Kenneth Perry

1
Dataflow Based Systems

• Kenneth Perry

2
Outline

• Overview of dataflow systems (architecture and
  software)
• Usages today (both hardware/software/inspired)
• Issues with dataflow

3
Dataflow Concepts

• Parallization at a low level
• System takes care of parallization neither the
  user or the programmer has to.
• Ability to scale to huge numbers of
  processors/cores.
• Based around graphs which link inputs and
  outputs.
• Generic approach applicable to both
  higher/lower levels in software/hardware.

4
Dataflow Graphs

• Programs are arranged into graphs based on their
  inputs and outputs.
• As inputs become available, outputs are
  automatically triggered causing it to be
  parallized.

5
Dataflow Architecture

• Von neuman model was considered unsuitable for
  parallelism, due to the global program counter
  and overhead for process switching
• Dennis and Misiunas created the dataflow
  architecture, which is based around parallelism
  at the lowest level.

6
Dataflow Architecture (cont)

• Two major types of dataflow architectures
• Static dataflow Original design by Dennis
• Uses memory addresses as data dependency tags
• Does not allow multiple instances of the same
  routine to be ran simulatinously, since the tags
  are not unique.
• Dynamic dataflow
• Uses associative memory instead to determine
  tags.
• Manchester Dataflow Machine and MIT Tagged Token
  Arch are two examples.

7
Programming for Dataflow

• Problems arose writing for the new architecture
  with current languages, since they were also
  based on the von neuman model.
• Came up with dataflow languages, which are
  designed around the different architecture.
• Some languages were converted by limiting
  functionality (single assignment C).

8
Programming for Dataflow (cont)

• Also known as reactive programming
• Used outside dataflow architectures, where data
  changes will affect other data (such as in a
  spreadsheet), or as a simple way to paralize
  applications at a higher level then architectures
  (component level instead of at an instruction
  level).

9
Modern Usage Spreadsheets

• Classic example of dataflow
• Any changes in one cell automatically updates
  every other cell that depends on it.

10
Modern Usage VPLs

• Visual programming languages show the inputs and
  outputs directly.
• Easily can create a dataflow graph from this
• Examples LabView, Microsoft VPL, KTechLab

11
Modern Usage TRIPS

• Uses a new instruction set (EDGE) to transfer
  dataflow graphs to the processor.
• Under specific tests, TRIPS at 366Mhz was
  comparrible to a Pentium IV at 3600Mhz.
• http//www.cs.utexas.edu/trips/

12
Modern Usage TRIPS (continued)

• Compiler creates dataflow graphs specifically for
  the hardware. Programmer doesnt have to handle
  the concurency at all.
• Very low power usage optimal for high
  performance at low power.
• Still being optimized hand optimization shows
  about double performance gain from current
  compiler is still possible.
• Goal is to reach one trillion calculations per
  second by 2012.

13
Modern Usage Multimedia Systems

• Both audio and video systems can be represented
  as a graph of filters and components
• Dataflow techniques can then be applied to
  parallelize multimedia processing

14
Multimedia Systems (cont)

• Multimedia systems are often on multiple
  computers and use multiple sensors to determine
  information.
• Can use dataflow techniques to manage acquiring
  data and processing throughout the system. (NIST
  Data Flow System II)

15
Grid Architectures

• Allowing dataflows to be specified, which is then
  converted to a dataflow graph and ran using ideas
  from the tagged dataflow architecture.
• Instead of doing it per instruction as in
  dataflow, performs it at a higher level to reduce
  overhead.
• Parallizes each actor in the system, which is
  transparent to the workflow designers.

16
Intel Research

• Threading Building Blocks
• Library designed to easily paralize programs
• Define everything as 'tasks' and build the
  dependency graph, library takes care of
  scheduling and running
• Larrabee
• An 80 core CPU currently in research (40x power
  of current intel quad cores)
• Designed to use dataflow ideas through TBB and Ct
  (a modified C/C designed around threading).
• Intel claims it will bring the 'death of the GPU'.

17
Issues with Dataflow

• Computers still heavily based around a few fast
  cores then a lot of smaller cores
• Dataflow can be used with this, but it has a lot
  of overhead due to the overhead of simulating the
  dataflow architecture.
• Programming languages have to be written/modified
  to take advantage of the full power of dataflow
  architectures.
• Most programming languages used currently are
  imperitive, and do not translate well directly to
  dataflow.

18
More issues

• As programs are parallelized more, more resources
  are needed to exploit the parallelization
  otherwise saturation/deadlocks will occur.

19
References

• The NIST Data Flow System II A Standardized
  Interface for Distributed Multimedia
  Applications. Antoine Fillinger, Lukas Diduch,
  Imad Hamchi, Mathieu Hoarau, Stéphane Degré,
  Vincent Stanford. IEEE International Symposium on
  a World of Wireless Mobile and MultiMedia
  Networks (WoWMoM), Newport Beach, California,
  June 2008.
• Gebhart, M., Maher, B. A., Coons, K. E., Diamond,
  J., Gratz, P., Marino, M., Ranganathan, N.,
  Robatmili, B., Smith, A., Burrill, J., Keckler,
  S. W., Burger, D., and McKinley, K. S. 2009. An
  evaluation of the TRIPS computer system. In
  Proceeding of the 14th international Conference
  on Architectural Support For Programming
  Languages and Operating Systems (Washington, DC,
  USA, March 07 - 11, 2009). ASPLOS '09. ACM, New
  York, NY, 1-12. DOI http//doi.acm.org/10.1145/15
  08244.1508246
• Arumí, P., García, D., and Amatriain, X. 2006. A
  dataflow pattern catalog for sound and music
  computing. In Proceedings of the 2006 Conference
  on Pattern Languages of Programs (Portland,
  Oregon, October 21 - 23, 2006). PLoP '06. ACM,
  New York, NY, 1-23. DOI http//doi.acm.org/10.114
  5/1415472.1415503
• Abramson, D., Enticott, C., and Altinas, I. 2008.
  Nimrod/K towards massively parallel dynamic grid
  workflows. In Proceedings of the 2008 ACM/IEEE
  Conference on Supercomputing (Austin, Texas,
  November 15 - 21, 2008). Conference on High
  Performance Networking and Computing. IEEE Press,
  Piscataway, NJ, 1-11. DOI http//doi.acm.org/10.1
  145/1413370.1413395
• http//techresearch.intel.com/UserFiles/en-us/File
  /terascale/Whitepaper-Ct.pdf

20
Questions?