SIMD-MIMD Real-Time Comparisons (Chapter 7)

1
SIMD-MIMD Real-Time Comparisons(Chapter 7)

• References
• Stankovic, Spuri, Ramamritham, Buttazzo,
  Deadline Scheduling for Real-Time Systems,
  Kluwer, 1998, ISBN 0-7923-8269-2.
• Stankovic, Spuri, Natale, Buttazzo, Implications
  of Classical Scheduling Results for Real-Time
  Systems, IEEE Computer, Vol. 28, No 6, pp.
  16-25, June 1995.
• Meilander, Jin, Baker, Tractable Real-Time Air
  Traffic Control Automation, Fourteenth IASTED
  International Conference on Parallel and
  Distributed Computing and Systems (PDCS02), pp.
  483-488, November 2002.
• Importance of SIMD Computation Reconsidered,
  Meilander, Baker, Jin, International Parallel and
  Distributed Processing Symposium (IPDPS),
  Workshop on Massively Parallel Processing.
• Initially we will go through most of the slides
  used for the presentation of the paper,
  Tractable Real-Time Air Traffic Control
  Automation.
• Presentation was given by Will Meilander at PDCS
  Conference in November 2002.
• The PDCS slides will follow this set of slides.
• This slides describe a polynomial time solution
  for the ATC problem.
• A polynomial time multiprocessor (MP) solution to
  the ATC is not believed possible. In particular,
• MP solutions to virtually all real-time problems
  today include an online solution to one or more
  dynamic scheduling problems.

2

• Most dynamic scheduling problems are NP-hard.
• The MP cannot use static scheduling to solve most
  real-time problems (to avoid using dynamic
  scheduling).
• Observation A widely accepted principle in
  parallel computing is that MIMDs (or MPs) are
  more powerful than SIMDs. From a modeling point
  of view, this would mean that a MIMD could
  simulate a SIMD (of the same relative size) in
  constant time.
• Since SIMD processors are small ALU units, a fair
  interpretation of same size is needed.
• In general, only sketchy reasons are given for
  this claim such as
• A MIMD is a SIMD with fewer restrictions (i.e. no
  synchronization requirements) so anything a SIMD
  can do, a MIMD can do in no more time.
• Many MIMDs have extra hardware to provide fast
  synchronization to allow efficient simulation of
  SIMDs. (See 25, Kumar et. al. SIMD topics)
• Each MIMD processor can execute the SIMD program
  and synchronize at points where this is required.

3

• A discussion along this line occurs in 31,
  Smith, The Design and Analysis of Parallel
  Algorithms, pgs 62-65.
• Due to mass production, general purpose computers
  can be used as MIMD processors, but SIMD
  computers (due to low usage) have to be specially
  designed.

4

• If an AP (an enhanced SIMD) can solve the ATC
  problem in polynomial time, but no such solution
  is expected for a MIMD computer, then this casts
  doubt on a MIMD being able to efficiently
  simulate an AP.
• Since a SIMD can simulate an AP relatively
  efficiently, the preceding bullet seems to raise
  doubts as to whether the simulation of a SIMD by
  a MIMD can be very efficient.
• This work raises other open questions as well,
  both practical and theoretical.
• Can a MASC model be built that can efficiently
  execute current data parallel solutions for MPs
  but which avoid some additional work that MPs do
  (in addition to executing the solution steps).
• Dynamic scheduling of tasks
• Load balancing
• Synchronization
• Cache/memory coherency problems involving keeping
  multiple copies of data.
• In particular, how can the multiple instruction
  streams for MASC interact so as to avoid this MP
  difficulty?