Design Review

1
Design Review 1
Team 1 Cowboy and Indians
March 1st, 2005
Mark Hairgrove
Jay Patel
Prameet Shah
Samir Shah
2
Design Goals

• A robust security system that would provide
  notification upon the release of a bio-terror
  agent in a public domain such as an Airport.
• Distributed network of sensor nodes reporting to
  a centralized server
• Network communication with two-way authentication
  protocol
• Server side virus alert system
• Variable scan granularity depending on current
  National Terrorist Attack Threat level
• Database upgradeability via network uplink
• Real-time sensor diagnostic monitoring
• Robust signal processing to handle noisy signals
  from a vast open space
• Run BLAST fast

3
Design Exploration
BLAST Algorithm

• BLAST utilizes most of the computation time
• Want to maximize throughput
• Optimize for
• Speed
• Memory/Database Usage
• Ease of Implementation
• How?

4
Design Exploration
BLAST Algorithm
index time self children called
name
ltspontaneousgt 1 100.0 0.01 79.31
main 1 74.75
0.00 38154/38154 get_hits 2
0.87 2.93 38154/38154 extend_hits
3 0.27 0.48 1/1
generate_nbr_words 6 0.00
0.00 1/1 check_qry
10 ---------------------------------------------
-- 74.75 0.00 38154/38154
main 1 2 94.2 74.75 0.00 38154
get_hits 2 -------------------------------
---------------- 0.87 2.93
38154/38154 main 1 3 4.8 0.87
2.93 38154 extend_hits 3
0.40 1.61 3171601/4522804 blosum62
4 0.35 0.24 150213/150213
save_hsps 8 0.06 0.28
150213/334213 b62_score 7 -----------------
------------------------------
25 Char Query, 14M DB
5
Design Exploration
BLAST Algorithm
94.2
4.8
0.9
6
Design Exploration
BLAST Algorithm

• BLAST spends most of its time finding hits
  (string comparisons are expensive)
• In order to get it to run fast, we have 2 options
• Amdahl says make string compare fast as we spend
  most of our time finding hits
• We could also introduce some parallelism
• Split algorithm and pipeline?
• Split Database?

7
Design Exploration
Architecture

• Design 1 NIOS Arbiter

8
Design Exploration
Architecture

• Design 1 NIOS Arbiter

• Pros
• NIOS cores are very fast gt can run at speeds of
  up to 180 MHz
• String comparison in custom hardware can be done
  in two-level logic and is thus faster than
  software
• Cons
• We have a clear load imbalance with this design
  (ARM will be on vacation most of the time)
• Implementing BLAST across NIOS and ARM is
  complicated

9
Design Exploration
Architecture

• Design 2 Dual ARM

10
Design Exploration
Architecture

• Design 2 Dual ARM

• Pros
• Parallel searching halves computation time
• Increases throughput
• Cons
• Memory transfer is slow because NIOS is
  intermediate medium
• We have a clear load imbalance with this design
  (ARM will be on vacation most of the time)
• Implementing BLAST across NIOS and ARM is
  complicated

11
Design Exploration
Architecture

• Design 3 Dual ARM, Split Database

12
Design Exploration
Architecture

• Design 3 Dual ARM, Split Database

• Pros
• Allows unimpeded parallel computing
• String comparison throughput is doubled from
  parallel ARM
• Cons
• Extra cost of hardware
• Increased complexity of the system

13
Verification and Test Plan

• We will test our system on two levels unit and
  system
• Server, Network and Distributed Nodes
• This will be tested with a series of nodes that
  are simulated in software
• Custom Hardware
• Test with verilog testbench in ModelSim and cross
  reference with MATLAB
• BLAST Algorithm
• Test with a sequence of test vectors
• Memory
• Test with programs that erase, write and read
  from memory

14
Demonstration Plan
Demo 1

• BLAST Algorithm ported to C from MATLAB
• Convert MATLAB code to C in a manner that allows
  easy transfer to hardware
• Test and prove correctness
• Custom Hardware Tested
• Complete the Amino Acid mapping hardware in
  Verilog
• Use a test bench as proof that it works
• Simple Networking
• NIOS Ethernet connectivity installed and
  configured
• Simple socket communication/message transfer to
  PC and back
• DSP Design Exploration
• Explore matching algorithms
• Write MATLAB code to convert from waveforms to
  nucleotides
• Develop encoding scheme for nucleotides
• Explore real-time vs. post signal processing

15
Schedule
16
Schedule
17
Division of Labor

• Mark Hairgrove
• Network integration, authentication protocol,
  simulated testing, network diagnostic tool
• Jay Patel
• Custom hardware design, memory interfaces and
  control, database upgradeability feature
• Prameet Shah
• Deciphering input signals from gene sequencer.
• Samir Shah
• Porting BLAST to C, optimizing algorithm, running
  on NIOS, splitting across ARM and integrating
  with memory

18
Questions?