number of active hardware threads per core. Can be 1 or 2 (1 is default) Must be used if using --threads-per-core=2 in your srun command.
-d	#SBATCH -d afterok:12345	Specify job dependency (in this example, this job cannot start until job 12345 exits with an exit code of 0. See the Job Dependency section for more information
-C	#SBATCH -C nvme	Request the burst buffer/NVMe on each node be made available for your job. See the Burst Buffers section for more information on using them.
-J	#SBATCH -J MyJob123	Specify the job name (this will show up in queue listings)
-o	#SBATCH -o jobout.%j	File where job STDOUT will be directed (%j will be replaced with the job ID). If no -e option is specified, job STDERR will be placed in this file, too.
-e	#SBATCH -e joberr.%j	File where job STDERR will be directed (%j will be replaced with the job ID). If no -o option is specified, job STDOUT will be placed in this file, too.
--mail-type	#SBATCH --mail-type=END	Send email for certain job actions. Can be a comma-separated list. Actions include BEGIN, END, FAIL, REQUEUE, INVALID_DEPEND, STAGE_OUT, ALL, and more.
--mail-user	#SBATCH --mail-user=user@somewhere.com	Email address to be used for notifications.
--reservation	#SBATCH --reservation=MyReservation.1	Instructs Slurm to run a job on nodes that are part of the specified reservation.

Slurm Environment Variables

Slurm reads a number of environment variables, many of which can provide the same information as the job options noted above. We recommend using the job options rather than environment variables to specify job options, as it allows you to have everything self-contained within the job submission script (rather than having to remember what options you set for a given job).

Slurm also provides a number of environment variables within your running job. The following table summarizes those that may be particularly useful within your job (e.g. for naming output log files):

Variable	Description
$SLURM_SUBMIT_DIR	The directory from which the batch job was submitted. By default, a new job starts in your home directory. You can get back to the directory of job submission with cd $SLURM_SUBMIT_DIR. Note that this is not necessarily the same directory in which the batch script resides.
$SLURM_JOBID	The job’s full identifier. A common use for $SLURM_JOBID is to append the job’s ID to the standard output and error files.
$SLURM_JOB_NUM_NODES	The number of nodes requested.
$SLURM_JOB_NAME	The job name supplied by the user.
$SLURM_NODELIST	The list of nodes assigned to the job.

Job States

A job will transition through several states during its lifetime. Common ones include:

State Code	State	Description
CA	Canceled	The job was canceled (could’ve been by the user or an administrator)
CD	Completed	The job completed successfully (exit code 0)
CG	Completing	The job is in the process of completing (some processes may still be running)
PD	Pending	The job is waiting for resources to be allocated
R	Running	The job is currently running

Job Reason Codes

In addition to state codes, jobs that are pending will have a “reason code” to explain why the job is pending. Completed jobs will have a reason describing how the job ended. Some codes you might see include:

Reason	Meaning
Dependency	Job has dependencies that have not been met
JobHeldUser	Job is held at user’s request
JobHeldAdmin	Job is held at system administrator’s request
Priority	Other jobs with higher priority exist for the partition/reservation
Reservation	The job is waiting for its reservation to become available
AssocMaxJobsLimit	The job is being held because the user/project has hit the limit on running jobs
ReqNodeNotAvail	The requested a particular node, but it’s currently unavailable (it’s in use, reserved, down, draining, etc.)
JobLaunchFailure	Job failed to launch (could due to system problems, invalid program name, etc.)
NonZeroExitCode	The job exited with some code other than 0

Many other states and job reason codes exist. For a more complete description, see the squeue man page (either on the system or online).

Scheduling Policy

In a simple batch queue system, jobs run in a first-in, first-out (FIFO) order. This can lead to inefficient use of the system. If a large job is the next to run, a strict FIFO queue can cause nodes to sit idle while waiting for the large job to start. Backfilling would allow smaller, shorter jobs to use those resources that would otherwise remain idle until the large job starts. With the proper algorithm, they would do so without impacting the start time of the large job. While this does make more efficient use of the system, it encourages the submission of smaller jobs.

The DOE Leadership-Class Job Mandate

As a DOE Leadership Computing Facility, OLCF has a mandate that a large portion of Frontier’s usage come from large, leadership-class (a.k.a. capability) jobs. To ensure that OLCF complies with this directive, we strongly encourage users to run jobs on Frontier that are as large as their code will allow. To that end, OLCF implements queue policies that enable large jobs to run in a timely fashion.

Note

The OLCF implements queue policies that encourage the submission and timely execution of large, leadership-class jobs on Frontier.

The basic priority mechanism for jobs waiting in the queue is the time the job has been waiting in the queue. If your jobs require resources outside these policies such as higher priority or longer walltimes, please contact help@olcf.ornl.gov

Job Priority by Node Count

Jobs are aged according to the job’s requested node count (older age equals higher queue priority). Each job’s requested node count places it into a specific bin. Each bin has a different aging parameter, which all jobs in the bin receive.

Bin	Min Nodes	Max Nodes	Max Walltime (Hours)	Aging Boost (Days)
1	5,645	9,408	24.0	15
2	1,882	5,644	24.0	10
3	184	1,881	12.0	0
4	92	183	6.0	0
5	1	91	2.0	0

batch Queue Policy

The batch queue is the default queue for production work on Frontier. Most work on Frontier is handled through this queue. The following policies are enforced for the batch queue:

• Limit of four eligible-to-run jobs per user. (Jobs in excess of this number will be held, but will move to the eligible-to-run state at the appropriate time.)
• Users may have only 100 jobs queued in the batch queue at any time (this includes jobs in all states). Additional jobs will be rejected at submit time.

Allocation Overuse Policy

Projects that overrun their allocation are still allowed to run on OLCF systems, although at a reduced priority. Like the adjustment for the number of processors requested above, this is an adjustment to the apparent submit time of the job. However, this adjustment has the effect of making jobs appear much younger than jobs submitted under projects that have not exceeded their allocation. In addition to the priority change, these jobs are also limited in the amount of wall time that can be used. For example, consider that job1 is submitted at the same time as job2. The project associated with job1 is over its allocation, while the project for job2 is not. The batch system will consider job2 to have been waiting for a longer time than job1. Additionally, projects that are at 125% of their allocated time will be limited to only 3 running jobs at a time. The adjustment to the apparent submit time depends upon the percentage that the project is over its allocation, as shown in the table below:

% of Allocation Used	Priority Reduction
< 100%	none
>=100% but <=125%	30 days
> 125%	365 days

System Reservation Policy

Projects may request to reserve a set of nodes for a period of time by contacting help@olcf.ornl.gov. If the reservation is granted, the reserved nodes will be blocked from general use for a given period of time. Only users that have been authorized to use the reservation can utilize those resources. Since no other users can access the reserved resources, it is crucial that groups given reservations take care to ensure the utilization on those resources remains high. To prevent reserved resources from remaining idle for an extended period of time, reservations are monitored for inactivity. If activity falls below 50% of the reserved resources for more than (30) minutes, the reservation will be canceled and the system will be returned to normal scheduling. A new reservation must be requested if this occurs.

The requesting project’s allocation is charged according to the time window granted, regardless of actual utilization. For example, an 8-hour, 2,000 node reservation on Frontier would be equivalent to using 16,000 Frontier node-hours of a project’s allocation.

Note

Reservations should not be confused with priority requests. If quick turnaround is needed for a few jobs or for a period of time, a priority boost should be requested. A reservation should only be requested if users need to guarantee availability of a set of nodes at a given time, such as for a live demonstration at a conference.

Job Dependencies

Oftentimes, a job will need data from some other job in the queue, but it’s nonetheless convenient to submit the second job before the first finishes. Slurm allows you to submit a job with constraints that will keep it from running until these dependencies are met. These are specified with the -d option to Slurm. Common dependency flags are summarized below. In each of these examples, only a single jobid is shown but you can specify multiple job IDs as a colon-delimited list (i.e. #SBATCH -d afterok:12345:12346:12346). For the after dependency, you can optionally specify a +time value for each jobid.

Flag	Meaning (for the dependent job)
#SBATCH -d after:jobid[+time]	The job can start after the specified jobs start or are canceled. The optional +time argument is a number of minutes. If specified, the job cannot start until that many minutes have passed since the listed jobs start/are canceled. If not specified, there is no delay.
#SBATCH -d afterany:jobid	The job can start after the specified jobs have ended (regardless of exit state)
#SBATCH -d afternotok:jobid	The job can start after the specified jobs terminate in a failed (non-zero) state
#SBATCH -d afterok:jobid	The job can start after the specified jobs complete successfully (i.e. zero exit code)
#SBATCH -d singleton	Job can begin after any previously-launched job with the same name and from the same user have completed. In other words, serialize the running jobs based on username+jobname pairs.

Monitoring and Modifying Batch Jobs

scontrol hold and scontrol release: Holding and Releasing Jobs

Sometimes you may need to place a hold on a job to keep it from starting. For example, you may have submitted it assuming some needed data was in place but later realized that data is not yet available. This can be done with the scontrol hold command. Later, when the data is ready, you can release the job (i.e. tell the system that it’s now OK to run the job) with the scontrol release command. For example:

scontrol hold 12345	Place job 12345 on hold
scontrol release 12345	Release job 12345 (i.e. tell the system it’s OK to run it)

scontrol update: Changing Job Parameters

There may also be occasions where you want to modify a job that’s waiting in the queue. For example, perhaps you requested 2,000 nodes but later realized this is a different data set and only needs 1,000 nodes. You can use the scontrol update command for this. For example:

scontrol update NumNodes=1000 JobID=12345	Change job 12345’s node request to 1000 nodes
scontrol update TimeLimit=4:00:00 JobID=12345	Change job 12345’s max walltime to 4 hours

scancel: Cancel or Signal a Job

Slurm allows you to signal a job with the scancel command. Typically, this is used to remove a job from the queue. In this use case, you do not need to specify a signal and can simply provide the jobid. For example, scancel 12345.

In addition to removing a job from the queue, the command gives you the ability to send other signals to the job with the -s option. For example, if you want to send SIGUSR1 to a job, you would use scancel -s USR1 12345.

squeue: View the Queue

The squeue command is used to show the batch queue. You can filter the level of detail through several command-line options. For example:

squeue -l	Show all jobs currently in the queue
squeue -l -u $USER	Show all of your jobs currently in the queue

sacct: Get Job Accounting Information

The sacct command gives detailed information about jobs currently in the queue and recently-completed jobs. You can also use it to see the various steps within a batch jobs.

sacct -a -X	Show all jobs (-a) in the queue, but summarize the whole allocation instead of showing individual steps (-X)
sacct -u $USER	Show all of your jobs, and show the individual steps (since there was no -X option)
sacct -j 12345	Show all job steps that are part of job 12345
sacct -u $USER -S 2022-07-01T13:00:00 -o "jobid%5,jobname%25,nodelist%20" -X	Show all of your jobs since 1 PM on July 1, 2022 using a particular output format

scontrol show job: Get Detailed Job Information

In addition to holding, releasing, and updating the job, the scontrol command can show detailed job information via the show job subcommand. For example, scontrol show job 12345.

Srun

The default job launcher for Frontier is srun . The srun command is used to execute an MPI binary on one or more compute nodes in parallel.

Srun Format

srun  [OPTIONS... [executable [args...]]]

Single Command (non-interactive)

$ srun -A <project_id> -t 00:05:00 -p <partition> -N 2 -n 4 --ntasks-per-node=2 ./a.out
<output printed to terminal>

The job name and output options have been removed since stdout/stderr are typically desired in the terminal window in this usage mode.

srun accepts the following common options:

-N	Number of nodes
-n	Total number of MPI tasks (default is 1)
-c, --cpus-per-task=<ncpus>	Logical cores per MPI task (default is 1) When used with --threads-per-core=1: -c is equivalent to physical cores per task
--cpu-bind=threads	Bind tasks to CPUs. threads - (default, recommended) Automatically generate masks binding tasks to threads.
--threads-per-core=<threads>	In task layout, use the specified maximum number of hardware threads per core (default is 1; there are 2 hardware threads per physical CPU core). Must also be set in salloc or sbatch if using --threads-per-core=2 in your srun command.
-m, --distribution=<value>:<value>:<value>	Specifies the distribution of MPI ranks across compute nodes, sockets (L3 regions), and cores, respectively. The default values are block:cyclic:cyclic, see man srun for more information. Currently, the distribution setting for cores (the third “<value>” entry) has no effect on Frontier
--ntasks-per-node=<ntasks>	If used without -n: requests that a specific number of tasks be invoked on each node. If used with -n: treated as a maximum count of tasks per node.
--gpus	Specify the number of GPUs required for the job (total GPUs across all nodes).
--gpus-per-node	Specify the number of GPUs per node required for the job.
--gpu-bind=closest	Binds each task to the GPU which is on the same NUMA domain as the CPU core the MPI rank is running on.
--gpu-bind=map_gpu:<list>	Bind tasks to specific GPUs by setting GPU masks on tasks (or ranks) as specified where <list> is <gpu_id_for_task_0>,<gpu_id_for_task_1>,.... If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. (For example map_gpu:0*4,1*4)
--ntasks-per-gpu=<ntasks>	Request that there are ntasks tasks invoked for every GPU.

Below is a comparison table between srun and jsrun.

Option	jsrun (Summit)	srun (Frontier)
Number of nodes	-nnodes	-N, --nnodes
Number of tasks	defined with resource set	-n, --ntasks
Number of tasks per node	defined with resource set	--ntasks-per-node
Number of CPUs per task	defined with resource set	-c, --cpus-per-task
Number of resource sets	-n, --nrs	N/A
Number of resource sets per host	-r, --rs_per_host	N/A
Number of tasks per resource set	-a, --tasks_per_rs	N/A
Number of CPUs per resource set	-c, --cpus_per_rs	N/A
Number of GPUs per resource set	-g, --gpus_per_rs	N/A
Bind tasks to allocated CPUs	-b, --bind	--cpu-bind
Performance binding preference	-l,--latency_priority	--hint
Specify the task to resource mapping pattern	--launch_distribution	-m, --distribution

Process and Thread Mapping Examples

This section describes how to map processes (e.g., MPI ranks) and process threads (e.g., OpenMP threads) to the CPUs and GPUs on Frontier.

Users are highly encouraged to use the CPU- and GPU-mapping programs used in the following sections to check their understanding of the job steps (i.e., srun commands) they intend to use in their actual jobs.

• For the CPU Mapping and Multithreading sections:

  A simple MPI+OpenMP “Hello, World” program (hello_mpi_omp) will be used to clarify the mappings.

• For the GPU Mapping section:

  An MPI+OpenMP+HIP “Hello, World” program (hello_jobstep) will be used to clarify the GPU mappings.

CPU Mapping

This subsection covers how to map tasks to the CPU without the presence of additional threads (i.e., solely MPI tasks – no additional OpenMP threads).

The intent with both of the following examples is to launch 8 MPI ranks across the node where each rank is assigned its own logical (and, in this case, physical) core. Using the -m distribution flag, we will cover two common approaches to assign the MPI ranks – in a “round-robin” (cyclic) configuration and in a “packed” (block) configuration. Slurm’s Interactive Jobs method was used to request an allocation of 1 compute node for these examples: salloc -A <project_id> -t 30 -p <parition> -N 1

Note

There are many different ways users might choose to perform these mappings, so users are encouraged to clone the hello_mpi_omp program and test whether or not processes and threads are running where intended.

8 MPI Ranks (round-robin)

Assigning MPI ranks in a “round-robin” (cyclic) manner across L3 cache regions (sockets) is the default behavior on Frontier. This mode will assign consecutive MPI tasks to different sockets before it tries to “fill up” a socket.

Recall that the -m flag behaves like: -m <node distribution>:<socket distribution>. Hence, the key setting to achieving the round-robin nature is the -m block:cyclic flag, specifically the cyclic setting provided for the “socket distribution”. This ensures that the MPI tasks will be distributed across sockets in a cyclic (round-robin) manner.

The below srun command will achieve the intended 8 MPI “round-robin” layout:

$ export OMP_NUM_THREADS=1
$ srun -N1 -n8 -c1 --cpu-bind=threads --threads-per-core=1 -m block:cyclic ./hello_mpi_omp | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier144
MPI 001 - OMP 000 - HWT 008 - Node frontier144
MPI 002 - OMP 000 - HWT 016 - Node frontier144
MPI 003 - OMP 000 - HWT 024 - Node frontier144
MPI 004 - OMP 000 - HWT 032 - Node frontier144
MPI 005 - OMP 000 - HWT 040 - Node frontier144
MPI 006 - OMP 000 - HWT 048 - Node frontier144
MPI 007 - OMP 000 - HWT 056 - Node frontier144

Breaking down the srun command, we have:

• -N1: indicates we are using 1 node
• -n8: indicates we are launching 8 MPI tasks
• -c1: indicates we are assigning 1 logical core per MPI task. In this case, because of --threads-per-core=1, this also means 1 physical core per MPI task.
• --cpu-bind=threads: binds tasks to threads
• --threads-per-core=1: use a maximum of 1 hardware thread per physical core (i.e., only use 1 logical core per physical core)
• -m block:cyclic: distribute the tasks in a block layout across nodes (default), and in a cyclic (round-robin) layout across L3 sockets
• ./hello_mpi_omp: launches the “hello_mpi_omp” executable
• | sort: sorts the output

Note

Although the above command used the default settings -c1, --cpu-bind=threads, --threads-per-core=1 and -m block:cyclic, it is always better to be explicit with your srun command to have more control over your node layout. The above command is equivalent to srun -N1 -n8.

As you can see in the node diagram above, this results in the 8 MPI tasks (outlined in different colors) being distributed “vertically” across L3 sockets.

8 MPI Ranks (packed)

Instead, you can assign MPI ranks so that the L3 regions are filled in a “packed” (block) manner. This mode will assign consecutive MPI tasks to the same L3 region (socket) until it is “filled up” or “packed” before assigning a task to a different socket.

Recall that the -m flag behaves like: -m <node distribution>:<socket distribution>. Hence, the key setting to achieving the round-robin nature is the -m block:block flag, specifically the block setting provided for the “socket distribution”. This ensures that the MPI tasks will be distributed in a packed manner.

The below srun command will achieve the intended 8 MPI “packed” layout:

$ export OMP_NUM_THREADS=1
$ srun -N1 -n8 -c1 --cpu-bind=threads --threads-per-core=1 -m block:block ./hello_mpi_omp | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier144
MPI 001 - OMP 000 - HWT 001 - Node frontier144
MPI 002 - OMP 000 - HWT 002 - Node frontier144
MPI 003 - OMP 000 - HWT 003 - Node frontier144
MPI 004 - OMP 000 - HWT 004 - Node frontier144
MPI 005 - OMP 000 - HWT 005 - Node frontier144
MPI 006 - OMP 000 - HWT 006 - Node frontier144
MPI 007 - OMP 000 - HWT 007 - Node frontier144

Breaking down the srun command, the only difference than the previous example is:

• -m block:block: distribute the tasks in a block layout across nodes (default), and in a block (packed) socket layout

As you can see in the node diagram above, this results in the 8 MPI tasks (outlined in different colors) being distributed “horizontally” within a socket, rather than being spread across different L3 sockets like with the previous example.

Multithreading

Because a Frontier compute node has two hardware threads available (2 logical cores per physical core), this enables the possibility of multithreading your application (e.g., with OpenMP threads). Although the additional hardware threads can be assigned to additional MPI tasks, this is not recommended. It is highly recommended to only use 1 MPI task per physical core and to use OpenMP threads instead on any additional logical cores gained when using both hardware threads.

The following examples cover multithreading with hybrid MPI+OpenMP applications. In these examples, Slurm’s Interactive Jobs method was used to request an allocation of 1 compute node: salloc -A <project_id> -t 30 -p <parition> -N 1

Note

There are many different ways users might choose to perform these mappings, so users are encouraged to clone the hello_mpi_omp program and test whether or not processes and threads are running where intended.

2 MPI ranks - each with 2 OpenMP threads

In this example, the intent is to launch 2 MPI ranks, each of which spawn 2 OpenMP threads, and have all of the 4 OpenMP threads run on different physical CPU cores.

First (INCORRECT) attempt

To set the number of OpenMP threads spawned per MPI rank, the OMP_NUM_THREADS environment variable can be used. To set the number of MPI ranks launched, the srun flag -n can be used.

$ export OMP_NUM_THREADS=2
$ srun -N1 -n2 ./hello_mpi_omp | sort

WARNING: Requested total thread count and/or thread affinity may result in
oversubscription of available CPU resources!  Performance may be degraded.
Explicitly set OMP_WAIT_POLICY=PASSIVE or ACTIVE to suppress this message.
Set CRAY_OMP_CHECK_AFFINITY=TRUE to print detailed thread-affinity messages.
WARNING: Requested total thread count and/or thread affinity may result in
oversubscription of available CPU resources!  Performance may be degraded.
Explicitly set OMP_WAIT_POLICY=PASSIVE or ACTIVE to suppress this message.
Set CRAY_OMP_CHECK_AFFINITY=TRUE to print detailed thread-affinity messages.

MPI 000 - OMP 000 - HWT 000 - Node frontier001
MPI 000 - OMP 001 - HWT 000 - Node frontier001
MPI 001 - OMP 000 - HWT 008 - Node frontier001
MPI 001 - OMP 001 - HWT 008 - Node frontier001

The first thing to notice here is the WARNING about oversubscribing the available CPU cores. Also, the output shows each MPI rank did spawn 2 OpenMP threads, but both OpenMP threads ran on the same logical core (for a given MPI rank). This was not the intended behavior; each OpenMP thread was meant to run on its own physical CPU core.

The problem here arises from two default settings; 1) each MPI rank is only allocated 1 logical core (-c 1) and, 2) only 1 hardware thread per physical CPU core is enabled (--threads-per-core=1). When using --threads-per-core=1 and --cpu-bind=threads (the default setting), 1 logical core in -c is equivalent to 1 physical core. So in this case, each MPI rank only has 1 physical core (with 1 hardware thread) to run on - including any threads the process spawns - hence the WARNING and undesired behavior.

Second (CORRECT) attempt

Recall that in this scenario, because of the --threads-per-core=1 setting, 1 logical core is equivalent to 1 physical core when using -c. Therefore, in order for each OpenMP thread to run on its own physical CPU core, each MPI rank should be given 2 physical CPU cores (-c 2). Now the OpenMP threads will be mapped to unique hardware threads on separate physical CPU cores.

$ export OMP_NUM_THREADS=2
$ srun -N1 -n2 -c2 ./hello_mpi_omp | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier001
MPI 000 - OMP 001 - HWT 001 - Node frontier001
MPI 001 - OMP 000 - HWT 008 - Node frontier001
MPI 001 - OMP 001 - HWT 009 - Node frontier001

Now the output shows that each OpenMP thread ran on its own physical CPU core. More specifically (see the Frontier Compute Node diagram), OpenMP thread 000 of MPI rank 000 ran on logical core 000 (i.e., physical CPU core 00), OpenMP thread 001 of MPI rank 000 ran on logical core 001 (i.e., physical CPU core 01), OpenMP thread 000 of MPI rank 001 ran on logical core 008 (i.e., physical CPU core 08), and OpenMP thread 001 of MPI rank 001 ran on logical core 009 (i.e., physical CPU core 09) - as intended.

Third attempt - Using multiple threads per core

To use both available hardware threads per core, the job must be allocated with --threads-per-core=2 (as opposed to only the job step - i.e., srun command). That value will then be inherited by srun unless explcitly overridden with --threads-per-core=1. Because we are using --threads-per-core=2, the usage of -c goes back to purely meaning the amount of logical cores (i.e., it is no longer equivalent to 1 physical core).

$ salloc -N1 -A <project_id> -t <time> -p <partition> --threads-per-core=2

$ export OMP_NUM_THREADS=2
$ srun -N1 -n2 -c2 ./hello_mpi_omp | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier001
MPI 000 - OMP 001 - HWT 064 - Node frontier001
MPI 001 - OMP 000 - HWT 008 - Node frontier001
MPI 001 - OMP 001 - HWT 072 - Node frontier001

Comparing this output to the Frontier Compute Node diagram, we see that each pair of OpenMP threads is contained within a single physical core. MPI rank 000 ran on logical cores 000 and 064 (i.e. physical CPU core 00) and MPI rank 001 ran on logical cores 008 and 072 (i.e. physical CPU core 08).

GPU Mapping

In this sub-section, an MPI+OpenMP+HIP “Hello, World” program (hello_jobstep) will be used to clarify the GPU mappings. Again, Slurm’s Interactive Jobs method was used to request an allocation of 2 compute nodes for these examples: salloc -A <project_id> -t 30 -p <parition> -N 2. The CPU mapping part of this example is very similar to the example used above in the Multithreading sub-section, so the focus here will be on the GPU mapping part.

In general, GPU mapping can be accomplished in different ways. For example, an application might map MPI ranks to GPUs programmatically within the code using, say, hipSetDevice. In this case, since all GPUs on a node are available to all MPI ranks on that node by default, there might not be a need to map to GPUs using Slurm (just do it in the code). However, in another application, there might be a reason to make only a subset of GPUs available to the MPI ranks on a node. It is this latter case that the following examples refer to.

Note

There are many different ways users might choose to perform these mappings, so users are encouraged to clone the hello_jobstep program and test whether or not processes and threads are running where intended.

Warning

Due to the unique architecture of Frontier compute nodes and the way that Slurm currently allocates GPUs and CPU cores to job steps, it is suggested that all 8 GPUs on a node are allocated to the job step to ensure that optimal bindings are possible.

Mapping 1 task per GPU

In the following examples, each MPI rank (and its OpenMP threads) will be mapped to a single GPU.

Example 0: 1 MPI rank with 1 OpenMP thread and 1 GPU (single-node)

Somewhat counterintuitively, this common test case is currently among the most difficult. Slurm ignores GPU bindings for nodes with only a single task, so we do not use --gpu-bind here. We must allocate only a single GPU to ensure that only one GPU is available to the task, and since we get the first GPU available we should bind the task to the CPU closest to the allocated GPU.

$ export OMP_NUM_THREADS=1
$ srun -N1 -n1 -c1 --cpu-bind=map_cpu:48 --gpus=1 ./hello_jobstep

MPI 000 - OMP 000 - HWT 048 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1

Example 1: 8 MPI ranks - each with 2 OpenMP threads and 1 GPU (single-node)

This example launches 8 MPI ranks (-n8), each with 2 physical CPU cores (-c2) to launch 2 OpenMP threads (OMP_NUM_THREADS=2) on. In addition, each MPI rank (and its 2 OpenMP threads) should have access to only 1 GPU. To accomplish the GPU mapping, two new srun options will be used:

• --gpus-per-node specifies the number of GPUs required for the job
• --gpu-bind=closest binds each task to the GPU which is closest.

Note

Without these additional flags, all MPI ranks would have access to all GPUs (which is the default behavior).

$ export OMP_NUM_THREADS=2
$ srun -N1 -n8 -c2 --gpus-per-node=8 --gpu-bind=closest ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 000 - OMP 001 - HWT 001 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 001 - OMP 001 - HWT 009 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 002 - OMP 001 - HWT 017 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 003 - OMP 001 - HWT 025 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 004 - OMP 001 - HWT 033 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 005 - OMP 001 - HWT 041 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 006 - OMP 001 - HWT 049 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 007 - OMP 001 - HWT 057 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

The output from the program contains a lot of information, so let’s unpack it. First, there are different IDs associated with the GPUs so it is important to describe them before moving on. GPU_ID is the node-level (or global) GPU ID, which is labeled as one might expect from looking at the Frontier Node Diagram: 0, 1, 2, 3, 4, 5, 6, 7. RT_GPU_ID is the HIP runtime GPU ID, which can be though of as each MPI rank’s local GPU ID number (with zero-based indexing). So in the output above, each MPI rank has access to only 1 unique GPU - where MPI 000 has access to “global” GPU 4, MPI 001 has access to “global” GPU 5, etc., but all MPI ranks show a HIP runtime GPU ID of 0. The reason is that each MPI rank only “sees” one GPU and so the HIP runtime labels it as “0”, even though it might be global GPU ID 0, 1, 2, 3, 4, 5, 6, or 7. The GPU’s bus ID is included to definitively show that different GPUs are being used.

Here is a summary of the different GPU IDs reported by the example program:

• GPU_ID is the node-level (or global) GPU ID read from ROCR_VISIBLE_DEVICES. If this environment variable is not set (either by the user or by Slurm), the value of GPU_ID will be set to N/A by this program.
• RT_GPU_ID is the HIP runtime GPU ID (as reported from, say hipGetDevice).
• Bus_ID is the physical bus ID associated with the GPUs. Comparing the bus IDs is meant to definitively show that different GPUs are being used.

So the job step (i.e., srun command) used above gave the desired output. Each MPI rank spawned 2 OpenMP threads and had access to a unique GPU. The --gpus-per-node=8 allocated 8 GPUs for node and the --gpu-bind=closest ensured that the closest GPU to each rank was the one used.

Note

This example shows an important peculiarity of the Frontier nodes; the “closest” GPUs to each MPI rank are not in sequential order. For example, MPI rank 000 and its two OpenMP threads ran on hardware threads 000 and 001. As can be seen in the Frontier node diagram, these two hardware threads reside in the same L3 cache region, and that L3 region is connected via Infinity Fabric (blue line in the diagram) to GPU 4. This is an important distinction that can affect performance if not considered carefully.

Example 2: 16 MPI ranks - each with 2 OpenMP threads and 1 GPU (multi-node)

This example will extend Example 1 to run on 2 nodes. As the output shows, it is a very straightforward exercise of changing the number of nodes to 2 (-N2) and the number of MPI ranks to 16 (-n16).

$ export OMP_NUM_THREADS=2
$ srun -N2 -n16 -c2 --gpus-per-node=8 --gpu-bind=closest ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 000 - OMP 001 - HWT 001 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 001 - OMP 001 - HWT 009 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 002 - OMP 001 - HWT 017 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 003 - OMP 001 - HWT 025 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 004 - OMP 001 - HWT 033 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 005 - OMP 001 - HWT 041 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 006 - OMP 001 - HWT 049 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 007 - OMP 001 - HWT 057 - Node frontier001 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 008 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 008 - OMP 001 - HWT 001 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 009 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 009 - OMP 001 - HWT 009 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 010 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 010 - OMP 001 - HWT 017 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 011 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 011 - OMP 001 - HWT 025 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 012 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 012 - OMP 001 - HWT 033 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 013 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 013 - OMP 001 - HWT 041 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 014 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 014 - OMP 001 - HWT 049 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 015 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 015 - OMP 001 - HWT 057 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

Example 3: 8 MPI ranks - each with 2 OpenMP threads and 1 *specific* GPU (single-node)

This example will be very similar to Example 1, but instead of using --gpu-bind=closest to map each MPI rank to the closest GPU, --gpu-bind=map_gpu will be used to map each MPI rank to a specific GPU. The map_gpu option takes a comma-separated list of GPU IDs to specify how the MPI ranks are mapped to GPUs, where the form of the comma-separated list is <gpu_id_for_task_0>, <gpu_id_for_task_1>,....

$ export OMP_NUM_THREADS=2
$ srun -N1 -n8 -c2 --gpus-per-node=8 --gpu-bind=map_gpu:4,5,2,3,6,7,0,1 ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 000 - OMP 001 - HWT 001 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 001 - OMP 001 - HWT 009 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 002 - OMP 001 - HWT 017 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 003 - OMP 001 - HWT 025 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 004 - OMP 001 - HWT 033 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 005 - OMP 001 - HWT 041 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 006 - OMP 001 - HWT 049 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 007 - OMP 001 - HWT 057 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

Here, the output is the same as the results from Example 1. This is because the 8 GPU IDs in the comma-separated list happen to specify the GPUs within the same L3 cache region that the MPI ranks are in. So MPI 000 is mapped to GPU 4, MPI 001 is mapped to GPU 5, etc.

While this level of control over mapping MPI ranks to GPUs might be useful for some applications, it is always important to consider the implication of the mapping. For example, if the order of the GPU IDs in the map_gpu option is reversed, the MPI ranks and the GPUs they are mapped to would be in different L3 cache regions, which could potentially lead to poorer performance.

Example 4: 16 MPI ranks - each with 2 OpenMP threads and 1 *specific* GPU (multi-node)

Extending Examples 2 and 3 to run on 2 nodes is also a straightforward exercise by changing the number of nodes to 2 (-N2) and the number of MPI ranks to 16 (-n16).

$ export OMP_NUM_THREADS=2
$ srun -N2 -n16 -c2 --gpus-per-node=8 --gpu-bind=map_gpu:4,5,2,3,6,7,0,1 ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 000 - OMP 001 - HWT 001 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 001 - OMP 001 - HWT 009 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 002 - OMP 001 - HWT 017 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 003 - OMP 001 - HWT 025 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 004 - OMP 001 - HWT 033 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 005 - OMP 001 - HWT 041 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 006 - OMP 001 - HWT 049 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 007 - OMP 001 - HWT 057 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 008 - OMP 000 - HWT 000 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 008 - OMP 001 - HWT 001 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 009 - OMP 000 - HWT 008 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 009 - OMP 001 - HWT 009 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 010 - OMP 000 - HWT 016 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 010 - OMP 001 - HWT 017 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 011 - OMP 000 - HWT 024 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 011 - OMP 001 - HWT 025 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 012 - OMP 000 - HWT 032 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 012 - OMP 001 - HWT 033 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 013 - OMP 000 - HWT 040 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 013 - OMP 001 - HWT 041 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 014 - OMP 000 - HWT 048 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 014 - OMP 001 - HWT 049 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 015 - OMP 000 - HWT 056 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 015 - OMP 001 - HWT 057 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

Mapping multiple MPI ranks to a single GPU

In the following examples, 2 MPI ranks will be mapped to 1 GPU. For the sake of brevity, OMP_NUM_THREADS will be set to 1, so -c1 will be used unless otherwise specified.

Note

On AMD’s MI250X, multi-process service (MPS) is not needed since multiple MPI ranks per GPU is supported natively.

Example 5: 16 MPI ranks - where 2 ranks share a GPU (round-robin, single-node)

This example launches 16 MPI ranks (-n16), each with 1 physical CPU core (-c1) to launch 1 OpenMP thread (OMP_NUM_THREADS=1) on. The MPI ranks will be assigned to GPUs in a round-robin fashion so that each of the 8 GPUs on the node are shared by 2 MPI ranks. To accomplish this GPU mapping, a new srun options will be used:

• --ntasks-per-gpu specifies the number of MPI ranks that will share access to a GPU.

$ export OMP_NUM_THREADS=1
$ srun -N1 -n16 -c1 --ntasks-per-gpu=2 --gpu-bind=closest ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 008 - OMP 000 - HWT 001 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 009 - OMP 000 - HWT 009 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 010 - OMP 000 - HWT 017 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 011 - OMP 000 - HWT 025 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 012 - OMP 000 - HWT 033 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 013 - OMP 000 - HWT 041 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 014 - OMP 000 - HWT 049 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 015 - OMP 000 - HWT 057 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

The output shows the round-robin (cyclic) distribution of MPI ranks to GPUs. In fact, it is a round-robin distribution of MPI ranks to L3 cache regions (the default distribution). The GPU mapping is a consequence of where the MPI ranks are distributed; --gpu-bind=closest simply maps the GPU in an L3 cache region to the MPI ranks in the same L3 region.

Example 6: 32 MPI ranks - where 2 ranks share a GPU (round-robin, multi-node)

This example is an extension of Example 5 to run on 2 nodes.

$ export OMP_NUM_THREADS=1
$ srun -N2 -n32 -c1 --ntasks-per-gpu=2 --gpu-bind=closest ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 002 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 003 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 004 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 005 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 006 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 007 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 008 - OMP 000 - HWT 004 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 009 - OMP 000 - HWT 012 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 010 - OMP 000 - HWT 020 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 011 - OMP 000 - HWT 028 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 012 - OMP 000 - HWT 036 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 013 - OMP 000 - HWT 044 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 014 - OMP 000 - HWT 052 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 015 - OMP 000 - HWT 060 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 016 - OMP 000 - HWT 000 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 017 - OMP 000 - HWT 008 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 018 - OMP 000 - HWT 016 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 019 - OMP 000 - HWT 024 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 020 - OMP 000 - HWT 034 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 021 - OMP 000 - HWT 041 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 022 - OMP 000 - HWT 048 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 023 - OMP 000 - HWT 056 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 024 - OMP 000 - HWT 006 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 025 - OMP 000 - HWT 012 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 026 - OMP 000 - HWT 021 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 027 - OMP 000 - HWT 028 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 028 - OMP 000 - HWT 036 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 029 - OMP 000 - HWT 044 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 030 - OMP 000 - HWT 052 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 031 - OMP 000 - HWT 060 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

Example 7: 16 MPI ranks - where 2 ranks share a GPU (packed, single-node)

This example launches 16 MPI ranks (-n16), each with 4 physical CPU cores (-c4) to launch 1 OpenMP thread (OMP_NUM_THREADS=1) on. The MPI ranks will be assigned to GPUs in a packed fashion so that each of the 8 GPUs on the node are shared by 2 MPI ranks. Similar to Example 5, -ntasks-per-gpu=2 will be used, but a new srun flag will be used to change the default round-robin (cyclic) distribution of MPI ranks across NUMA domains:

• --distribution=<value>[:<value>][:<value>] specifies the distribution of MPI ranks across compute nodes, sockets (L3 cache regions on Frontier), and cores, respectively. The default values are block:cyclic:cyclic, which is where the cyclic assignment comes from in the previous examples.

Note

In the job step for this example, --distribution=*:block is used, where * represents the default value of block for the distribution of MPI ranks across compute nodes and the distribution of MPI ranks across L3 cache regions has been changed to block from its default value of cyclic.

Note

Because the distribution across L3 cache regions has been changed to a “packed” (block) configuration, caution must be taken to ensure MPI ranks end up in the L3 cache regions where the GPUs they intend to be mapped to are located. To accomplish this, the number of physical CPU cores assigned to an MPI rank was increased - in this case to 4. Doing so ensures that only 2 MPI ranks can fit into a single L3 cache region. If the value of -c was left at 1, all 8 MPI ranks would be “packed” into the first L3 region, where the “closest” GPU would be GPU 4 - the only GPU in that L3 region.

Notice that this is not a workaround like in Example 6, but a requirement due to the block distribution of MPI ranks across NUMA domains.

$ export OMP_NUM_THREADS=1
$ srun -N1 -n16 -c4 --ntasks-per-gpu=2 --gpu-bind=closest --distribution=*:block ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 004 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 002 - OMP 000 - HWT 008 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 003 - OMP 000 - HWT 012 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 004 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 005 - OMP 000 - HWT 020 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 006 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 007 - OMP 000 - HWT 028 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 008 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 009 - OMP 000 - HWT 036 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 010 - OMP 000 - HWT 040 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 011 - OMP 000 - HWT 044 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 012 - OMP 000 - HWT 048 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 013 - OMP 000 - HWT 052 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 014 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 015 - OMP 000 - HWT 060 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

The overall effect of using --distribution=*:block and increasing the number of physical CPU cores available to each MPI rank is to place the first two MPI ranks in the first L3 cache region with GPU 4, the next two MPI ranks in the second L3 cache region with GPU 5, and so on.

Example 8: 32 MPI ranks - where 2 ranks share a GPU (packed, multi-node)

This example is an extension of Example 7 to use 2 compute nodes. With the appropriate changes put in place in Example 7, it is a straightforward exercise to change to using 2 nodes (-N2) and 32 MPI ranks (-n32).

$ export OMP_NUM_THREADS=1
$ srun -N2 -n32 -c4 --ntasks-per-gpu=2 --gpu-bind=closest --distribution=*:block ./hello_jobstep | sort

MPI 000 - OMP 000 - HWT 000 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 001 - OMP 000 - HWT 004 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 002 - OMP 000 - HWT 010 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 003 - OMP 000 - HWT 012 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 004 - OMP 000 - HWT 016 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 005 - OMP 000 - HWT 021 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 006 - OMP 000 - HWT 024 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 007 - OMP 000 - HWT 028 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 008 - OMP 000 - HWT 032 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 009 - OMP 000 - HWT 037 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 010 - OMP 000 - HWT 041 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 011 - OMP 000 - HWT 044 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 012 - OMP 000 - HWT 049 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 013 - OMP 000 - HWT 052 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 014 - OMP 000 - HWT 056 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 015 - OMP 000 - HWT 060 - Node frontier002 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 016 - OMP 000 - HWT 000 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 017 - OMP 000 - HWT 004 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 4 - Bus_ID d1
MPI 018 - OMP 000 - HWT 008 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 019 - OMP 000 - HWT 013 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 5 - Bus_ID d6
MPI 020 - OMP 000 - HWT 016 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 021 - OMP 000 - HWT 020 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID c9
MPI 022 - OMP 000 - HWT 024 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 023 - OMP 000 - HWT 028 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID ce
MPI 024 - OMP 000 - HWT 034 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 025 - OMP 000 - HWT 036 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 6 - Bus_ID d9
MPI 026 - OMP 000 - HWT 040 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 027 - OMP 000 - HWT 044 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 7 - Bus_ID de
MPI 028 - OMP 000 - HWT 048 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 029 - OMP 000 - HWT 052 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c1
MPI 030 - OMP 000 - HWT 056 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6
MPI 031 - OMP 000 - HWT 060 - Node frontier004 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID c6

Example 9: 4 independent and simultaneous job steps in a single allocation

This example shows how to run multiple job steps simultaneously in a single allocation. The example below demonstrates running 4 independent, single rank MPI executions on a single node, however the example could be extrapolated to more complex invocations using the above examples.

Submission script:

#!/bin/bash
#SBATCH -A <projid>
#SBATCH -N 1
#SBATCH -t 10

srun -N1 -c1 --gpus-per-task=1 --exact ./hello_jobstep &
srun -N1 -c1 --gpus-per-task=1 --exact ./hello_jobstep &
srun -N1 -c1 --gpus-per-task=1 --exact ./hello_jobstep &
srun -N1 -c1 --gpus-per-task=1 --exact ./hello_jobstep &
wait

Output:

MPI 000 - OMP 000 - HWT 003 - Node frontier25 - RT_GPU_ID 0 - GPU_ID 3 - Bus_ID 09
MPI 000 - OMP 000 - HWT 001 - Node frontier25 - RT_GPU_ID 0 - GPU_ID 1 - Bus_ID 87
MPI 000 - OMP 000 - HWT 002 - Node frontier25 - RT_GPU_ID 0 - GPU_ID 2 - Bus_ID 48
MPI 000 - OMP 000 - HWT 000 - Node frontier25 - RT_GPU_ID 0 - GPU_ID 0 - Bus_ID c9

Note

The --exact parameter is important to avoid the error message srun: Job <job id> step creation temporarily disabled, retrying (Requested nodes are busy). The wait command is also critical, or your job script and allocation will immediately end after launching your jobs in the background.

Note

This may result in a sub-optimal alignment of CPU and GPU on the node, as shown in the example output. Unfortunately, at the moment there is not a workaround for this, however improvements are possible in future SLURM updates.

Multiple GPUs per MPI rank

As mentioned previously, all GPUs are accessible by all MPI ranks by default, so it is possible to programatically map any combination of GPUs to MPI ranks. It should be noted however that Cray MPICH does not support GPU-aware MPI for multiple GPUs per rank, so this binding is not suggested.

Software

Visualization and analysis tasks should be done on the Andes cluster. There are a few tools provided for various visualization tasks, as described in the Visualization tools section of the Andes User Guide.

For a full list of software availability and latest news at the OLCF, please reference the Software section in OLCF’s User Documentation.

Debugging

ARM DDT

Arm DDT is an advanced debugging tool used for scalar, multi-threaded, and large-scale parallel applications. In addition to traditional debugging features (setting breakpoints, stepping through code, examining variables), DDT also supports attaching to already-running processes and memory debugging. In-depth details of DDT can be found in the Official DDT User Guide, and instructions for how to use it on OLCF systems can be found on the Forge (DDT/MAP) Software Page. DDT is the OLCF’s recommended debugging software for large parallel applications.

One of the most useful features of DDT is its remote debugging feature. This allows you to connect to a debugging session on Frontier from a client running on your workstation. The local client provides much faster interaction than you would have if using the graphical client on Frontier. For guidance in setting up the remote client see this tutorial.

Profiling Applications

Getting Started with the HPE Performance Analysis Tools (PAT)

The Performance Analysis Tools (PAT), formerly CrayPAT, are a suite of utilities that enable users to capture and analyze performance data generated during program execution. These tools provide an integrated infrastructure for measurement, analysis, and visualization of computation, communication, I/O, and memory utilization to help users optimize programs for faster execution and more efficient computing resource usage.

There are three programming interfaces available: (1) Perftools-lite, (2) Perftools, and (3) Perftools-preload.

Below are two examples that generate an instrumented executable using Perftools, which is an advanced interface that provides full-featured data collection and analysis capability, including full traces with timeline displays.

The first example generates an instrumented executable using a PrgEnv-amd build:

module load PrgEnv-amd
module load craype-accel-amd-gfx90a
module load rocm
module load perftools

export PATH="${PATH}:${ROCM_PATH}/llvm/bin"
export CXX='CC -x hip'
export CXXFLAGS='-ggdb -O3 -std=c++17 –Wall'
export LD='CC'
export LDFLAGS="${CXXFLAGS} -L${ROCM_PATH}/lib"
export LIBS='-lamdhip64'

make clean
make

pat_build -g hip,io,mpi -w -f <executable>

The second example generates an instrumened executable using a hipcc build:

module load perftools
module load craype-accel-amd-gfx90a
module load rocm

export CXX='hipcc'
export CXXFLAGS="$(pat_opts include hipcc) \
  $(pat_opts pre_compile hipcc) -g -O3 -std=c++17 -Wall \
  --offload-arch=gfx90a -I${CRAY_MPICH_DIR}/include \
  $(pat_opts post_compile hipcc)"
export LD='hipcc'
export LDFLAGS="$(pat_opts pre_link hipcc) ${CXXFLAGS} \
  -L${CRAY_MPICH_DIR}/lib ${PE_MPICH_GTL_DIR_amd_gfx908}"
export LIBS="-lmpi ${PE_MPICH_GTL_LIBS_amd_gfx908} \
  $(pat_opts post_link hipcc)"

make clean
make

pat_build -g hip,io,mpi -w -f <executable>

The pat_build command in the above examples generates an instrumented executable with +pat appended to the executable name (e.g., hello_jobstep+pat).

When run, the instrumented executable will trace HIP, I/O, MPI, and all user functions and generate a folder of results (e.g., hello_jobstep+pat+39545-2t).

To analyze these results, use the pat_report command, e.g.:

pat_report hello_jobstep+pat+39545-2t

The resulting report includes profiles of functions, profiles of maximum function times, details on load imbalance, details on program energy and power usages, details on memory high water mark, and more.

More detailed information on the HPE Performance Analysis Tools can be found in the HPE Performance Analysis Tools User Guide.

Note

When using perftools-lite-gpu, there is a known issue causing ld.lld not to be found. A workaround this issue can be found here.

Getting Started with HPCToolkit

HPCToolkit is an integrated suite of tools for measurement and analysis of program performance on computers ranging from multicore desktop systems to the nation’s largest supercomputers. HPCToolkit provides accurate measurements of a program’s work, resource consumption, and inefficiency, correlates these metrics with the program’s source code, works with multilingual, fully optimized binaries, has very low measurement overhead, and scales to large parallel systems. HPCToolkit’s measurements provide support for analyzing a program execution cost, inefficiency, and scaling characteristics both within and across nodes of a parallel system.

Programming models supported by HPCToolkit include MPI, OpenMP, OpenACC, CUDA, OpenCL, DPC++, HIP, RAJA, Kokkos, and others.

Below is an example that generates a profile and loads the results in their GUI-based viewer.

module use /gpfs/alpine/csc322/world-shared/modulefiles/x86_64
module load hpctoolkit

# 1. Profile and trace an application using CPU time and GPU performance counters
srun <srun_options> hpcrun -o <measurement_dir> -t -e CPUTIME -e gpu=amd <application>

# 2. Analyze the binary of executables and its dependent libraries
hpcstruct <measurement_dir>

# 3. Combine measurements with program structure information and generate a database
hpcprof -o <database_dir> <measurement_dir>

# 4. Understand performance issues by analyzing profiles and traces with the GUI
hpcviewer <database_dir>

More detailed information on HPCToolkit can be found in the HPCToolkit User’s Manual.

Note

HPCToolkit does not require a recompile to profile the code. It is recommended to use the -g optimization flag for attribution to source lines.

Getting Started with the ROCm Profiler

rocprof gathers metrics on kernels run on AMD GPU architectures. The profiler works for HIP kernels, as well as offloaded kernels from OpenMP target offloading, OpenCL, and abstraction layers such as Kokkos. For a simple view of kernels being run, rocprof --stats --timestamp on is a great place to start. With the --stats option enabled, rocprof will generate a file that is named results.stats.csv by default, but named <output>.stats.csv if the -o flag is supplied. This file will list all kernels being run, the number of times they are run, the total duration and the average duration (in nanoseconds) of the kernel, and the GPU usage percentage. More detailed infromation on rocprof profiling modes can be found at ROCm Profiler documentation.

Roofline Profiling with the ROCm Profiler

The Roofline performance model is an increasingly popular way to demonstrate and understand application performance. This section documents how to construct a simple roofline model for a single kernel using rocprof. This roofline model is designed to be comparable to rooflines constructed by NVIDIA’s NSight Compute. A roofline model plots the achieved performance (in floating-point operations per second, FLOPS/s) as a function of arithmetic (or operational) intensity (in FLOPS per Byte). The model detailed here calculates the bytes moved as they move to and from the GPU’s HBM.

Note

Integer instructions and cache levels are currently not documented here.

To get started, you will need to make an input file for rocprof, to be passed in through rocprof -i <input_file> --timestamp on -o my_output.csv <my_exe>. Below is an example, and contains the information needed to roofline profile GPU 0, as seen by each rank:

pmc : SQ_INSTS_VALU_ADD_F16 SQ_INSTS_VALU_MUL_F16 SQ_INSTS_VALU_FMA_F16 SQ_INSTS_VALU_TRANS_F16
pmc : SQ_INSTS_VALU_ADD_F32 SQ_INSTS_VALU_MUL_F32 SQ_INSTS_VALU_FMA_F32 SQ_INSTS_VALU_TRANS_F32
pmc : SQ_INSTS_VALU_ADD_F64 SQ_INSTS_VALU_MUL_F64 SQ_INSTS_VALU_FMA_F64 SQ_INSTS_VALU_TRANS_F64
pmc : SQ_INSTS_VALU_MFMA_MOPS_F16 SQ_INSTS_VALU_MFMA_MOPS_BF16 SQ_INSTS_VALU_MFMA_MOPS_F32 SQ_INSTS_VALU_MFMA_MOPS_F64
pmc : TCC_EA_RDREQ_32B_sum TCC_EA_RDREQ_sum TCC_EA_WRREQ_sum TCC_EA_WRREQ_64B_sum
gpu: 0

Note

In an application with more than one kernel, you should strongly consider filtering by kernel name by adding a line like: kernel: <kernel_name> to the rocprof input file.

This provides the minimum set of metrics used to construct a roofline model. To gather metrics across multiple MPI ranks, you will need to use a command that redirects the output of rocprof to a unique file for each task. For example:

srun -N 2 -n 16 --ntasks-per-node=8 --gpus-per-node=8 --gpu-bind=closest bash -c 'rocprof -o ${SLURM_JOBID}_${SLURM_PROCID}.csv -i <input_file> --timestamp on <exe>'

Note

The gpu: filter in the rocprof input file identifies GPUs by the number the MPI rank would see them as. In the srun example above, each MPI rank only has 1 GPU, so each rank sees its GPU as GPU 0.

Theoretical Roofline

The theoretical (not attainable) peak roofline constructs a theoretical maximum performance for each operational intensity.

Note

theoretical peak is determined by the hardware specifications and is not attainable in practice. attaiable peak is the performance as measured by in-situ microbenchmarks designed to best utilize the hardware. achieved performance is what the profiled application actually achieves.

The theoretical roofline can be constructed as:

\[FLOPS_{peak} = minimum(ArithmeticIntensity * BW_{HBM}, TheoreticalFLOPS)\]

On Frontier, the memory bandwidth for HBM is 1.6 TB/s, and the theoretical peak floating-point FLOPS/s when using vector registers is calculated by:

\[TheoreticalFLOPS = 128 FLOP/cycle/CU * 110 CU * 1700000000 cycles/second = 23.9 TFLOP/s\]

However, when using MFMA instructions, the theoretical peak floating-point FLOPS/s is calculated by:

\[TheoreticalFLOPS = flop\_per\_cycle(precision) FLOP/cycle/CU * 110 CU * 1700000000 cycles/second\]

where flop_per_cycle(precision) is the published floating-point operations per clock cycle, per compute unit. Those values are:

Data Type	Flops/Clock/CU
FP64	256
FP32	256
FP16	1024
BF16	1024
INT8	1024

Note

Attainable peak rooflines are constructed using microbenchmarks, and are not currently discussed here. Attainable rooflines consider the limitations of cooling and power consumption and are more representative of what an application can achieve.

Achieved FLOPS/s

We calculate the achieved performance at the desired level (here, double-precision floating point, FP64), by summing each metric count and weighting the FMA metric by 2, since a fused multiply-add is considered 2 floating point operations. Also note that these SQ_INSTS_VALU_<ADD,MUL,TRANS> metrics are reported as per-simd, so we mutliply by the wavefront size as well. The SQ_INSTS_VALU_MFMA_MOPS_* instructions should be multiplied by the Flops/Cycle/CU value listed above. We use this equation to calculate the number of double-precision FLOPS:

\[\begin{split}FP64\_FLOPS = 64 *&(SQ\_INSTS\_VALU\_ADD\_F64 \\\\ &+ SQ\_INSTS\_VALU\_MUL\_F64 \\\\ &+ SQ\_INSTS\_VALU\_TRANS\_F64 \\\\ &+ 2 * SQ\_INSTS\_VALU\_FMA\_F64) \\\\ + 256 *&(SQ\_INSTS\_VALU\_MFMA\_MOPS\_F64)\end{split}\]

When SQ_INSTS_VALU_MFMA_MOPS_*_F64 instructions are used, then 47.8 TF/s is considered the theoretical maximum FLOPS/s. If only SQ_INSTS_VALU_<ADD,MUL,TRANS> are found, then 23.9 TF/s is the theoretical maximum FLOPS/s. Then, we divide the number of FLOPS by the elapsed time of the kernel to find FLOPS per second. This is found from subtracting the rocprof metrics EndNs by BeginNs, provided by --timestamp on, then converting from nanoseconds to seconds by dividing by 1,000,000,000 (power(10,9)).

Note

For ROCm/5.2.0 and earlier, there is a known issue with the timings provided by --timestamp on. See Known Issues.

Calculating for all precisions

The above formula can be adapted to compute the total FLOPS across all floating-point precisions (INT excluded).

\[\begin{split}TOTAL\_FLOPS = 64 *&(SQ\_INSTS\_VALU\_ADD\_F16 \\\\ &+ SQ\_INSTS\_VALU\_MUL\_F16 \\\\ &+ SQ\_INSTS\_VALU\_TRANS\_F16 \\\\ &+ 2 * SQ\_INSTS\_VALU\_FMA\_F16) \\\\ + 64 *&(SQ\_INSTS\_VALU\_ADD\_F32 \\\\ &+ SQ\_INSTS\_VALU\_MUL\_F32 \\\\ &+ SQ\_INSTS\_VALU\_TRANS\_F32 \\\\ &+ 2 * SQ\_INSTS\_VALU\_FMA\_F32) \\\\ + 64 *&(SQ\_INSTS\_VALU\_ADD\_F64 \\\\ &+ SQ\_INSTS\_VALU\_MUL\_F64 \\\\ &+ SQ\_INSTS\_VALU\_TRANS\_F64 \\\\ &+ 2 * SQ\_INSTS\_VALU\_FMA\_F64) \\\\ + 1024 &*(SQ\_INSTS\_VALU\_MFMA\_MOPS\_F16) \\\\ + 1024 &*(SQ\_INSTS\_VALU\_MFMA\_MOPS\_BF16) \\\\ + 256 *&(SQ\_INSTS\_VALU\_MFMA\_MOPS\_F32) \\\\ + 256 *&(SQ\_INSTS\_VALU\_MFMA\_MOPS\_F64) \\\\\end{split}\]

Arithmetic Intensity

Arithmetic intensity calculates the ratio of FLOPS to bytes moved between HBM and L2 cache. We calculated FLOPS above (FP64_FLOPS). We can calculate the number of bytes moved using the rocprof metrics TCC_EA_WRREQ_64B, TCC_EA_WRREQ_sum, TCC_EA_RDREQ_32B, and TCC_EA_RDREQ_sum. TCC refers to the L2 cache, and EA is the interface between L2 and HBM. WRREQ and RDREQ are write-requests and read-requests, respectively. Each of these requests is either 32 bytes or 64 bytes. So we calculate the number of bytes traveling over the EA interface as:

\[BytesMoved = BytesWritten + BytesRead\]

where

\[BytesWritten = 64 * TCC\_EA\_WRREQ\_64B\_sum + 32 * (TCC\_EA\_WRREQ\_sum - TCC\_EA\_WRREQ\_64B\_sum)\]

\[BytesRead = 32 * TCC\_EA\_RDREQ\_32B\_sum + 64 * (TCC\_EA\_RDREQ\_sum - TCC\_EA\_RDREQ\_32B\_sum)\]

---

Known Issues

Open Issues

Open Issues w/Workaround

Running

OLCFDEV-1479: rocFFT default cache file is in NFS shared directory - scalability issue

This behavior is fixed in ROCm/5.4.0. It affects all previous versions of ROCm.

By default rocFFT uses a location in $HOME to store its cache file. This is highly unsuitable for parallel runs with many processes all processes to try access the same file in $HOME, which is NFS share in our systems. This triggers a segfault when running more than 512 processes on Frontier/Crusher.

Workaround:

The following line forces rocFFT to use in-memory cache file:

export ROCFFT_RTC_CACHE_PATH=/dev/null

OLCFDEV-1485: Slurm “Unspecified error” when tasks per node is greater than 32

When running greater than 32 tasks per node, the following error can occur:

$ srun --ntasks-per-node=63 -n 126 -N 2 --gpus-per-node=8 --gpu-bind=closest ./hello_jobstep
srun: error: Task launch for StepId=136281.21 failed on node crusher001: Unspecified error
srun: error: Task launch for StepId=136281.21 failed on node crusher002: Unspecified error
srun: error: Application launch failed: Unspecified error
srun: Job step aborted

Workaround

The first thing to try would be to change your distribution flag to -m *:fcyclic or -m *:block. If the number of tasks per node is less than 64, then try setting -c 1 instead. If it is equal to 64 per node, then you can try combining those settings and set -c 1 -m *:fcyclic or -c 1 -m *:block.

This is a new behavior in Slurm 22.05 which was installed on July 5, 2022.

Resolved Issues