TLB Consistency on Highly-Parallel Shared-Memory Multiprocessors

Multiprocessors that store the same shared data in different private caches must ensure these caches have consistent copies. Almost all known solutions to this cache consistency problem are only suitable for architectures with a few tens of processors (PEs). Efficient solutions to the TLB (translation lookaside buffer) consistency problem, a special case of the cache consistency problem, can be found for highly-parallel, shared memory multiprocessors (HPSMMs) with many hundreds of PEs for the following reasons: the number of references to address translation information per modification is very large; the cache for storing translation information can be present anywhere on the path from the PEs to memory; when the memory mapping needs to be modified, one can often select which translation information to change; and obsolete mapping information can be used until permanent changes must be made. We present three general methods that exploit these features and can be used on HPSMMs to main..