It seems that Spack automatically generates very optimized binaries (probably with something like gcc -march=native).
This is usually a good thing, but if a cluster consists of several generations of CPUs, this would cause a problem. In my case, the front node is newer than compute nodes, and the spack modules die on compute nodes.
To avoid this, you may want to add the following in ./etc/spack/defaults/packages.yaml.
Then the compiled binaries are for generic x86_64.
packages: all: target: [x86_64]
Appended on 30/July/2021
It should be better to have a user level configuration rather than changing the global file
(the global config file will be overwritten with git pull or you will be complained).
$ cat ~/.spack/packages.yaml
packages:
all:
target: [x86_64]
https://spack.readthedocs.io/en/latest/configuration.html
0 件のコメント:
コメントを投稿