Werner Feldmann

Compute Nodes

This page lists hardware specifications for the compute nodes currently used in the MOGON clusters, as well as our partitioning.

Available Resources#

Partitioning#

Individual compute nodes are grouped together into larger subsets of the cluster to form so-called partitions. A partition groups nodes based on characteristics or policies to ensure fairness and responsiveness.

Nodes: CPU-Nodes

Partitions supporting Accelerators#

Hidden Partitions#

Information on hidden partitions can be viewed by anyone. These partitions are set to be hidden to avoid cluttering the output for every poll - these partitions are “private” to certain projects/groups and only of interest to respective groups.

To visualize all jobs for a user in all partitions supply the -a flag:

squeue -u $USER -a

Likewise sinfo can be supplemented with -a to gather informations. All other commands work without this flag as expected.

Slurm Query#

The tables on this page list key attributes of MOGON’s compute nodes grouped by partition. For a more comprehensive listing, you can also query this information with the following Slurm command:

scontrol show partition <name-of-partition> --clusters=<cluster>

For example:

scontrol show partition ki-parallel --clusters=mogonki

scontrol show partition parallel --clusters=mogonnhr

Slurm will display defaults as well as minimal and maximal settings for reservation time, memory capacity, etc. of the partition.

Memory Limits#

You can also list all of our partitions with relevant limits using the sinfo command:

sinfo -e -o "%20P %16F %8z %.8m %.11l %G" -S "+P+m" --clusters=all

What are Nodes, Sockets, Cores and Threads?#

To explain the performance and structure the available Mogon clusters, their performance is indicated by stating facts about the number of machines (nodes) within the cluster and their specifications - sockets (S), physical cores (C) and threads/logical cores (T). The following illustrations give a rough overview of what these specifications mean:

While this already explains how the MOGON cluster manages to work on very large data sets and perform multiple tasks simultaneously, there are still some optimizations to be had. For this, we need to look more closely at how individual cores conduct their work. Looking at it from a simplified angle, every core starts working on a task, finishes this task, and then moves on to the next one. However, this process does not always work in such an idealized way. In many cases, the core will need to wait for other work to be finished before it can start calculating – like when data needs to be loaded or processed by another component on the node before the core can continue to calculate with it. Because of this, the core cannot be used as efficiently as theoretically possible.

To solve this problem, the task that needs to wait for additional input is usually simply paused, and another task is performed in the time that would usually be taken up by waiting. However, since this approach involves moving all the data for the relevant task in and out of the processors registers, it has a noticeable performance overhead.

To improve on this, each physical core of the CPU can be split into two logical cores (threads), which significantly reduces the amount of times this operation has to be performed. In this case, the operating system and software on the node only see the logical cores, not the physical ones. Despite this, the logical cores are no actual component within the processor – rather they can be imagined as virtual implementations of the workings of the traditional physical components.

The process of splitting physical cores into multiple threads is often called hyperthreading.