Hi Philippe, > In a previous community call, Zhao asked us how his work will scale > in the heterogeneous context. > > My first idea is CPUs must belong to a cluster.
Thank you for considering this! At present, cluster is a arch-specific topology level used by ARM. So maybe we need call this abstraction as another name not "cluster"? I guess the cluster you mentioned is the cluster device used in TCG, right? I also tried to eliminate differences between cluster devices and the cluster level in CPU topology [1]. My previous proposal introduced a abstract topology device [2]. And all topology specific levels are derived from the underlying topology device, including CPU. I feel like this topology device abstraction seems close to your idea, am I understanding it correctly? ;-) > For machines without > explicit cluster, we could always create the first one. Then -smp > would become a sugar property of the first cluster. Next -smp could > also be sugar property of the next cluster. Could you please explain the above ideas more? It feels we need to split -smp for each cluster. But I'm not sure if sugar property means defining smp-like properties for each cluster. Or is there a command line example? ;-) [1]: https://lore.kernel.org/qemu-devel/20231130144203.2307629-23-zhao1....@linux.intel.com/ [2]: https://lore.kernel.org/qemu-devel/20231130144203.2307629-9-zhao1....@linux.intel.com/ Thanks, Zhao