Skip to content

CPU Cluster Romeo

Overview

The HPC system Romeo is a general purpose cluster based on AMD Rome CPUs. From 2019 till the end of 2023, it was available as partition romeo within Taurus. With the decommission of Taurus, Romeo has been re-engineered and is now a homogeneous, standalone cluster with own Slurm batch system and own login nodes. This maintenance also comprised:

  • change the ethernet access (new VLANs),
  • complete integration of Romeo into the Barnard InfiniBand network to get full bandwidth access to all new filesystems,
  • configure and deploy stand-alone Slurm batch system,
  • newly build software within separate software and module system.

Hardware Resources

The hardware specification is documented on the page HPC Resources.

Usage

There is a total of 128 physical cores in each node. SMT is also active, so in total, 256 logical cores are available per node.

Note

Multithreading is disabled per default in a job. To make use of it include the Slurm parameter --hint=multithread in your job script or command line, or set the environment variable SLURM_HINT=multithread before job submission.

Each node brings 512 GB of main memory, so you can request roughly 1972 MB per logical core (using --mem-per-cpu). Note that you will always get the memory for the logical core sibling too, even if you do not intend to use SMT.

Note

If you are running a job here with only ONE process (maybe multiple cores), please explicitly set the option -n 1!

Be aware that software built with Intel compilers and -x* optimization flags will not run on those AMD processors! That's why most older modules built with Intel toolchains are not available on partition romeo.

We provide the script ml_arch_avail that can be used to check if a certain module is available on rome architecture.

Example, running CP2K on Rome

First, check what CP2K modules are available in general: module spider CP2K or module avail CP2K.

You will see that there are several different CP2K versions avail, built with different toolchains. Now let's assume you have to decided you want to run CP2K version 6 at least, so to check if those modules are built for rome, use:

marie@login$ ml_arch_avail CP2K/6
CP2K/6.1-foss-2019a: haswell, rome
CP2K/6.1-foss-2019a-spglib: haswell, rome
CP2K/6.1-intel-2018a: sandy, haswell
CP2K/6.1-intel-2018a-spglib: haswell

There you will see that only the modules built with toolchain foss are available on architecture rome, not the ones built with intel. So you can load, e.g. ml CP2K/6.1-foss-2019a.

Then, when writing your batch script, you have to specify the partition romeo. Also, if e.g. you wanted to use an entire ROME node (no SMT) and fill it with MPI ranks, it could look like this:

#!/bin/bash
#SBATCH --partition=romeo
#SBATCH --ntasks-per-node=128
#SBATCH --nodes=1
#SBATCH --mem-per-cpu=1972

srun cp2k.popt input.inp

Using the Intel Toolchain on Rome

Currently, we have only newer toolchains starting at intel/2019b installed for the Rome nodes. Even though they have AMD CPUs, you can still use the Intel compilers on there and they don't even create bad-performing code. When using the Intel Math Kernel Library (MKL) up to version 2019, though, you should set the following environment variable to make sure that AVX2 is used:

export MKL_DEBUG_CPU_TYPE=5

Without it, the MKL does a CPUID check and disables AVX2/FMA on non-Intel CPUs, leading to much worse performance.

Note

In version 2020 and above, Intel has removed this environment variable and added separate Zen codepaths to the library. However, they are still incomplete and do not cover every BLAS function. Also, the Intel AVX2 codepaths still seem to provide somewhat better performance, so a new workaround would be to overwrite the mkl_serv_intel_cpu_true symbol with a custom function:

int mkl_serv_intel_cpu_true() {
    return 1;
}

and preloading this in a library:

marie@login$ gcc -shared -fPIC -o libfakeintel.so fakeintel.c
marie@login$ export LD_PRELOAD=libfakeintel.so

As for compiler optimization flags, -xHOST does not seem to produce best-performing code in every case on Rome. You might want to try -mavx2 -fma instead.

Intel MPI

We have seen only half the theoretical peak bandwidth via InfiniBand between two nodes, whereas Open MPI got close to the peak bandwidth, so you might want to avoid using Intel MPI on partition rome if your application heavily relies on MPI communication until this issue is resolved.