Fig 1.
Simplified organisation of a typical HPC system.
Users log in remotely via SSH to a login node from a client machine. Jobs are scheduled by job scheduler software running on the master node and run on compute nodes and GPU nodes. Gateway node(s) facilitate the transfer of large files into and out of the HPC system. GPU, graphics processing unit; HPC, high-performance computing; SSH, (Secure Shell Protocol).
Fig 2.
Parallelism in a typical HPC cluster.
The job scheduler runs on the master node and controls the highest level of parallelism in running jobs on each node (blue arrows). For example, in a variant calling pipeline, each node may perform the calling for an individual chromosome and could use multiple threads per core—3 threads per core are depicted, for a total of 12 per chromosomal job. In MPI jobs, threads on one node can communicate with threads on other nodes through inter-process communication (depicted in red dashed lines). HPC, high-performance computing; MPI, Message Passing Interface.