By: Doug S (foo.delete@this.bar.bar), October 15, 2021 10:44 am
Room: Moderated Discussions
rwessel (rwessel.delete@this.yahoo.com) on October 15, 2021 4:49 am wrote:
> While you don't want an infinite range of hardware in your cluster, you do need to allow
> some variation, or you cease to be able to upgrade it as new hardware comes along. For example,
> IBM's clustering for Z allows "N-2" systems in a cluster. So you could have (today) z13,
> z14s and z15s existing in a cluster, but no earlier systems. When the z16s come out, you'll
> need to remove any remaining z13s from your cluster before adding the new box.
IBM mainframes are a wholly different market though, that's more comparable to Apple in that they control the entire hardware/software stack. Given how much they cost, they pretty much have to support different generations in the same cluster. The problems they face doing it, i.e. what to do about a feature that exists in z15 but not z13, are not much of an issue for them because they provide the clustering software.
For e.g. Intel where you have some servers in a VM cluster that support AVX512 and some that don't, for instance, the software either has to claim the vCPUs don't support AVX512 on hardware that does, or limit migration of code using that to the subset of servers that support AVX512. And since Intel doesn't write the hypervisors, third parties like Microsoft and VMware are left to choose their own path between those options.
> While you don't want an infinite range of hardware in your cluster, you do need to allow
> some variation, or you cease to be able to upgrade it as new hardware comes along. For example,
> IBM's clustering for Z allows "N-2" systems in a cluster. So you could have (today) z13,
> z14s and z15s existing in a cluster, but no earlier systems. When the z16s come out, you'll
> need to remove any remaining z13s from your cluster before adding the new box.
IBM mainframes are a wholly different market though, that's more comparable to Apple in that they control the entire hardware/software stack. Given how much they cost, they pretty much have to support different generations in the same cluster. The problems they face doing it, i.e. what to do about a feature that exists in z15 but not z13, are not much of an issue for them because they provide the clustering software.
For e.g. Intel where you have some servers in a VM cluster that support AVX512 and some that don't, for instance, the software either has to claim the vCPUs don't support AVX512 on hardware that does, or limit migration of code using that to the subset of servers that support AVX512. And since Intel doesn't write the hypervisors, third parties like Microsoft and VMware are left to choose their own path between those options.
