Add-on memory coherence directory

1. A method for memory coherence in a multiple processor system, the method comprising:

responsive to determining a number of sharers of a given memory region is greater than one and less than a threshold, marking the memory region as directory-based in an effective-to-real address translation table and storing a new directory entry for the memory region in an add-on cache directory;

responsive to a memory access resulting in a cache miss in a given processor, determining, by the given processor, whether the accessed memory region is marked as directory-based in the effective-to-real address translation table;

responsive to the given processor determining the accessed memory region is not marked as directory-based, performing the memory access using a snooping protocol; and

responsive to the given processor determining the accessed memory region is marked as directory-based in the effective-to-real address translation table, access a directory entry corresponding to the accessed memory region from the add-on cache directory to identify a home chip for the accessed memory region using a directory-based protocol and forwarding the memory access request from the given processor to the home chip to perform the memory access.

2. The method of claim 1 , wherein the directory entry comprises a sharing vector, wherein the sharing vector comprises a bit vector having a bit for each chip in the multiple processor system and wherein each bit in the bit vector is asserted if the corresponding chip is sharing the accessed memory region.

3. The method of claim 1 , further comprising caching the directory entry in a directory cache of the given processor.

4. The method of claim 3 , wherein the directory entry comprises least recently used (LRU) information for identifying candidate directory entries in the cache to replace.

5. The method of claim 1 , wherein the memory access request is a load request and wherein the home chip forwards a read request to a sharing node, the sharing node services the read request, and the sharing node forwards data for the read request directly to the given processor.

6. The method of claim 5 , wherein the sharing node services the read request on a given chip locally using a directory protocol.

7. The method of claim 5 , wherein the home node updates a sharing vector in the directory entry.

8. The method of claim 1 , wherein the memory access request is a store request and wherein the given processor forwards a read with intent to modify (RWITM) request to the home chip, the home chip forwards the RWITM request and a number of sharers to a sharing node, the home chip sends invalidate notifications to the sharers, the sharing node services the RWITM request, and the sharing node forwards data for the RWITM request directly to the given processor.

9. The method of claim 8 , wherein the sharing node services the RWITM request on a given chip locally using a directory protocol.

10. The method of claim 8 , further comprising:

receiving, by the given processor, the data for the RWITM request and the number of sharers;

counting, by the given processor, confirmation notifications indicating the sharers have invalidated their copies of the data; and

responsive to the number of confirmation notifications equaling the number of sharers, sending a final confirmation message to the home node.

11. The method of claim 8 , wherein the home node updates a sharing vector in the directory entry.

12. The method of claim 1 , further comprising:

attempting, by the given processor, to service the memory access locally using a snooping protocol in parallel with forwarding the memory access request to the home chip to perform the directory-based memory access.

13. The method of claim 1 , wherein storing the new directory entry for the memory region in the add-on cache directory comprises:

identifying a directory entry to be replaced from the add-on cache directory; and

replacing the identified directory entry with the new directory entry in the add-on cache directory.

14. The method of claim 13 , wherein identifying the directory entry to be replaced comprises identifying a least recently used directory entry in the add-on cache directory.

15. The method of claim 13 , wherein identifying the directory entry to be replaced comprises identifying a directory entry having a highest number of sharers.

16. The method of claim 13 , wherein identifying the directory entry to be replaced comprises identifying a subset of directory entries in the add-on cache directory having a number of sharers greater than the threshold and identifying a least recently used directory entry within the subset of directory entries.