Multi-level coherency management for high-performance shared virtual memory multicomputers

A multicomputer with a thousand nodes is likely to include tens of giga-bytes of physical memory and an even larger virtual address space. The total number of blocks in such a system will be very large, potentially leading to unacceptably high storage overhead for the required mapping tables. The authors propose a solution to this problem, based on the assumption, which is validated by trace-driven simulations, that at any one time only a small percentage of the address space of every process is actually shared. Hence, the mapping tables are partitioned into tables for strictly local, non-shared pages and special tables for shared pages, where it is possible to specify different locations for different blocks in the page. During normal execution, pages are dynamically moved between the different mapping tables, as the need arises. The proposed coherency algorithm introduces the use of ownership (see R.H. Katz et al., 1985) at both the page level and the block level in order to provide block-level sharing and also support dynamic page migration to reduce message traffic and the size of mapping tables