Multiprocessors

1
Multiprocessors

• To create powerful computers simply by connecting
  many existing smaller ones

Multiprocessors must be scalable
Multiprocessors can be used in a fault-tolerant
manner, i.e., increase performance and reliability
Multiprocessors now enbody a significant market
2
Key Design Issues

• How do parallel processors share data?
• How do parallel processors coordinate?
• How many processors?

3
Data Sharing

• Shared-Memory
• Single address space
• need synchronization
• The lock approach only one processor at a time
  can access the data
• Message Passing
• Every machine has its own private memory
• coordination is done through send receive

4
Connection Organizations

• Connected through a Singl Bus
• Connected through Network

Processor
Processor
Processor
Cache
Cache
Cache
Single bus
Memory
I/O
5
Multiprocessors connected by a network
Processor
Processor
Processor
Cache
Cache
Cache
Memory
Memory
Memory
Network
6
Programming Multiprocessors

• The obstacle is not the price, not the
  unavailability of appropriate applications or
  programming languages, the difficulty has been on
  developing parallel processing programs that has
  better performance results.
• Communication overhead
• You must get good performance, otherwise why
  bother
• programmer must know a good deal about the
  hardware (not portable in most cases)
• Amdahls law

7
Multiprocessor Cache Coherency

• The Cache Coherency Problem
• Snooping a mechanism to maintain cache coherency
• Snooping protocols
• Write Invalidate
• Write Update (Write through)

8
Network Topologies

• Ring
• Mesh
• n-Cube
• Crossbar

9
A Ring Network Example
10
Mesh
Example http//comsoc.ieee-cr.org/opticos2003/pdf
/coasin.pdf
11
Cube
http//www.geocities.com/CapeCanaveral/7997/hyperc
ube.html
12
Crossbar
http//www.npbook.cs.purdue.edu/figs/figure-10.4.p
df