The Basic Idea

Consider three variables x, y, and z, whose size is smaller than that of a page. In a page-based DSM system, the memory for these variables may be allocated on the same page, resulting in false sharing. Consider a mapping of the page that contains these variables to three different, non-overlapping, virtual address regions, starting at addresses v₁, v₂, and v₃, so that each of the locations in the page can be viewed via three different virtual addresses. Each of the regions is called a a view. Since access permission is controlled through the virtual memory mechanism, it follows that protection and fault handling can now be applied in three different and independent ways, one for each view. Consequently, the mechanism for maintaining page consistency can also arbitrate between three different policies, each using the access capability setting of one of the views. Figure 1 depicts this situation.

  

Figure 1: Mapping three virtual views to a single memory page in which three variables reside.

Now suppose the application uses different views to access different variables, say, the variable x is accessed via $v_1+\mbox{offset}(x)$, y via $v_2+\mbox{offset}(y)$, and z via $v_3+\mbox{offset}(z)$. Then, although the variables reside in successive memory addresses on the same physical page, they are seen by the application as if placed in three different (virtual) pages. It is now possible to manage a separate access policy for each variable using its respective view. Moreover, an independent consistency protocol can be implemented for each variable, despite the fact that all three reside on the same page.