Update 202405: Difference between revisions

Latest revision as of 17:02, 28 May 2024

Recent node purchases and the discontinuation of CentOS have prompted the need for major updates on the Magnolia HPC cluster.

The operating system of the Magnolia HPC cluster will be transitioned from CentOS 7 to Rocky Linux 9. This will require some custom packages to be recompiled, since the libraries of Rocky 9 are more recent than those of CentOS 7. These include those listed in Modules, and also from user installed packages included with miniconda.

Roadmap

Changes are required for the transition to Rocky Linux and the addition of new nodes; these include:

Topic	Completed?	Description
Install new compute node hardware - Part I	Yes	Vendor Installed 6 Compute nodes, each with: 2 Intel(R) Xeon(R) Gold 6448H Processors 512 MiB RAM 4 High memory compute nodes, each with: 2 Intel(R) Xeon(R) Gold 6426Y Processors 2 TiB RAM 3 GPU compute nodes, each with: 2 Intel(R) Xeon(R) Gold 6426Y Processors 256 GiB RAM 2 NVIDIA A100 80GB PCIe GPUs
Install new compute node hardware - Part II	No	Hardware not yet shipped by vendor.
Install operating system on new compute nodes	Yes	Using Rocky Linux 9 since CentOS 7 reaching end of life (EOL) on June 30, 2024. Took a lot of time getting the compute nodes with the new Rocky Linux operating system to integrate with the much older CentOS operating system nodes; Finally done.
Update backup and cloning system	Yes	Significant programming used to enable backup and cloning software to work with both Rocky Linux and CentOS.
Update Slurm	Yes	Updated to version that works on both Rocky Linux 9 and CentOS 7. Required very short downtime.
Add new compute nodes to Slurm partitions	Yes	Two new partitions created: kisame suliaoma These are currently only available to those who purchased the nodes, but will be available to all users after preemption (see below) is enabled. Four nodes were added to the himem node, which is available to all users.
Update module setup	No	WIP. See the New Modules Layout section for details on how to preview the new setup.
Update modules for Rocky Linux 9	No	WIP
Install CUDA drivers for new GPU compute nodes	Yes
Preemption partitions	No	TODO
Update CentOS 7 compute nodes to Rocky Linux 9	No	TODO
Update login nodes to Rocky Linux 9	No	TODO

Efforts are being made to make these changes as transparent to the general user as possible, however there will come a point where this will no longer be possible. Ample of notice will be give when it comes time for the general user to make changes in how they utilize the Magnolia HPC cluster.

Contacts

Upgrades may inadvertently cause some software packages to fail and miss checks by administraton; if you suspect such an event, please contact Brian Olson with details.

Slurm Partitions

As mentioned in the roadmap, two new partitions have been created. The standard slurm commands can be used to show current partition usage, or an overview of cores and RAM available on each partition can be shown with the coresavail command:

user $coresavail

Number of nodes in partition with N available cores and RAM.

Nodes  Partition  Available Cores  Available RAM (MiB)
6      kisame     64               515134
3      suliaoma   32               257094
4      himem      32               2063430
24     node       20               128000
4      himem      20               512000
2      gpu        20               128000
17     lomem      16               64000

This example output shows that the himem partition has 4 compute nodes each with 32 cores available and just under 2TiB RAM free. There are another 4 himem compute nodes free, but these each have 20 cores and 0.5 TiB RAM free.

The kisame and suliaoma partitions are currently only accessible to those who purchased the hardwaare (see roadmap for details).

Modules

During the upgrade, the cluster will have some compute nodes running on CentOS 7, and other on Rock Linux 9. Since many of the software packages installed as a Module on the cluster will not run on Rocky Linux, because of library version mismatches, a change in the layout of modules is required (see roadmap). This change will make it easier to run software on the appropriate operating system, and also allow for easier upgrades in the future.

New Modules Layout

The new module layout may be previewed by loading the newsetup module:

user $module load newsetup

user $module avail

------------------------ /usr/share/Modules/modulefiles ------------------------
dot         module-git  module-info modules     null        use.own

--------------------------- /modules/node/common/MPI ---------------------------
impi/2017.4.196

------------------------ /modules/node/common/Programs -------------------------
bamtools/2.5.1             ffmpeg/4.4
bedtools/2.31.0            gaussian/16B.01-avx2
bowtie/1.2.2               gaussian/16C.01-avx2
bowtie2/2.3.4.3            gaussian/16C.01-LINDA-avx2
bowtie2/2.5.1              nco/4.7.6
cmake/3.15.3               nco/4.9.3
cmake/3.24.2(default)      openssl/1.0.2k
cmake/3.9.1                openssl/3.0.9
fastx-toolkit/0.0.14       salmon/0.12.0
ffmpeg/3.3.3               salmon/1.1.0

------------------------ /modules/node/common/Libraries ------------------------
isl/0.22              mpfr/4.0.1            trimmomatic/0.39
lapack/3.7.1          ncurses/5.9           x264/20171213
libjpeg-turbo/2.1.5.1 newsetup/0.1          zstd/1.5.5
libpng/1.5.30         openssl/1.0.2k
mkl/2017.0.3          openssl/3.0.9

------------------------ /modules/node/common/Languages ------------------------
cuda-toolkit/10.1.243(default) gcc/6.4.0
cuda-toolkit/11.6.2            gcc/7.3.0
cuda-toolkit/8.0.61            gcc/8.3.0
gcc/11.4.0                     gcc/9.2.0
gcc/12.3.0                     intel/2017.4.196
gcc/13.2.0

----------------------------- /modules/node/Types ------------------------------
centos-7-63/0.1   centos-7-79/0.1   rocky-9.3-143/0.1 thisnode/0.1

This new module layout will show modules in the /modules/node/common path, which will run on both Rocky Linux 9 and CentOS 7. The last module path on the list is /modules/node/Types, which, when loaded, will reveal packages that will run on that particular node type; the naming scheme for these modules is distribution name-distribution version-CPU model.

The special thisnode module determines which type of node it is currently on and loads the appropriate node module. After the upgrade, this will most likely be in the Slurm submission script of all users. To show which type of node you are currently one, use the module help thisnode commands:

user $module help thisnode

----------- Module Specific Help for 'thisnode/0.1' ---------------

Adds node specific modulefiles directory to MODULEPATH.

	Node Type: centos-7-79

Rocky Linux Modules

To show packages that will run on Rocky Linux nodes, run the following commands:

user $module load rocky-9.3-143

user $module avail

------------------------ /usr/share/Modules/modulefiles ------------------------
dot         module-git  module-info modules     null        use.own

--------------------------- /modules/node/common/MPI ---------------------------
impi/2017.4.196

------------------------ /modules/node/common/Programs -------------------------
bamtools/2.5.1             ffmpeg/4.4
bedtools/2.31.0            gaussian/16B.01-avx2
bowtie/1.2.2               gaussian/16C.01-avx2
bowtie2/2.3.4.3            gaussian/16C.01-LINDA-avx2
bowtie2/2.5.1              nco/4.7.6
cmake/3.15.3               nco/4.9.3
cmake/3.24.2(default)      openssl/1.0.2k
cmake/3.9.1                openssl/3.0.9
fastx-toolkit/0.0.14       salmon/0.12.0
ffmpeg/3.3.3               salmon/1.1.0

------------------------ /modules/node/common/Libraries ------------------------
isl/0.22              mpfr/4.0.1            trimmomatic/0.39
lapack/3.7.1          ncurses/5.9           x264/20171213
libjpeg-turbo/2.1.5.1 newsetup/0.1          zstd/1.5.5
libpng/1.5.30         openssl/1.0.2k
mkl/2017.0.3          openssl/3.0.9

------------------------ /modules/node/common/Languages ------------------------
cuda-toolkit/10.1.243(default) gcc/6.4.0
cuda-toolkit/11.6.2            gcc/7.3.0
cuda-toolkit/8.0.61            gcc/8.3.0
gcc/11.4.0                     gcc/9.2.0
gcc/12.3.0                     intel/2017.4.196
gcc/13.2.0

----------------------------- /modules/node/Types ------------------------------
centos-7-63/0.1   centos-7-79/0.1   rocky-9.3-143/0.1 thisnode/0.1

----------------------- /modules/node/rocky-9.3-143/MPI ------------------------
openmpi-gcc/4.1.5

--------------------- /modules/node/rocky-9.3-143/Programs ---------------------
lammps/20220623

-------------------- /modules/node/rocky-9.3-143/Libraries ---------------------
hdf5/1.12.1    hdf5/1.8.19    netcdf/4.4.1.1 netcdf/4.8.1

Along with the modules previously listed, additional modules for packages that will run on Rocky Linux nodes are now shown. These can be loaded and used like regular modules, but they will fail to run on CentOS 7 nodes.

This list is current, as of May 28, 2024.

Installation Priority

The bulk of my time is currently spent installing modules for the nodes running Rocky Linux. I am prioritizing modules that have been used 10 or more times over the last 90 days:

All nodes over last 90 days:

Module         Count
hdf5           2089
netcdf         2079
intel          1381
mkl            1317
lammps         1294
mpfr           1225
gcc            1222
libpng         1109
isl            1057
x264           804
ffmpeg         792
esmf           749
MCT            744
scrip-coawst   742
fftw           737
nco            683
ncview         676
gaussian       645
python         433
cuda-toolkit   317
bedtools       258
samtools       258
openmpi        221
bowtie         203
R              165
orca           150
fastx-toolkit  146
digimat        109
icu            96
readline       93
ncurses        88
matlab         69
impi           56
cmake          54
openmpi-gcc    52
lapack         45
salmon         34
git            27
trimmomatic    20
newsetup       18
gmake          14
bowtie2        13
lammps-tools   13
mopac          13
gromacs        10
STAR           10

Every effort will be taken to install the same version of module packages as on CentOS 7, however some older versions will simply not compile against the newer libraries of Rocky Linux.

If you require a module not on this list, please make a request to those listed in the contacts section.

Using Rocky Linux Nodes

For users that wish to use the nodes running Rocky Linux, you should modify your slurm submission scripts so that they load the newsetup and thisnode modules. A minimal slurm batch script to run on the new suliaoma partition with 1 GPU is shown here:

FILE sample.sh

#!/bin/bash -l

#SBATCH --partition=suliaoma
#SBATCH --gres=gpu:1
#SBATCH --qos=bxmarg
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=5
#SBATCH --cpus-per-task=1
#SBATCH --mem=1000G
#SBATCH --time=00:10:00
#SBATCH --job-name=example
#SBATCH --output=myoutput-%j.txt

cd "$SLURM_SUBMIT_DIR"

module purge
module load newsetup
module load thisnode

srun hostname

@@ Line 1: / Line 1: @@
-Update to Rocky Linux 9
+Recent node purchases and the discontinuation of CentOS have prompted the need for major updates on the Magnolia HPC cluster.
+__TOC__
+The operating system of the Magnolia HPC cluster will be transitioned from CentOS 7 to [//www.rockylinux.org Rocky Linux 9]. This will require some custom packages to be recompiled, since the libraries of Rocky 9 are more recent than those of CentOS 7. These include those listed in [[#Modules|Modules]], and also from user installed packages included with [//docs.anaconda.com/free/miniconda/index.html miniconda].
+== Roadmap ==
+Changes are required for the transition to Rocky Linux and the addition of new nodes; these include:
+{| class="table table-striped"
+|-
+! Topic !! Completed? !! Description
+|-
+| Install new compute node hardware - Part I ||{{Yes}}
+| Vendor Installed
+* 6 Compute nodes, each with:
+** 2 Intel(R) Xeon(R) Gold 6448H Processors
+** 512 MiB RAM
+* 4 High memory compute nodes, each with:
+** 2 Intel(R) Xeon(R) Gold 6426Y Processors
+** 2 TiB RAM
+* 3 GPU compute nodes, each with:
+** 2 Intel(R) Xeon(R) Gold 6426Y Processors
+** 256 GiB RAM
+** 2 NVIDIA A100 80GB PCIe GPUs
+|-
+| Install new compute node hardware - Part II|| {{No}} || Hardware not yet shipped by vendor.
+|-
+| Install operating system on new compute nodes || {{Yes}}
+| Using [//www.rockylinux.org Rocky Linux 9] since CentOS 7 reaching [//www.redhat.com/en/topics/linux/centos-linux-eol end of life (EOL)] on June 30, 2024. Took a lot of time getting the compute nodes with the new Rocky Linux operating system to integrate with the much older CentOS operating system nodes; Finally done.
+|-
+| Update backup and cloning system || {{Yes}} || Significant programming used to enable backup and cloning software to work with both Rocky Linux and CentOS.
+|-
+| Update Slurm || {{Yes}}
+| Updated to version that works on both Rocky Linux 9 and CentOS 7. Required very short downtime.
+|-
+| Add new compute nodes to Slurm partitions || {{Yes}}
+| Two new partitions created:
+* kisame
+* suliaoma
+These are currently only available to those who purchased the nodes, but will be available to all users after preemption (see below) is enabled.
+Four nodes were added to the {{C|himem}} node, which is available to all users.
+|-
+| Update module setup || {{No}} || WIP. See the [[#New Modules Layout|New Modules Layout]] section for details on how to preview the new setup.
+|-
+| Update modules for Rocky Linux 9 || {{No}} || WIP
+|-
+| Install CUDA drivers for new GPU compute nodes || {{Yes}} ||
+|-
+| Preemption partitions || {{No}} || TODO
+|-
+| Update CentOS 7 compute nodes to Rocky Linux 9 || {{No}} || TODO
+|-
+| Update login nodes to Rocky Linux 9 || {{No}} || TODO
+|}
+Efforts are being made to make these changes as transparent to the general user as possible, however there will come a point where this will no longer be possible. Ample of notice will be give when it comes time for the general user to make changes in how they utilize the Magnolia HPC cluster.
+== Contacts ==
+Upgrades may inadvertently cause some software packages to fail and miss checks by administraton; if you suspect such an event, please contact [mailto:Brian.Olson@usm.edu?Subject=Magnolia%20HPC%20Cluster Brian Olson] with details.
+== Slurm Partitions ==
+As mentioned in the [[#roadmap|roadmap]], two new partitions have been created. The standard slurm commands can be used to show current partition usage, or an overview of cores and RAM available on each partition can be shown with the {{C|coresavail}} command:
+{{Cmd|coresavail|output=<pre>
+Number of nodes in partition with N available cores and RAM.
+Nodes  Partition  Available Cores  Available RAM (MiB)
+      kisame     64               515134
+      suliaoma   32               257094
+      himem      32               2063430
+     node       20               128000
+      himem      20               512000
+      gpu        20               128000
+     lomem      16               64000
+</pre>}}
+This example output shows that the {{C|himem}} partition has 4 compute nodes each with 32 cores available and just under 2TiB RAM free. There are another 4 himem compute nodes free, but these each have 20 cores and 0.5 TiB RAM free.
+The {{C|kisame}} and {{C|suliaoma}} partitions are currently only accessible to those who purchased the hardwaare (see [[#Roadmap|roadmap]] for details).
+== Modules ==
+During the upgrade, the cluster will have some compute nodes running on CentOS 7, and other on Rock Linux 9. Since many of the software packages installed as a Module on the cluster will not run on Rocky Linux, because of library version mismatches, a change in the layout of modules is required (see [[#Roadmap|roadmap]]). This change will make it easier to run software on the appropriate operating system, and also allow for easier upgrades in the future.
+=== New Modules Layout ===
+The new module layout may be previewed by loading the {{C|newsetup}} module:
+{{Cmd|module load newsetup|module avail|output=<pre>
+------------------------ /usr/share/Modules/modulefiles ------------------------
+dot         module-git  module-info modules     null        use.own
+--------------------------- /modules/node/common/MPI ---------------------------
+impi/2017.4.196
+------------------------ /modules/node/common/Programs -------------------------
+bamtools/2.5.1             ffmpeg/4.4
+bedtools/2.31.0            gaussian/16B.01-avx2
+bowtie/1.2.2               gaussian/16C.01-avx2
+bowtie2/2.3.4.3            gaussian/16C.01-LINDA-avx2
+bowtie2/2.5.1              nco/4.7.6
+cmake/3.15.3               nco/4.9.3
+cmake/3.24.2(default)      openssl/1.0.2k
+cmake/3.9.1                openssl/3.0.9
+fastx-toolkit/0.0.14       salmon/0.12.0
+ffmpeg/3.3.3               salmon/1.1.0
+------------------------ /modules/node/common/Libraries ------------------------
+isl/0.22              mpfr/4.0.1            trimmomatic/0.39
+lapack/3.7.1          ncurses/5.9           x264/20171213
+libjpeg-turbo/2.1.5.1 newsetup/0.1          zstd/1.5.5
+libpng/1.5.30         openssl/1.0.2k
+mkl/2017.0.3          openssl/3.0.9
+------------------------ /modules/node/common/Languages ------------------------
+cuda-toolkit/10.1.243(default) gcc/6.4.0
+cuda-toolkit/11.6.2            gcc/7.3.0
+cuda-toolkit/8.0.61            gcc/8.3.0
+gcc/11.4.0                     gcc/9.2.0
+gcc/12.3.0                     intel/2017.4.196
+gcc/13.2.0
+----------------------------- /modules/node/Types ------------------------------
+centos-7-63/0.1   centos-7-79/0.1   rocky-9.3-143/0.1 thisnode/0.1
+</pre>}}
+This new module layout will show modules in the {{Path|/modules/node/common}} path, which will run on both Rocky Linux 9 and CentOS 7. The last module path on the list is {{Path|/modules/node/Types}}, which, when loaded, will reveal packages that will run on that particular node type; the naming scheme for these modules is {{C|distribution name-distribution version-CPU model}}.
+The special {{C|thisnode}} module determines which type of node it is currently on and loads the appropriate node module. After the upgrade, this will most likely be in the Slurm submission script of all users. To show which type of node you are currently one, use the {{C|module help thisnode}} commands:
+{{Cmd|module help thisnode|output=<pre>
+----------- Module Specific Help for 'thisnode/0.1' ---------------
+Adds node specific modulefiles directory to MODULEPATH.
+	Node Type: centos-7-79
+</pre>}}
+=== Rocky Linux Modules ===
+To show packages that will run on Rocky Linux nodes, run the following commands:
+{{Cmd|module load rocky-9.3-143|module avail|output=<pre>
+------------------------ /usr/share/Modules/modulefiles ------------------------
+dot         module-git  module-info modules     null        use.own
+--------------------------- /modules/node/common/MPI ---------------------------
+impi/2017.4.196
+------------------------ /modules/node/common/Programs -------------------------
+bamtools/2.5.1             ffmpeg/4.4
+bedtools/2.31.0            gaussian/16B.01-avx2
+bowtie/1.2.2               gaussian/16C.01-avx2
+bowtie2/2.3.4.3            gaussian/16C.01-LINDA-avx2
+bowtie2/2.5.1              nco/4.7.6
+cmake/3.15.3               nco/4.9.3
+cmake/3.24.2(default)      openssl/1.0.2k
+cmake/3.9.1                openssl/3.0.9
+fastx-toolkit/0.0.14       salmon/0.12.0
+ffmpeg/3.3.3               salmon/1.1.0
+------------------------ /modules/node/common/Libraries ------------------------
+isl/0.22              mpfr/4.0.1            trimmomatic/0.39
+lapack/3.7.1          ncurses/5.9           x264/20171213
+libjpeg-turbo/2.1.5.1 newsetup/0.1          zstd/1.5.5
+libpng/1.5.30         openssl/1.0.2k
+mkl/2017.0.3          openssl/3.0.9
+------------------------ /modules/node/common/Languages ------------------------
+cuda-toolkit/10.1.243(default) gcc/6.4.0
+cuda-toolkit/11.6.2            gcc/7.3.0
+cuda-toolkit/8.0.61            gcc/8.3.0
+gcc/11.4.0                     gcc/9.2.0
+gcc/12.3.0                     intel/2017.4.196
+gcc/13.2.0
+----------------------------- /modules/node/Types ------------------------------
+centos-7-63/0.1   centos-7-79/0.1   rocky-9.3-143/0.1 thisnode/0.1
+----------------------- /modules/node/rocky-9.3-143/MPI ------------------------
+openmpi-gcc/4.1.5
+--------------------- /modules/node/rocky-9.3-143/Programs ---------------------
+lammps/20220623
+-------------------- /modules/node/rocky-9.3-143/Libraries ---------------------
+hdf5/1.12.1    hdf5/1.8.19    netcdf/4.4.1.1 netcdf/4.8.1
+</pre>}}
+Along with the modules previously listed, additional modules for packages that will run on Rocky Linux nodes are now shown. These can be loaded and used like regular modules, but they will fail to run on CentOS 7 nodes.
+This list is current, as of May 28, 2024.
+=== Installation Priority ===
+The bulk of my time is currently spent installing modules for the nodes running Rocky Linux. I am prioritizing modules that have been used 10 or more times over the last 90 days:
+{{GenericCmd|output=<pre>
+All nodes over last 90 days:
+Module         Count
+hdf5           2089
+netcdf         2079
+intel          1381
+mkl            1317
+lammps         1294
+mpfr           1225
+gcc            1222
+libpng         1109
+isl            1057
+x264           804
+ffmpeg         792
+esmf           749
+MCT            744
+scrip-coawst   742
+fftw           737
+nco            683
+ncview         676
+gaussian       645
+python         433
+cuda-toolkit   317
+bedtools       258
+samtools       258
+openmpi        221
+bowtie         203
+R              165
+orca           150
+fastx-toolkit  146
+digimat        109
+icu            96
+readline       93
+ncurses        88
+matlab         69
+impi           56
+cmake          54
+openmpi-gcc    52
+lapack         45
+salmon         34
+git            27
+trimmomatic    20
+newsetup       18
+gmake          14
+bowtie2        13
+lammps-tools   13
+mopac          13
+gromacs        10
+STAR           10
+</pre>}}
+Every effort will be taken to install the same version of module packages as on CentOS 7, however some older versions will simply not compile against the newer libraries of Rocky Linux.
+If you require a module not on this list, please make a request to those listed in the [[#Contacts|contacts]] section.
+== Using Rocky Linux Nodes ==
+For users that wish to use the nodes running Rocky Linux, you should modify your slurm submission scripts so that they load the {{C|newsetup}} and {{C|thisnode}} modules. A minimal slurm batch script to run on the new {{C|suliaoma}} partition with 1 GPU is shown here:
+{{FileBox|filename=sample.sh|1=
+#!/bin/bash -l
+#SBATCH --partition=suliaoma
+#SBATCH --gres=gpu:1
+#SBATCH --qos=bxmarg
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=5
+#SBATCH --cpus-per-task=1
+#SBATCH --mem=1000G
+#SBATCH --time=00:10:00
+#SBATCH --job-name=example
+#SBATCH --output=myoutput-%j.txt
+cd "$SLURM_SUBMIT_DIR"
+module purge
+module load newsetup
+module load thisnode
+srun hostname
+}}