I am currently investigating how to enhance performance isolation and predictability on the latest Intel platforms.
To this aim, I am envisioning to exploit the Cluster-on-die mechanism, but I have some questions.
The first one is about the models where it is available: on multiple internet sources I read that it is available only in models with 10 or more cores, since they arrange cores and LLC slices on two ring buses, each one with a System Agent. Can someone confirm this?
The second one is about its functioning. If I enable Cluster-on-die, LLC misses are handled by the SA on the same ring of the core that generated the miss, right?
Finally, the major question is about data placement in LLC with COD enabled: are data placed on LLC slices on the same ring of the core fetching them, or are data spread to all slices across both rings?
Thanks in advance for any information!