02 Lecture Flynn IN
02 Lecture Flynn IN
Higher Performance
Computing
LECTURE 2: FLYNN’S TAXONOMY +
INTERCONNECTION NETWORKS
# Chapter Subtitle
Agenda
• Motivation for parallel computing
• Flynn’s Taxonomy
• Interconnection networks
• Dynamic Networks
• Static Networks
Motivation for Parallel Computing
• A multiprocessor is expected to reach a faster speed
than the fastest single-processor system.
• A multiprocessor is more cost-effective than a high-
performance single processor.
• If a processor fails, the remaining processors should be
able to provide continued service, albeit with degraded
performance.
Four Decades of Computing
Motivation for Parallel Computing
SISD (SIMD)
Single instruction stream Single instruction stream
Single data stream Multiple data stream
MISD (MIMD)
Multiple instruction stream Multiple instruction stream
Single data stream Multiple data stream
n data items
control unit
n ALUs
28
MIMD Architecture – Message Passing
• Message Passing Organization
• Each processor has access to its own local memory. No shared
memory
• Communications are performed via send-and-receive
operations. (data copy – consistency issues)
• Message-passing multiprocessors employ a variety of static
networks in local communications.
MIMD Architecture
Figure 2.4
• Mode of Operation
– Synchronous:
• a single global clock is used by all components in the system
(lock-step manner).
– Asynchronous:
• No global clock required
• Hand shaking signals are used to coordinate the operation of
asynchronous systems.
• Control Strategy
– Centralized: one central control unit is used to control
the operations of the components of the system.
• Switching Techniques
– Circuit switching: a complete path has to be
established prior to the start of communication
between a source and a destination.
• Topology
– Describes how to connect processors and memories
to other processors and memories.
– Static: direct fixed links are established among nodes
to form a fixed network.
– Dynamic: connections are established when needed.
Interconnection Network
Static Dynamic
p1 p2 ••• pN −1 pN
P1 P2 P3 P4 P5 P6
M1 M2 M3 M4
P1 P2 P3 P4 P5 P6
M1 M2 M3 M4
P1 P2 P3 P4 P5 P6
M1 M2 M3 M4
P1 P2 P3 P4 P5 P6
M1 M2 M3 M4 M5 M6
(a)
A crossbar switch connecting 4 processors
(Pi) and 4 memory modules (Mj)
(b)
Configuration of internal switches in a
crossbar
ISC 1 ISC
x-1
For example:
From 101 to 011
000 000
001 001
010 010
011 011
100 100
101 101
110 110
111 111
010
011
100
101
110
111
Hesham El-Rewini & ADVANCED COMPUTER ARCHITECTURE AND
Mostafa Abd-El-Barr PARALLEL PROCESSING
2.5 Analysis and Performance Metrics
• Dynamic Networks
• The network cost is the number of switching points
• The delay (latency) measured in terms of the amount of the
input to output delay
• Nonblocking network allows multiple output connection
patterns (permutation) to be achieved
• A fault-tolerant system can be simply defined as a system that
can still function even in the presence of faulty components
inside the system.
6 3 completely connected
network.
5 4
Linear arrays
Tree networks
Two-dimensional arrays
Cube network
D
D
S S
• Static Networks
• Degree of a node, d, is defined as the number of channels
incident on the node.
• Diameter, D, of a network having N nodes is defined as the
longest path, p, of the shortest paths between any two nodes
• A network is said to be symmetric if it is isomorphic to itself
with any node labeled as the origin
• Cost means the total number of links in the network