Skip to content

New Tier-2 clusters: donphan and gallade#

In April 2023, two new clusters were added to the HPC-UGent Tier-2 infrastructure: donphan and gallade.

This page provides some important information regarding these clusters, and how they differ from the clusters they are replacing (slaking and kirlia, respectively).

If you have any questions on using donphan or gallade, you can contact the HPC-UGent team.

For software installation requests, please use the request form.


donphan: debug/interactive cluster#

donphan is the new debug/interactive cluster.

It replaces slaking, which will be retired on Monday 22 May 2023.

It is primarily intended for interactive use: interactive shell sessions, using GUI applications through the HPC-UGent web portal, etc.

This cluster consists of 12 workernodes, each with:

  • 2x 18-core Intel Xeon Gold 6240 (Cascade Lake @ 2.6 GHz) processor;
  • one shared NVIDIA Ampere A2 GPU (16GB GPU memory)
  • ~738 GiB of RAM memory;
  • 1.6TB NVME local disk;
  • HDR-100 InfiniBand interconnect;
  • RHEL8 as operating system;

To start using this cluster from a terminal session, first run:

module swap cluster/donphan

You can also start (interactive) sessions on donphan using the HPC-UGent web portal.

Differences compared to slaking#

CPUs#

The most important difference between donphan and slaking workernodes is in the CPUs: while slaking workernodes featured Intel Haswell CPUs, which support SSE*, AVX, and AVX2 vector instructions, donphan features Intel Cascade Lake CPUs, which also support AVX-512 instructions, on top of SSE*, AVX, and AVX2.

Although software that was built on a slaking workernode with compiler options that enable architecture-specific optimizations (like GCC's -march=native, or Intel compiler's -xHost) should still run on a donphan workernode, it is recommended to recompile the software to benefit from the support for AVX-512 vector instructions.

Cluster size#

The donphan cluster is significantly bigger than slaking, both in terms of number of workernodes and number of cores per workernode, and hence the potential performance impact of oversubscribed cores (see below) is less likely to occur in practice.

User limits and oversubscription on donphan#

By imposing strict user limits and using oversubscription on this cluster, we ensure that anyone can get a job running without having to wait in the queue, albeit with limited resources.

The user limits for donphan include: * max. 5 jobs in queue; * max. 3 jobs running; * max. of 8 cores in total for running jobs; * max. 27GB of memory in total for running jobs;

The job scheduler is configured with to allow oversubscription of the available cores, which means that jobs will continue to start even if all cores are already occupied by running jobs. While this prevents waiting time in the queue, it does imply that performance will degrade when all cores are occupied and additional jobs continue to start running.

Shared GPU on donphan workernodes#

Each donphan workernode includes a single NVIDIA A2 GPU that can be used for light compute workloads, and to accelerate certain graphical tasks.

This GPU is shared across all jobs running on the workernode, and does not need to be requested explicitly (it is always available, similar to the local disk of the workernode).

Warning

Due to the shared nature of this GPU, you should assume that any data that is loaded in the GPU memory could potentially be accessed by other users, even after your processes have completed.

There are no strong security guarantees regarding data protection when using this shared GPU!


gallade: large-memory cluster#

gallade is the new large-memory cluster.

It replaces kirlia, which will be retired on Monday 22 May 2023.

This cluster consists of 12 workernodes, each with:

  • 2x 64-core AMD EPYC 7773X (Milan-X @ 2.2 GHz) processor;
  • ~940 GiB of RAM memory;
  • 1.5TB NVME local disk;
  • HDR-100 InfiniBand interconnect;
  • RHEL8 as operating system;

To start using this cluster from a terminal session, first run:

module swap cluster/gallade

You can also start (interactive) sessions on gallade using the HPC-UGent web portal.

Differences compared to kirlia#

CPUs#

The most important difference between gallade and kirlia workernodes is in the CPUs: while kirlia workernodes featured Intel Cascade Lake CPUs, which support vector AVX-512 instructions (next to SSE*, AVX, and AVX2), gallade features AMD Milan-X CPUs, which implement the Zen3 microarchitecture and hence do not support AVX-512 instructions (but do support SSE*, AVX, and AVX2).

As a result, software that was built on a kirlia workernode with compiler options that enable architecture-specific optimizations (like GCC's -march=native, or Intel compiler's -xHost) may not work anymore on a gallade workernode, and will produce Illegal instruction errors.

Therefore, you may need to recompile software in order to use it on gallade. Even if software built on kirlia does still run on gallade, it is strongly recommended to recompile it anyway, since there may be signficant peformance benefits.

Memory per core#

Although gallade workernodes have signficantly more RAM memory (~940 GiB) than kirlia workernodes had (~738 GiB), the average amount of memory per core is significantly lower on gallade than it was on kirlia, because a gallade workernode has 128 cores (so ~7.3 GiB per core on average), while a kirlia workernode had only 36 cores (so ~20.5 GiB per core on average).

It is important to take this aspect into account when submitting jobs to gallade, especially when requesting all cores via ppn=all. You may need to explictly request more memory (see also here).