[Neolith-users] New options when specifying job size

[Pär Andersson]

Hello Neolith test pilots,

Thanks to recent patches to both SLURM and mpprun it is now possible to use 
the option --ntasks-per-node in job specifications on Neolith. Especially 
users used to PBS (nodes=X:ppn=Y) might like this.

The following three batch scripts will all result in a job running on 32 
nodes, with 4 ranks per node, total 128 ranks:

---
#!/bin/bash
#SBATCH --nodes 32
#SBATCH --ntasks-per-node 4
mpprun my-mpi-binary
---

---
#!/bin/bash
#SBATCH --nodes 32
#SBATCH --tasks 128
mpprun my-mpi-binary
---

---
#!/bin/bash
#SBATCH --ntasks-per-node 4
#SBATCH --tasks 128
mpprun my-mpi-binary
---

If you don't specify --tasks or --ntasks-per-node the default is to run 8 MPI 
ranks on each node as each node have 8 cores. This means that the following 
specifications are also equivalent and will run on 2 nodes, 8 MPI ranks per 
node, 16 ranks total: 

---
#!/bin/bash
#SBATCH --nodes 2
mpprun my-mpi-binary
---

---
#!/bin/bash
#SBATCH --tasks 16
mpprun my-mpi-binary
---

---
#!/bin/bash
#SBATCH --nodes 2
#SBATCH --ntasks-per-node 8
mpprun my-mpi-binary
---

---
#!/bin/bash
#SBATCH --nodes 2
#SBATCH --tasks 16
mpprun my-mpi-binary
---

Best regards,

/Pär

-- 
Pär Andersson
National Supercomputer Centre
Sweden