Efficient Associative SIMD Processing for Non-Tabular Data

1
Efficient Associative SIMD Processing for
Non-Tabular Data

• Jalpesh K. Chitalia and Robert A. Walker
• Computer Science Department
• Kent State University

2
Presentation Outline

• ASC Processor Architecture
• Associative Features
• Structure Codes
• Represent Data Structures
• Structure Code Operations
• Summary and Future Work

3
The ASC Processor

• A scalable design implemented on a million gate
  Altera FPGA
• SIMD-like architecture
• Currently, 36 8-bit Processing Elements (PE)
  available
• 8-bit Instruction Stream (IS) control unit with
  8-bit Instruction and Data addresses, 32-bit
  instructions

4
The ASC Architecture
5
The ASC Architecture

• Each PE listens to the IS through the broadcast
  and reduction network
• PEs can communicate amongst themselves using the
  PE Network
• PE may either execute or ignore the microcode
  instruction broadcast by IS under the control of
  the Mask Stack

6
The ASC Features

• Associative Search
• Each PE can search its local memory for a key
  under the control of IS
• Responder Resolution
• A special circuit signals if at least one
  record was found
• Masked Operation
• Local Mask Stacks can turn on or off the
  execution of instruction from IS

7
The ASC Example

• Select from Students where Grade gt 90

8
The ASC Features

• Constant Time Associative Operations
• Associative Search
• Finding minimum or maximum in a field
• Ideal for Database processing
• Data is organized in tabular format
• Each tuple in a table can be processed by one PE
• PE Network
• Many parallel algorithms require all PEs to move
  contents in a regular pattern
• E.g. Image Convolution, Matrix Multiplication

9
Non Tabular Data Structures

• Many applications use linked list based data
  structures
• For example, plain HTML parsing can be done using
  tree-structure
• Similarly, XML or Object-relational databases can
  be represented using trees only
• Game programming uses tree-based algorithms, and
  demand much of processing power
• Complier construction uses directed acyclic
  graphs and tree structures

10
Data Structure Codes

• A unique coding scheme
• Allows representation of any data structure into
  a tabular format
• Tabular format allows division of data amongst
  the PEs
• Also known as structure code
• Different coding schemes for different data
  structures
• Uses Associative Search feature of the ASC
  Processor

11
Simple List-based Structures

• The left figure shows a possible representation
  of 1D and 2D arrays. Representations of this kind
  should be particularly useful for adjacency
  matrix or sparse matrix
• The right figure represents stack and queue,
  depending on the use of appropriate functions

12
Complex List-based Structures

• Each digit-position indicates the level of a tree
• Each value in that position indicates the
  position of that child from the left
• Discussions henceforth are confined to trees
• Graphs are read in a slightly different manner

13
Structure Code Operations

• Two sets of constant-time operations scalar and
  parallel
• Scalar Instructions are simple mathematic
  operations
• Search parent, child or root
• Finds value of structure code for next or
  previous nodes
• Parallel Instructions use complex associative
  operations
• Finds code for next, previous or both siblings
• Locates the required node for further processing

14
Scalar Operations

• fstcd leftmost child
• nxtcd right sibling
• prvcd left sibling
• trncd truncate (parent) node
• trnacd truncate all (root) node
• Can be used to allocate a new node
• Limited use in searching records

15
Parallel Operations

• Index instructions
• Flags a node of the result to ease further
  processing
• nxtdex (next or right), prvdex (previous or left)
  and sibdex (siblings or both left and right)
• Value instructions
• Returns the exact structure code of the result
• nxtval (next or right) and prvval (previous or
  left)
• Can be used to locate nodes in any tree
• Uses parallel and associative hardware resources

16
Summary

• SIMD-based computers are more suited for database
  processing
• ASC processor with its associative operations
  makes them more efficient
• Structure codes translate non-tabular data
  structures into a tabular format
• Tree and Graphs can be represented and evenly
  divided like records in a table
• Object-relational databases, XML databases, plain
  HTML processing
• Stacks, Queues, multi-dimensional arrays can be
  represented
• Structure codes not required for even division

17
Future Work

• Efforts are in place to clean the architecture
• To allow multiplier/divider on each PE
• To accommodate RISC instruction set
• Unpacked bytes are used to support
  variable-length structure code
• Can be avoid with an efficient divider unit
• Developing applications that use structure codes

18
Questions?
Thank You!