Draft . Do Not Copy

Many operating systems allow multiple processes to share the same physical object. This introduces aliases into the system, where different virtual addresses map to the same physical address. Aliases increase the size of the page tables, exerting increased pressure on TLBs and virtually-tagged caches. Use of segments and a global address space model is one approach to avoiding aliases. Most UNIX implementations, however, use a private address space model and allow for aliases. We present in this paper a common-mask scheme that allows multiple translations from different processes to the same physical address, to share a single translation entry in a TLB or page table. The scheme allows for a limited number of common regions that processes can share. A common region is a set of physical pages that a process maps using a fixed virtual address and a fixed set of permissions for each physical page in the set. The process context id is extended with a bit vector that identifies a set of common regions that a process shares with other processes. TLB and page table entries support shared entries by storing a bit vector that can identify common regions instead of a single process context id. The scheme requires aliases that use the common region to use the same virtual address, and can only support a few common regions. Further, the common-mask TLBs and page tables continue to work with conventional private mappings exactly as in the original system. We show how the common mask scheme can be applied to all levels of the translation hierarchy— hardware TLBs (fully-associative and set-associative), memory-based TLBs and page tables. The scheme we propose is quite general and can be applied in domains other than address translation where multiple keys map to the same data e.g., virtually-tagged caches, associative processors [26], relational databases and any hash table]. We show that the use of the common mask scheme in hardware TLBs can substantially reduce the number of TLB misses in a multiuser workload. We show a 50% reduction in the number of TLB misses for a 256-entry fully-associative TLB. However, the scheme does not work very well in TLBs with a small TLB reach. We expect the common mask scheme to be attractive when combined with other techniques that improve TLB reach—superpages, subblocking, and larger TLBs. A set-associative common-mask TLB would require accessing two sets and may increase TLB hit time due …