Parallel Computing

Flynn’s Taxonomy

Classification of computer architectures

Four classifications:

• SISD - Single Instruction Single Data

• MISD - Multiple Instruction Single Data

• SIMD - Single Instruction Multiple Data

• MIMD - Multiple Instruction Multiple Data

SIMD Extensions

SIMD processing - do same thing to many data objects

64-245 bit words are large enough for many separate bytes inside

Needs special software and CPU hardware e.g. streaming SIMD extensions (SSE)

Streaming SIMD Extensions (SSE) uses

Uses:

• Image processing - multiple pixels at a time

• Video processing - processing blocks

• Array/vector processing

• Text processing - especially XML

• Offers general speed ups - does depend on application

Symmetric Multiprocessors (SMP)

MIMD system

Multiple CPUs share main memory and I/O

Increases performance

Scalable until bus is saturated

Heterogeneous Multi-processing

Combine big performance cores with small energy efficient cores

Core used determined by task requirements

Simultaneous Multithreading (SMT)

Hardware multi-threading on superscalar CPUs

• One physical core used logically as two

Executes multiple instructions at the same time using redundant execution units in processor

Data parallelism

split data to make independent parallel tasks

Task parallelism

Split code up to run on separate threads, cores etc