Compilers

C, C++ and Fortran are supported on the Ascend cluster. Intel, oneAPI, GNU Compiler Collectio (GCC) and AOCC are available. The Intel development tool chain is loaded by default. Compiler commands and recommended options for serial programs are listed in the table below. See also our compilation guide.

The Rome/Milan processors from AMD that make up Ascend support the Advanced Vector Extensions (AVX2) instruction set, but you must set the correct compiler flags to take advantage of it. AVX2 has the potential to speed up your code by a factor of 4 or more, depending on the compiler and options you would otherwise use. However, bear in mind that clock speeds decrease as the level of the instruction set increases. So, if your code does not benefit from vectorization it may be beneficial to use a lower instruction set.

In our experience, the Intel compiler usually does the best job of optimizing numerical codes and we recommend that you give it a try if you’ve been using another compiler.

With the Intel/oneAPI compilers, use -xHost and -O2 or higher. With GCC, use -march=native and -O3. 

This advice assumes that you are building and running your code on Ascend. The executables will not be portable.  Of course, any highly optimized builds, such as those employing the options above, should be thoroughly validated for correctness.

Parallel Programming

MPI

By default, OSC systems use the MVAPICH implementation of the Message Passing Interface (MPI), which is optimized for high-speed InfiniBand interconnects. MPI is a standardized library designed for parallel processing in distributed-memory environments. OSC also supports OpenMPI and Intel MPI. For more information on building MPI applications, please visit the MPI software page.

MPI programs are started with the srun command. For example,

#!/bin/bash
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=48

srun [ options ] mpi_prog

The srun command will normally spawn one MPI process per task requested in a Slurm batch job. Use the --ntasks-per-node=n option to change that behavior. For example,

#!/bin/bash
#SBATCH --nodes=2
#SBATCh --exclusive

# Use the maximum number of CPUs of two nodes
srun ./mpi_prog

# Run 8 processes per node
srun -n 16 --ntasks-per-node=8  ./mpi_prog

The table below shows some commonly used options. Use srun -help for more information.

OpenMP

The Intel, and GNU compilers understand the OpenMP set of directives, which support multithreaded programming. For more information on building OpenMP codes on OSC systems, please visit the OpenMP documentation.

An OpenMP program by default will use a number of threads equal to the number of CPUs requested in a Slurm batch job. To use a different number of threads, set the environment variable OMP_NUM_THREADS. For example,

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8

# Run 8 threads
./omp_prog

# Run 4 threads
export OMP_NUM_THREADS=4
./omp_prog

Interactive job only

Please use -c, --cpus-per-task=X to request an interactive job. Both result in an interactive job with X CPUs available but only the former option automatically assigns the correct number of threads to the OpenMP program.

Hybrid (MPI + OpenMP)

An example of running a job for hybrid code:

#!/bin/bash
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=80

# Run 4 MPI processes on each node and 40 OpenMP threads spawned from a MPI process
export OMP_NUM_THREADS=40
srun -n 8 -c 40 --ntasks-per-node=4 ./hybrid_prog

Tuning Parallel Program Performance: Process/Thread Placement

To get the maximum performance, it is important to make sure that processes/threads are located as close as possible to their data, and as close as possible to each other if they need to work on the same piece of data, with given the arrangement of node, sockets, and cores, with different access to RAM and caches. 

While cache and memory contention between threads/processes are an issue, it is best to use scatter distribution for code. 

Processes and threads are placed differently depending on the computing resources you requste and the compiler and MPI implementation used to compile your code. For the former, see the above examples to learn how to run a job on exclusive nodes. For the latter, this section summarizes the default behavior and how to modify placement.

OpenMP only

For all three compilers (Intel, GCC and oneAPI), purely threaded codes do not bind to particular CPU cores by default. In other words, it is possible that multiple threads are bound to the same CPU core. 

The following table describes how to modify the default placements for pure threaded code:

1. Threads in the same socket might be bound to the same CPU core.

MPI Only

For MPI-only codes, MVAPICH first binds as many processes as possible on one socket, then allocates the remaining processes on the second socket so that consecutive tasks are near each other. Intel MPI and OpenMPI alternately bind processes on socket 1, socket 2, socket 1, socket 2 etc, as cyclic distribution.

For process distribution across nodes, all MPIs first bind as many processes as possible on one node, then allocates the remaining processes on the second node. 

The following table describe how to modify the default placements on single node for MPI-only code with the command srun:

1. MVP_CPU_BINDING_POLICY will not work if MVP_ENABLE_AFFINITY=0 is set.

To distribute processes evenly across nodes, please set SLURM_DISTRIBUTION=cyclic.

Hybrid (MPI + OpenMP)

For hybrid codes, each MPI process is allocated a number of cores defined by OMP_NUM_THREADS, and the threads of each process are bound to those cores. All MPI processes, along with the threads bound to them, behave similarly to what was described in the previous sections.

The following table describe how to modify the default placements on a single node for Hybrid code with the command srun:

Summary

The above tables list the most commonly used settings for process/thread placement. Some compilers and Intel libraries may have additional options for process and thread placement beyond those mentioned on this page. For more information on a specific compiler/library, check the more detailed documentation for that library.

GPU Programming

244 NVIDIA A100 GPUs are available on Ascend. Please visit our GPU documentation.

Reference

• Slurm: Support for Multi-core/Multi-thread Architecture
• Thread Placement and Thread Affinity

Memory limit

It is strongly suggested to consider the memory use to the available per-core memory when users request OSC resources for their jobs.

Summary

It is recommended to let the default memory apply unless more control over memory is needed.
Note that if an entire node is requested, then the job is automatically granted the entire node's main memory. On the other hand, if a partial node is requested, then memory is granted based on the default memory per core.

See a more detailed explanation below.

Default memory limits

A job can request resources and allow the default memory to apply. If a job requires 300 GB for example:

#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=30

This requests 30 cores, and each core will automatically be allocated 10.4 GB of memory (30 core * 10 GB memory = 300 GB memory).

Explicit memory requests

If needed, an explicit memory request can be added:

#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=300G

CPU only jobs

Dense gpu nodes on Ascend have 88 cores each. However, cpuonly partition jobs may only request 84 cores per node.

An example request would look like:

#!/bin/bash

#SBATCH --partition=cpuonly
#SBATCH --time=5:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=42

# requests 2 tasks * 42 cores each = 84 cores
<snip>

GPU Jobs

Jobs may request only parts of gpu node. These jobs may request up to the total cores on the node (88 cores).

Requests two gpus for one task:

#SBATCH --time=5:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=20
#SBATCH --gpus-per-task=2

Requests two gpus, one for each task:

#SBATCH --time=5:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=10
#SBATCH --gpus-per-task=1

Of course, jobs can request all the gpus of a dense gpu node as well. These jobs have access to all cores as well.

Request an entire dense gpu node:

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=88
#SBATCH --gpus-per-node=4

Partition time and job size limits

Here is the walltime and node limits per job for different queues/partitions available on Ascend:

Usually, you do not need to specify the partition for a job and the scheduler will assign the right partition based on the requested resources. To specify a partition for a job, either add the flag --partition=<partition-name> to the sbatch command at submission time or add this line to the job script:
#SBATCH --paritition=<partition-name>

Job/Core Limits

An individual user can have up to the max concurrently running jobs and/or up to the max processors/cores in use. However, among all the users in a particular group/project, they can have up to the max concurrently running jobs and/or up to the max processors/cores in use.

This page includes a summary of differences to keep in mind when migrating jobs from other clusters to Ascend. 

Guidance for Pitzer Users

Hardware Specifications

Guidance for Owens Users

Hardware Specifications

File Systems

Ascend accesses the same OSC mass storage environment as our other clusters. Therefore, users have the same home directory, project space, and scratch space as on the other clusters.

Software Environment

Ascend uses the same module system as other OSC Clusters.

Use   module load <package>   to add a software package to your environment. Use   module list   to see what modules are currently loaded and  module avail   to see the modules that are available to load. To search for modules that may not be visible due to dependencies or conflicts, use   module spider  . 

You can keep up to on the software packages that have been made available on Ascend by viewing the Software by System page and selecting the Ascend system.

Programming Environment

C, C++ and Fortran are supported on the Ascend cluster. Intel, oneAPI, GNU, nvhpc, and aocc compiler suites are available. The Intel development tool chain is loaded by default. To switch to a different compiler, use  module swap . Ascend also uses the MVAPICH2 implementation of the Message Passing Interface (MPI).

See the Ascend Programming Environment page for details. 

These are the public key fingerprints for Ascend:
ascend: ssh_host_rsa_key.pub = 2f:ad:ee:99:5a:f4:7f:0d:58:8f:d1:70:9d:e4:f4:16
ascend: ssh_host_ed25519_key.pub = 6b:0e:f1:fb:10:da:8c:0b:36:12:04:57:2b:2c:2b:4d
ascend: ssh_host_ecdsa_key.pub = f4:6f:b5:d2:fa:96:02:73:9a:40:5e:cf:ad:6d:19:e5

These are the SHA256 hashes:​
ascend: ssh_host_rsa_key.pub = SHA256:4l25PJOI9sDUaz9NjUJ9z/GIiw0QV/h86DOoudzk4oQ
ascend: ssh_host_ed25519_key.pub = SHA256:pvz/XrtS+PPv4nsn6G10Nfc7yM7CtWoTnkgQwz+WmNY
ascend: ssh_host_ecdsa_key.pub = SHA256:giMUelxDSD8BTWwyECO10SCohi3ahLPBtkL2qJ3l080

For more information about citations of OSC, visit https://www.osc.edu/citation.

To cite Ascend, please use the following Archival Resource Key:

ark:/19495/hpc3ww9d

Please adjust this citation to fit the citation style guidelines required.

Ohio Supercomputer Center. 2022. Ascend Supercomputer. Columbus, OH: Ohio Supercomputer Center. http://osc.edu/ark:/19495/hpc3ww9d

Here is the citation in BibTeX format:

@misc{Ascend2022,
ark = {ark:/19495/hpc3ww9d},
url = {http://osc.edu/ark:/19495/hpc3ww9d},
year  = {2022},
author = {Ohio Supercomputer Center},
title = {Ascend Supercomputer}
}

And in EndNote format:

%0 Generic
%T Ascend Supercomputer
%A Ohio Supercomputer Center
%R ark:/19495/hpc3ww9d
%U http://osc.edu/ark:/19495/hpc3ww9d
%D 2022

Users who would like to use the Ascend cluster will need to request access.  This is because of the particulars of the Ascend environment, which includes its size, GPUs, and scheduling policies.

Motivation

Access to Ascend is done on a case by case basis because:

• All nodes on Ascend are with 4 GPUs, and therefore it favors GPU work instead of CPU-only work
• It is a smaller machine than Pitzer and Owens, and thus has limited space for users

Good Ascend Workload Characteristics

Those interested in using Ascend should check that their work is well suited for it by using the following list.  Ideal workloads will exhibit one or more of the following characteristics:

• Needs access to Ascend specific hardware (GPUs, or AMD)
• Software:
  • Supports GPUs
  • Takes advantage of:
    • Long vector length
    • Higher core count
    • Improved memory bandwidth

Applying for Access

PIs of groups that would like to be considered for Ascend access should send the following in a email to OSC Help:

• Name
• Username
• Project code (group)
• Software/packages used on Ascend
• Evidence of workload being well suited for Ascend

The following are technical specifications for Ascend.  

Number of Nodes

24 nodes

Number of CPU Sockets

48 (2 sockets/node)

Number of CPU Cores

2,304 (96 cores/node)

Cores Per Node

96 cores/node (88 usable cores/node)

Internal Storage

12.8 TB NVMe internal storage

Compute CPU Specifications

AMD EPYC 7643 (Milan) processors for compute

• 2.3 GHz
• 48 cores per processor

Computer Server Specifications

24 Dell XE8545 servers

Accelerator Specifications

4 NVIDIA A100 GPUs with 80GB memory each, supercharged by NVIDIA NVLink

Number of Accelerator Nodes

24 total

Total Memory

~ 24 TB

Physical Memory Per Node

1 TB

Physical Memory Per Core

10.6 GB

Interconnect

Mellanox/NVIDA 200 Gbps HDR InfiniBand​