1. Distributed Shared Memory
CIS 720
Distributed Shared Memory

2. Shared Memory Shared memory programs are easier to write
Multiprocessor systems
Message passing systems:
- no physically shared memory
- need to provide an abstraction of shared memory: Distributed Shared Memory

3. Shared
Memory

4. Single copy of each variable at a fixed location
Multiple copies

5. Consistency Models w(x)v: write value v into x
r(x)v: read of x return value v
Uniprocess programs:
- all operations are totally ordered
- read operations return the value written
by the most recent write operation
w(y)2 w(x)2 r(y)2 w(x)1 r(x)

6. Migratory protocol Each page (variable) has a single copy
Initially, pages are distributed among the processes.
To read/write a variable: If page is locally available, perform the operation; Otherwise, DSM layer sends request for the page to be moved locally.

7. Migratory protocol can suffer from trashing.
Solution: Maintain multiple copies

8. Consistency model
In the presence of multiple copies, we need to look at values written by other processes

9. Definitions Program order x = 3 w = x x = 5 y = z y = 4 z = 3
x must be 3
We will first give some definitions. Then, I will go through some existing notions of DSM consistency.
Program order refers to the order in which a process issues operations. Program order for each process
Execution History:
…….,x = 5; x = 3; w = x; y = z; y = 4; z = 3; …….
- legal execution history

10. Atomic consistency
Any read to a memory location x must return the value stored by the most recent write on x that has been done.
The order of events must coincide with the real-time occurrence of non-overlapping events

12. Write-invalidate Protocol
Each page has an owner;
Protection modes: read, read_and_write, none
Read operation: if not locally available, then obtain a read-only copy. Set protection mode to read.
Write operation: Contact the current owner; get the page and its ownership; send invalidate messages to nodes that have copies; sets the protection to read_and_write;

14. Write-through Protocol
Multiprocessor with snooping cache
Read operation: if variable not in cache, then read from main memory and cache it. Else, read from the cache.
Write operation: update shared memory and invalidate cache entries.

16. Sequential Consistency
Lamport 1979
A multiprocessor system is sequentially consistent if the result of any execution is the same as if
the operations of all processors were
executed in some sequential order
and
the operations of each individual
processor appear in this sequence in the
order specified by its program.

17. Sequential Consistency
a = x
b = x
c = x
d = x
x = 3
w = x
x = 5
y = z 3 5
y = 4
z = 3
…….,x = 5; x = 3; w = x; y = z; y = 4; z = 3; …….
The processes issue the operations; The DSM must create a legal serialization that satisfies the following two properties: First is an intra-process constraint
The second is an inter-process constraint; no two processes can disagree on the order of two write operations

19. Brown’s Algorithm
Each process has a queue Ini of invalidation requests

21. Brown’s algorithm w(x)v: perform all invalidations in In queue;
update main memory;
place invalidation request in In
queue of each process
r(x): if x in cache then read x;
else perform all invalidation in Ini
read from the main memory

23. Whenever main memory is accessed, all outstanding invalidations must be performed.
Sequential consistency is maintained.

24. Distributed implementation
All processes maintain a local copy
Write w(x)v: send message to all processor updating x to v
Read r(x): read local copy