1 Introduction to Data Parallel Architectures Sima, Fountain and Kacsuk Chapter 10 CSE462
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Basic Concept of Data Parallelism Memory Register 1Register 2 8 Bit ALU
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Basic Concept of Data Parallelism 8 bit wide Memory cells 8 bit wide Registers 8 bit ALU made of single bit ALUs
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Why is this useful? Every cell of a matrix Every pixel of an image Every record of an database Process:
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Thinking Machines l Connection Machine l Up to 65,535 processors in CM-2
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Connectivity l Want to support basic computations required at cell level –E.g. A[i,j] = (A[i-1,j] + A[i+1,j] + A[i,j-1] + A[i,j+1])/4 l To achieve this, cells can be connected in a variety of ways –Near neighbours –Tree –Graph –Pyramid –Hypercube –Multistage –Reconfigurable –Crossbar –Bus.
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Nearest Neighbours l Mapping spatially coherence data onto SIMD systems –Spatially correlated like images l Common to connect to NSEW, but diagonal has also been implemented –Applied to massively parallel systems –Scalable –Simple to implement
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Trees and Graphs l Problems expressed as graphs –E.g. database searching, model matching, expert systems, etc l No mathematically regular structure l Reconfigurability required l Binary an Quad trees common l Data bottlenecks going through roots of sub trees
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley The Pyramid l Combination of mesh and tree –Supports nearest neighbour plus (quad) tree communication –Local communications of mesh –Global communication of tree Consider example of moving data from one corner to another l Useful for data stored at multiple resolutions –Like images
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Hypercubes l 2 N processors – each of which has N links. l Fault tolerant l Shorter pathways than mesh
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Different Data Parallel Architectures l SIMD Program PE Array
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Different Data Parallel Architectures l Systolic or Pipelined DataALU 1ALU 2 ALU 3ALU 4 PE Array
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Different Data Parallel Architectures l Vectorizing Vector of Vectors Vector ALU Automatic Sequence
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Different Data Parallel Architectures l Associative and Neural Comparator Category Object Database
David Abramson, 2004 Material from Sima, Fountain and Kacsuk, Addison Wesley Principal Characteristics of data- parallel systems PropertySIMDSystolicPipelineVectorizingNeuralAssociative ProgrammabilityGoodFixed GoodPoorGood AvailabilityGoodPoor GoodPoor ScalabilityGoodFixed Good ApplicabilityWideNarrow WideNarrowWide