====== When Should I Use the Low Latency Nodes? ======

If you are using MPI jobs then you may immediately think that the [[https://hpc.researchcomputing.ncl.ac.uk/calc/show/2/8|low latency nodes]] are what you should be using. 

Instinctively this would appear to be true, however the truth is a little more nuanced.

If your MPI job fits entirely within the resource constraints of a single node, then you will get __better__ performance by restricting the Slurm job to a single compute node, rather than distributing it across multiple physical nodes. The speed of process to process messaging within the same node is far higher than having to go out from the node and send a message over the (admittedly very fast) Infiniband connection to another node.

If you are running MPI jobs which count in the tens or dozens of cores range, than you are better off running those jobs on the **default_paid** partition, on a single node, as the job can be run entirely within the resources of that single machine.

   * //Running jobs on the low latency nodes which fit within the bounds of a single node will __not__ gain you any performance and may/will use your project funds at a faster rate//

Only if your job starts needing to run __hundreds__ of cores, should you move to the **low-latency_paid** partition.

----

[[:faq:index|Back to FAQ index]]