Portability query

User discussions
1

GROMACS version: 2023.4
GROMACS modification: No
Here post your question:

I am adding a version of GROMACS to our cluster following instructions provided to me that were written when GROMACS was in version 2019.6.

The cluster consists of several nodes, each of which has identical hardware. Typically, we (IT support) build applications on a login node, whose hardware (CPU/GPU) is not the same as the cluster nodes.

I had determined that, based on the hardware of our cluster nodes, I should run the following cmake command.

cmake .. -DGMX_BUILD_OWN_FFTW=ON -DREGRESSIONTEST_DOWNLOAD=ON -DGMX_GPU=CUDA -DGMX_SIMD=AVX2_256

In the documentation for the version I am installing, under portability, I read “This selection will be done by the build system based on the capabilities of the build host machine or otherwise specified to cmake during configuration.”

I can’t quite get my head around this statement and am doubting that my build is going to take advantage of the cluster nodes’ hardware. I think it says “build gromacs, with no cmake options, on the hardware on which it will run, and it will be configured correctly for that hardware, taking advantage of everything the hardware has to offer.” and then says “or build gromacs anywhere and tell cmake about the hardware on which it will run”. So that is either/or and by feeding my cmake command with the options I’ve selected, the build will be fine for our nodes (providing, of course, I’ve selected appropriate options.)

I’d be grateful if someone would be able to set me straight on this point.

All the best,

Nic

2

Hello @_cowson ,

I am no IT person, but when I compile locally on my HPC home directory GROMACS in different versions/flavours I usually ssh locally to a node, module load the module I know I need (gcc, CUDA, cmake, etc) and then compile from the node, so that the hardware can be correctly detected and all the dependencies are correctly linked.

3

It is primarily (perhaps only) the SIMD settings that matter for this. Since you specify the explicitly the build should take advantage of the nodes’ hardware.

4

Hi @obZehn - do you pass any hardware specific options to cmake?

5

Hi @MagnusL,

When you say:

since you specify…

do you mean:

since you have specified -DGMX_SIMD=AVX2_256 in your cmake command, your build will take advantage of the target nodes hardware, despite being built on hardware that does not support -DGMX_SIMD=AVX2_256

?

Many thanks,

Nic

6

Yes, that’s the intended behavior. We do that on our cluster: build different SIMD flavors on a single node for use on different partitions.

EDIT: The inverse is also true: if you build on a modern node with AVX-512 but set -DGMX_SIMD=SSE2, then the resulting binary will only require SSE2 instructions.

7

Many thanks, understood.