CSCI%204717/5717%20Computer%20Architecture

1
CSCI 4717/5717 Computer Architecture

• Topic Vector/Array Processors
• Reading Stallings, Section 18.7

2
Vector/Array Computing

• Optimized for calculation rather than
  multitasking and I/O
• Design focus is to perform parallel mathematical
  operations on a vector or array of data elements
• Scalar processor would need to handle one element
  at a time.
• Limited market -- Research, government agencies,
  meteorology

3
Vector/Array Computing (continued)

• Target applications
• data-intensive/scientific research such as
• Aerodynamics, seismology, meteorology
• Continuous field simulation
• specialized (high-performance) graphics
  applications
• Applicable because of ever-increasing need for
  improved resolution and model capabilities

4
Array Processor

• Alternative to supercomputer
• Configured as a peripheral to mainframe or
  minicomputer
• Processor is only responsible for running vector
  portion of problem
• The Sony PlayStation 3 uses a processor
  consisting of one scalar processor and eight
  vector processors. Developed by IBM, Toshiba and
  Sony. (Source http//en.wikipedia.org/wiki/Vecto
  r_computer)

5
Vector/Array Operation

• Power of vector computing comes in the form of
  special processing instructions (Single
  Instruction, Multiple Data or SIMD)
• Lock-step execution of code issuing single
  instruction to a large number of identical
  processors (or ALUs) with a large register set
  working on different data elements
• Single master CPU keeps control of the entire
  process

6
Speed-Up Not Linear

• As with any parallel processing architecture, the
  realized speed up of a vector processor is not
  linear because of
• Overhead for managing parallel computations
• Bottlenecks for communication and storage
• Load of application doesn't always match
  available processors
• These problems have an increasing effect with
  increases in the number of processors

7
Data Pipelining

• The sequential nature of instructions allows for
  an instruction pipeline
• Vector computing tends to have data that is well
  organized too
• This allows for pipelining the data too
• Single decode for instruction
• Stages to fetch data, process data, store result
  in register

8
Data Pipelining (continued)

• Example To add an array of numbers, processor
  must have the following information
• a single "add" instruction
• start address for the data
• end address for the data

9
Vector/Array Programming

• The programming goal is to divide a large dataset
  into independent sets that can be operated on in
  parallel
• Requires a deep understanding of the algorithm
  being applied to the data
• Distribute data to different processors
• Initiate parallel processing
• Bring everything back together when parallel
  processing is complete

10
Vector/Array Programming (continued)

• Example Count the number of times a specific
  value appears in a large array
• Begin by breaking up array into smaller arrays,
  one for each array processor
• Each array processor, in parallel, counts the
  number of occurrences of the value
• Final sum is then computed by adding the results
  from all of the processors

11
Vector/Array Applications

• Which of the following applications would be
  better served by a vector or array computer than
  an SMP, cluster, or scalar processor? What
  component of the problem is parallel?
• Web search indexing
• Generating Fibonacci Sequence f(i) f(i-1)
  f(i-2)
• Weather prediction
• Image processing for a game
• Web site server
• Photoshop-type image processing

12
Scalar Programming

• The following two slides are based on the
  multiplication of two 100X100 matrices A and B
• DO 100 I 1,N
• DO 100 J 1,N
• C(I,J) 0.0
• DO 100 K 1,N
• C(I,J) C(I,J) A(I,K)B(K,J) (J 1,N)
• 100 CONTINUE

13
(J 1,N) Vector Programming

• The notation (J 1,N) indicates that operations
  on all indices J are to be carried out on N
  processors as a single operation
• DO 100 I1,N
• C(I,J) 0.0 (J 1,N)
• DO 100 K 1,N
• C(I,J) C(I,J) A(I,K)B(K,J) (J 1,N)
• 100 CONTINUE

14
Fork/Join Parallel Programming

• One method of parallel programming is the
  fork-join.
• Programs start as a single process known as a
  master thread
• The operation "fork" is used to indicate the
  beginning of sections of the program that are to
  be executed in parallel
• The operation "join" is used to terminate the
  parallel threads created by "fork" to bring the
  program back to a single, master thread

15
Fork/Join Method (continued)

• DO 50 J1,N 1
• FORK 100
• 50 CONTINUE
• J N
• 100 DO 200 I1,N
• C(I,J) 0.0
• DO 200 K 1,N
• C(I,J) C(I,J) A(I,K)B(K,J)
• 200 CONTINUE

16
Neural Networks
17
What?! A Blank Slide?!It must be over!!!