Getting access to old architectures is going to get harder and harder. Already, there are only a very small number AWS regions that support all our target architectures (if we stick to EU region only we have to use a minimum of two). It seems like it could be a good approach to consider supporting a Qemu/KVM mode in our build scripts so that we support older CPUs on newer architectures.
This doesn't necessarily have to come with a performance penalty, if we group the CPUs into familes there doesn't need to be any translation between instructions (Intel Sapphaire Rapids can run all Intel CPUs down to generic x86, Zen5 can do all Zen families, Arm is a bit more complicated but the model still mostly works). For build clusters that we control this should be feasible, but of course where we run directly connected to an HPC system you won't have this option.
This approach should give us much better resilience, and make leveraging EOSC resources (for example) more attractive for us.
Getting access to old architectures is going to get harder and harder. Already, there are only a very small number AWS regions that support all our target architectures (if we stick to EU region only we have to use a minimum of two). It seems like it could be a good approach to consider supporting a Qemu/KVM mode in our build scripts so that we support older CPUs on newer architectures.
This doesn't necessarily have to come with a performance penalty, if we group the CPUs into familes there doesn't need to be any translation between instructions (Intel Sapphaire Rapids can run all Intel CPUs down to generic x86, Zen5 can do all Zen families, Arm is a bit more complicated but the model still mostly works). For build clusters that we control this should be feasible, but of course where we run directly connected to an HPC system you won't have this option.
This approach should give us much better resilience, and make leveraging EOSC resources (for example) more attractive for us.