100 likes | 218 Views
Parallel Architectures. Overview. This paper is a rehash of a paper that appeared in 1972 “Some Computer Organizations and Their Effectiveness.” IEEE Transactions on Computers, C-21(9):948-960, September 1972. Four basic parallel computer taxonomies SISD SIMD MISD MIMD
E N D
Overview • This paper is a rehash of a paper that appeared in 1972 • “Some Computer Organizations and Their Effectiveness.” IEEE Transactions on Computers, C-21(9):948-960, September 1972. • Four basic parallel computer taxonomies • SISD • SIMD • MISD • MIMD • Each has it’s purpose
Controller PE Fetch/Decode/Execute SISD • Single Instruction Single Data – sequential (unless you consider the micro-code level) • Pentium (of old) • VAX • SGI • Your stereo-typical Von Neumann machines
Fetch/Decode Controller Execute PE PE PE SIMD • Single Instruction Multiple Data – data level parallelism • All PEs do the same operations synchronously on their own separate data • Connection Machine CM-1, CM-2 • IUA CAAPP level • Vector processor within SISD machines such as the Cray
Controller Controller Controller PE PE PE Fetch/Decode/Execute Fetch/Decode/Execute Fetch/Decode/Execute MISD • Multiple Instruction Single Data – dataflow parallelism • Systolic array • Pipeline machines • Not talked about all that much
Controller Controller Controller PE PE PE Fetch/Decode/Execute Fetch/Decode/Execute Fetch/Decode/Execute MIMD • Multiple Instruction Multiple Data – control/process level parallelism • Cosmic Cube • Connection Machine CM-5 • IUA ICAP level
Controller Fetch PE PE PE Decode/Execute SPMD • Single Program Multiple Data – data and control level parallelism • Can be sub-classed from either SIMD or MIMD • PEs have identical programs but separate program counters (asynchronous SIMD)
Interesting Bits • Claims that SIMD is of “limited application” • I think this is unfair but maybe closer to the truth than what Thinking Machines, Inc. was pushing • “networking commodity workstations through a local area network” as a MIMD architecture • SETI started in 1984 which was prior to publication – I imagine he did not have such a statement in the 1972 paper • “However, we cannot expect that the ease of programming these improved configurations will advance” … “few of which are foreseeable.” • He speaks from experience and common sense
End Notes • Four basic parallel computer taxonomies • SISD • SIMD • MISD • MIMD • Each has it’s purpose • If someone builds/sells only 1 type, they will try to convince you that you don’t need the rest, no matter which type they build/sell