parallel processing architectures mimd varieties multiprocessor-shared ... · parallel processing...
TRANSCRIPT
1
Parallel Processing ArchitecturesMIMD - Multiprocessor
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
3BA5, 3rd Lecture, M. Manzke,Page: 1
MIMD VarietiesMultiprocessor-Shared Memory
3BA5, 3rd Lecture, M. Manzke,Page: 2
IN
P1 P2 Pn
M1 M2 Mm
Parallel Processing ArchitecturesMIMD - Multicomputer
3BA5, 2nd Lecture, M. Manzke,Page: 3
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
MIMD VarietiesMulticomputer - Passes Messages
IN
P1
M1
P2
M2
Pn
Mn
3BA5, 3rd Lecture, M. Manzke,Page: 4
2
Parallel Processing ArchitecturesMIMD – Dataflow Machine
3BA5, 2nd Lecture, M. Manzke,Page: 5
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
Dataflow Machine
3BA5, 3rd Lecture, M. Manzke,Page: 6
Dataflow attempts to solve the “von Neumann Bottleneck” problem by:
Linking data to instructionsIssuing the instructions to execute as
soon as all the necessary data binding is complete
Dataflow Machine - Example
3BA5, 3rd Lecture, M. Manzke,Page: 7
6 x
exp ×
+ f
× +
2 3
7
1 2 3 4 5
4
6
f ← 3x2+4x+7x ← 6
Dataflow MachinesClassified as MIMD in the Flynn taxonomy
3BA5, 3rd Lecture, M. Manzke,Page: 8Data (result) Instruction and data
PUDistributionnetwork
Arbitrationnetwork
PU
PU
IN
Instruction & Data Memory
Instruction & Data Memory
Instruction & Data Memory
IN
3
Dataflow - Conclusions
3BA5, 3rd Lecture, M. Manzke,Page: 9
The pseudo-random character of program assignments makes the efficient design of the arbitration and distribution networks a formidable task.
Further, dataflow architecture has no inherent ability to avail of the efficiencies possible in executing vector assignments, such as
y ? x + 3.7, where x and y are vectorswhich are frequent and execute very efficiently
on a SISD and more efficiently on a vector processor
So far, a number of dataflow machines have been constructed, but none have been marketed
Parallel Processing ArchitecturesMISD – Systolic Array
3BA5, 2nd Lecture, M. Manzke,Page: 10
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
Systolic Array
3BA5, 3rd Lecture, M. Manzke,Page: 11
Systolic, pertaining to or marked by systole, the regular contraction of the heart and arteries that drives the blood outward
These are two or three dimensional pipelines, suitable only for applications with very regular data movement.
For example, matrix multiplication
Parallel Processing ArchitecturesSpecial Purpose – ANN
3BA5, 2nd Lecture, M. Manzke,Page: 12
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
4
Non “Flynn-Classifiable” Parallel Systems
3BA5, 3rd Lecture, M. Manzke,Page: 13
Artificial Neural NetworksThese use a processing element
which implements a threshold function
Artificial Neural NetworksComputes:
3BA5, 3rd Lecture, M. Manzke,Page: 14
PE
x1
x2
xn-1
xn
y
c][Arithmeti 1
=×= ∑=
n
iia xas
if] logic[1 if 11 if 0
=
≥<
=ss
y
Block diagram of an artificial neural network
3BA5, 3rd Lecture, M. Manzke,Page: 15
PE
PE
PE
PE
PE
PE
PE
PE
PE
“Programming” consists of selecting the weight ai for each PEi and designing the interconnect between them
Parallel Processing ArchitecturesSpecial Purpose – Fuzzy Logic
3BA5, 2nd Lecture, M. Manzke,Page: 16
ParallelProcessingArchitectures
MIMD
SIMD
MISD
Hybrid
SpecialPurpose
MultiprocessorMulticomputerMulti-multiprocessorData flow machine
Array processorPipelinedvector processorSystolic array
MIMD-SIMD machines
MIMD-MISD machines
Artificial neural networkFuzzy logic processor
5
Fuzzy Logic
3BA5, 3rd Lecture, M. Manzke,Page: 17
The objective of Fuzzy logic, like ANNs, is to deal with noise and uncertainty
Deductive logic uses {TRUE, FALSE}Fuzzy logic uses truth values between
completely true and completely false and defines functions over these