Summary of March 18 class

Distributed Shared Memory

Goals

• Aim is appearance of shared memory among processors on different nodes
• Achieved by exchanging pages over a network
• This processs should be transparent to processor on node

Accomplished by mapping manager on each node

• Problem: assuring the consistency of shared memory
• 2 approaches to updating a page
  • Invalidation - On write, ask all nodes with read access to discard page On next read, will get new copy of the page
  • Write-broadcast - On write, ask all nodes with read access to update their copies

Manager strategies

• centralized - one node manages all pages
• fixed distributed - each node manages fixed set of pages
• dynamic distributed - each node manages pages it owns

Implementation - Code example of central manager

The mapping manager consists of

• Read fault handler - Requests page from manager or directly from owner if a manager
• Read server - If owner: services request for page; If manager: asks owner to send
• Write fault handler - Requests page as in read fault handler Gets ownership of page
• Write server - If owner: sends copy of page, sets own access to nil; If manager: asks owner to send, sets copyset to nil

• About this document ...