From: ahemavathy (ahemavathy_at_hotmail.com)
Date: Tue Feb 17 2004 - 21:48:25 PST
This was an interesting paper. It was easy to understand and read. The paper studies two solutions to memory coherence problems on loosely coupled multiprocessors which share virtual memory. The authors rule out writeback page synchronization as it is very expensive and choose invalidation. Page ownerships can be static, central or distributed. Here again static ownership is too expensive which leaves central or distributed page ownerships. The paper studies these two solution and also improvises each.
In the centralized method there exists one manager which acts like a monitor: it knows the owner of every page and which processors have copies of the page. An improvement over this is a method where each page owner is aware of page copies thus eliminating additional confirmation message to the manager.
Distributed Algorithm can be fixed or dynamic. In the fixed method, every processor manages a fixed set of pages or segments and thereby knows the owner of those pages while in dynamic method every processor maintains a ptable with probable owner for every page thereby eliminating the role of manager. If the owner info is not correct, the requests are forwarded to the probable owner in that ptable.This can further be improved by enforcing a broadcast owner message when the number of copies of a page reaches a limit thereby eliminating too many forwards or it can be improved by making each processor possessing a valid copy of the page pass a copy of the page if it receives a read request and also maintain its own copy-set. This can eliminate unnecessary forwarding to the owner.
The experimental results clearly show that dynamic distributed page ownerships have good performance over other methods and thus facilitate shared VM on multiprocessors.
This archive was generated by hypermail 2.1.6 : Tue Feb 17 2004 - 21:46:52 PST