1. Presented by Evan Yang

2. Overview of Munin  Distributed shared memory (DSM) system  Unique features Multiple consistency protocols Release consistency  Annotate data items according to how they are shared  Implemented on the V kernel

3. Review  Distributed shared memory (DSM) systems provide an abstraction for sharing data between processes that do not share physical memory  Spares programmer the concerns of message passing  Central problem is scalability

4. Consistency: other models  Sequential  Causal  Processor  Pipelined RAM  Entry (H)  Scope (H)  Weak (H)

5. Consistency: Munin  Release consistency Weaker than sequential Cheaper to implement  Each shared memory access is either synchronization or ordinary access For synch, either release or acquire  Sequential vs. release

6. Multiple consistency protocols  Annotate by expected access pattern  Choose consistency protocol suited to pattern  Why? No single consistency protocol is best suited for all parallel programs

7. Basics of Munin Programming  CreateThread(), DestroyThread() user_init() - # of threads and processors  Shared objects correspond to a single shared variable  CreateLock(), AcquireLock(), ReleaseLock(), CreateBarrier(), WaitAtBarrier()  Delayed Update Queue (DUQ)

8. Munin Protocol Parameters  I – Invalidate or update?  R – replicas allowed?  D – delayed operations allowed?  FO – fixed owner?  M – multiple writers allowed?  S – Stable sharing pattern  Fl – flush changes to owner?  W – Writable?

9. Annotations in Munin  Read-only  Migratory  Write-shared  Producer-consumer  Reduction  Result  Conventional  ChangeAnnotation()

11. Implementation and Performance of Munin  Munin vs. message passing Two programs: Matrix Multiply and Successive Over-Relaxation (SOR) Hand-coded the message passing versions Same hardware, identical computations  Assess the overhead for each approach

12. Matrix Multiply  Multiply two 400x400 matrices  Performs within 10% of message passing version  By reducing the number of access misses, Munin comes within 2% of message passing version

15. Successive Over-Relaxation  Used to model natural phenomena (determining temperature gradient over a square area)  Divide area into sections, compute iteratively

18. Summary  Approximately as efficient as message passing  What little is lost in efficiency is gained in decreased program complexity

19. Critique of Munin Study  Compare/contrast with other consistency models Only compared against hand-coded message passing  Didn’t challenge how Munin scales Researchers did say they will do another study where Munin is implemented on a high-speed network of supercomputer workstations