What does a global sharer bit name?

In the flat course model H names cores. This lesson groups cores into clusters. A cluster is a set of cores behind one local proxy: a controller that tracks its own cores and talks to the global home on their behalf. K0 = {C0, C2} with proxy H0, and K1 = {C1} with proxy H1. C2 is a third core, added because a local-sharing example needs two cores in one cluster; the course model allows this expansion when sharing needs it. The original H is the global home for X.

A local record (at H0 or H1) names cores in its cluster. The global record, labelled G, lives at H and names clusters. exclusive(K0) means K0 holds the only global permission for X. It does not mean H0 itself holds an M copy.

Start with X = 0 uncached. The rows are independent settled snapshots, not a sequence. Directory cells read G | H0 | H1, with clusters named at G and cores named at H0 and H1.

Local permission is bounded by global permission

H0 can grant its cores only what H granted K0. The global record for K0 limits the local states, and decides which local requests must go to H.

The last snapshot row shows the third case. K0 holds global write permission, so when C2 asked to read, H0 recalled the dirty value from C0, kept it, and gave S copies to both cores without contacting H. H0 is now K0's dirty-responsible holder, and G is unchanged because no message reached H. If C2 later asks to write, H0 invalidates C0 inside K0, again without H. Real two-level designs use the same idea: an inclusive shared L2 (one that holds every line its L1s hold) can hold M globally while its L1s share read-only copies.

Coverage crosses levels

Coverage means a record that leads a request to every cache that may hold a copy. If a core can hold a copy or has an admitted fill, the global level must keep coverage of its cluster, and the local level must cover the core. A local filter eviction cannot erase the last path by which a global request will find C0's M / 1. Records may over-cover; under-covering a cluster can lose a permission or the newest data.

This is an inclusive coverage policy. It does not require an inclusive shared data cache, and it is not the only hierarchical design. With K0 = {C0, C2} and K1 = {C1} the hierarchy saves no metadata (Hierarchy Trade-Offs counts it); it shows the scopes. Larger clusters let the global level summarize many cores, at the cost of local tracking and extra coordination.

For a concrete implementation to compare, gem5 MESI_Two_Level documents private L1 caches, a shared L2 with co-located directory information, and L1/L2 inclusion. That is one specific organisation, not the specification of our H0/H1 proxy extension.

Related topics

• Shared course model

• Tracking reclamation