Accelerating Bioinformatics Algorithms with Reconfigurable Computing

1
Accelerating Bioinformatics Algorithms with
Reconfigurable Computing

• Presentation to MAPLD Conference
• September 2004

2
Overview

• The Problem
• BioInformatics Algorithm Smith Waterman
• Current Implementations
• The Solution
• Viva as a Reconfigurable Computing SW HW Design
  Tool
• Hypercomputer Architecture for High-End RC
  applications
• The Implementation
• Smith Waterman Viva Code
• Smith Waterman Pipeline Design
• Smith Waterman Pipeline applied to Hypercomputer
  Architecture
• Smith Waterman Pipeline Primitives inside the
  FPGA
• The Results
• Visualization of Rat vs. Human Genetic Code
• Informal Benchmarks
• Other Potential Applications
• Seismic Data Processing Weather Modeling Image
  Rendering

Page 2
3
The ProblemEnormous Biosciences Problems

• Exploding Datasets in Biosciences
• DNA Sequencing
• Gene Expression
• Protein Identification

Page 3
4
The Need High-Speed High Sensitivity Algorithms

• High-Speed High-Sensitivity DNA and Protein
  Searching Algorithms
• Critical in virtually every branch of molecular
  biology.
• Smith-Waterman
• Theoretically optimal for sequence matching.
• BUT Compute Intensive!
• BLAST and FASTA
• Approximations.
• Faster than Smith Waterman, but less sensitive.

Page 4
5
The Need High-Speed High Sensitivity
Algorithms

• Comparative Genomics Comparing the genomes of
  related species
• Identifying genes, defining gene structure,
  elucidating evolutionary change, identifying
  regulatory elements and revealing combinatorial
  control of gene regulation
• Sequencing Effort
• Human sequence is completed other organisms now
  being sequenced
• Sequencing effort will require high sensitivity
  DNA searches and alignments
• SmithWaterman preferred method of choicemore
  accurate, specific
• NCBI BLAST, WU BLAWST not effective in
  low-coverage DNA situations
• RNA interference (RNAi) seeking novel therapies
  developing new drugs.
• The process Choosing the correct genetic
  sequence to effectively block a targeted
  messenger RNA (mRNA) without silencing additional
  genes
• Due to word length limitations, BLAST algorithms
  can miss sequences that have one or more
  mismatches compared to the query siRNA sequence
• Genome Annotation
• BLAST does not allow for long introns or
  frameshifts
• Smith-Waterman is both frameshift- and
  intron-tolerant.

Page5
6
The Need High-Speed Smith Waterman

• Large Matrix comparison
• Large datasets
• High level of detail for each SW calculation
• NOT heuristic approximations

Page 6
7
The Need High Performance Biosciences Platform

• Cluster Computingmost widely used platform. BUT
  there are diminishing returns
• Expensive to build, difficult to maintain
• Require significant power, air conditioning, and
  physical space
• Architecture inherently limits scalability and
  performance
• Reconfigurable Computing(RC)the promising
  alternative
• Advantages of a Custom Chip
• Implement algorithms directly in hardware
• Performance advantages of an ASIC, but without
  chip development cost
• Advantages of a General Purpose Platform
• Development time comparable to software
  development
• FPGAs can be reconfigured to perform other
  computational tasks.

Page 7
8
The SolutionFPGA-Programming Environment Viva

• VIVA GRAPHICAL LANGUAGE
• Capture natively parallel code
• Accommodate data of any type, size, or precision
• Tune algorithms for speed of execution or
  conservation of hardware resources
• VIVA EDITOR
• Call Viva algorithms from legacy code such as C,
  C, or Fortran
• Interactively debug code
• Import/Export EDIF files

• VIVA COMPILER/SYNTHESIZER
• Program multi-million gate designs
• Compile hardware designs quickly for efficient
  development
• VIVA LIBRARIES
• Reuse flexible Viva objects which accept any data
  type or size
• Target any hardware platform with a System
  Description
• Prototype Viva on any X-86-based Windows machine

Page 8
9
The Solution FPGA-based Hypercomputers
Page 9
10
Structure of an FPGA Processing Element
Page 10
11
Structure of a Processing Element Quad
Page 11
12
Structure of a Hypercomputer Accelerator Board
Page 12
13
The Prototype ImplementationSmith Waterman in
Viva Code
 
Page 13
14
Smith Waterman Program Flow

• As the query sequence is loaded, the Init_Cells
  object creates our initial column and stores it
  in SW_Cell_Mem.
• After this initialization period, SW_Cell_Mem
  will provide a cell to the chain SW_Iteration
  objects every clock cycle. It will also write a
  newly calculated cell every clock cycle.
• The SW_Cell_Mem object stores every nth column,
  where n is the number of SW_Iteration objects.

Page 14
15
Smith Waterman Cells

• There are as many cells as there are characters
  in the query sequence.
• The array of cells represent a column of the
  scoring matrix.
• The initial (zero) column is initialized and
  stored into the cell memory object, SW_Cell_Mem.
• Each cell contains the following four parameters
• Pattern a character from the query sequence
• Score the score of this cell in the current i,j
  position
• PatternStart the position in the query sequence
  from which the score was calculated
• DataStart the position in the reference
  sequence from which the score was calculated

Page 15
16
Cell Data Types

• Data Element size may be adjusted depending on
  usage
• Pattern contains as many bits as needed to
  encode characters from the sequences 4 bits for
  nucleotides.
• Score and PatternStart Equal in size. Must be
  large enough to encode the number of entries in
  the query sequence
• DataStart will be the largest data set as it
  must be able to encode any position in the
  reference sequence.
• Right size for the job
• Less circuitry is needed to calculate matches in
  smaller sequences.
• Smaller sequences may exploit more parallelism.

Page 16
17
Smith Waterman Data Sets
In this example, our Pattern contains 4 bits, for
modeling nucleotides. The Score and PatternStart
parameters contain 26 bits, so our query sequence
may contain up to 67,108,864 characters. The
DataStart parameter contains 27 bits, meaning our
reference sequence may contain up to 134,217,728
characters.
Page 17
18
Smith Waterman Iteration
Page 18
19
SW_Iteration Object

• Inputs
• Matrix_In receives a constant stream of cells.
  It is imperative for efficiency that the pipe
  remains full.
• Data receives a single character from the
  reference sequence. The cells computed will be
  for the column of the scoring matrix
  corresponding to the Data value.
• CountBy the radix of the algorithm (number of
  iteration objects)
• Init_J_In this iteration objects index in the
  chain of iteration objects
• ClkG System Clock
• Token_In a token pulse precedes a set of cells,
  allowing the iteration object to clear-out data
  from the previous set of cells
• Init initialization pulse utilized only before
  search commences
• G accompanies each valid cell

Page 19
20
SW_Iteration Object

• Outputs
• Matrix_Out newly-computed cell
• Token_Out passes token to next iteration object
• D accompanies each newly-computed cell
• Init_J_Out used by next iteration object
• I J current row and column used to report
  results

Page 20
21
Pipe Stages

• The SW_Iteration object contains four pipe
  stages.
• A cell is received by and produced by the
  SW_Iteration object every clock cycle.
• When a cell enters, it is coming from the
  previous column, so its values are those of the
  West neighbor.
• Since the cell in the row above any given cell is
  in the next pipe stage, access to both the North
  and Northwest neighbors values are possible.

Page 21
22
Parallelism

• If a given hardware system has enough physical
  resources to accommodate n SW_Iteration objects,
  the Smith Waterman program may operate on n
  columns in parallel.
• Hence n cells are computed every clock cycle.
• Each Virtex II 6000 can support 64 iteration
  objects

Page 22
23
The ImplementationPipeline Primitives Inside
the FPGA
Page 23
24
The Implementation Smith Waterman Pipeline
XPR Router
PE2
PE1 (Controller)
PE3
PE4
PE5
PE6
PE7
PE8
X86 System
 
Bus Controller
XPE Data Distribution
Page 24
25
The Results Rat vs. Human Genetic Code
Page 25
26
The Results Bacteria to Bacteria Comparison
Page 26
27
The Results Informal Statistics

• Total Operations / Second
• 1 Smith-Waterman Step includes
• 25 Logic Operations (Adds, compares, mostly 26-27
  bit ops, some single bit ops)
• 13 Data Reorder Operations (Move, Combine)
• 11 Data Stor (Assignment)
• Logic Operations Only
• 25 Ops 25Mhz 448 Smith-Waterman kernels
  280Billion Operations / Second
• Logic Data Operations
• 49 Ops 25Mhz 448 Smith-Waterman kernels
  550Billion Operations / Second
• Total Aggregate Communications Bandwidth of
  Systolic Array
• 12 88 25Mhz 26.4 Gb/s plus 7 22 50Mhz
  7Gb/s 34.1 Gb/s
• Resources Consumed / Resources Available
• PE2 PE7 60 to 70 consumed
• PE1 20 consumed XPE 5 XPR .1
• Compilation time
• Gates 70 Million Total
• Time to compile 20 Minutes
• Power Consumption
• Meter50 Watts

Page 27
28
Summary Conclusions

• This Viva prototype of the Smith-Waterman
  algorithm demonstrates that the algorithm can be
  parallelized for fast operation in an FPGA system
  and validates the usage of FPGAs to increase the
  speed of the Smith-Waterman algorithm compared to
  clusters
• Speed of the Prototype
• An HC-62 has the bandwidth to pass cells between
  7 FPGAs, allowing for 448 parallel SW_Iteration
  objects
• At a conservative 30 Mhz system clock speed, this
  gives 30,000 448 13.4 Billion Smith Waterman
  steps/second.
• Opportunities to further optimize the algorithm
  include
• Increasing the number of SW_Iterations that can
  be done in parallel (up to 100 Billion Smith
  Waterman steps/second)
• Increasing the clock speed of the hardware (up to
  1 Trillion Smith Waterman steps/second)

Page 28