Acceleration of stochastic algorithm on fpga system

 Simulation of stochastic model like SIR model is
very useful. However, the software simulation of
this kind of model becomes so slow with the
individual number increasing, so this project
provides a much faster hardware solution for it.
 A simulation is implemented on MATLAB to test
and verify the mathematical algorithm
 The final project is implemented on DE-115 board
using Verilog, which gives a much faster
performance than on PC
 To make the Verilog code more flexible and
variable, a MATLAB program is designed to
automatically generate the Verilog code

 In SIR model, each individual has to be one of the
three stages: S(Susceptible), I(Infectious) or
R(Recovered)
 Once the individual is recovered, it can’t be
infected again
 We can abstract the real world local relationship
network to a simplified one, where each individual
is connected to all the closely related ones, which
means they can infect others through connections
 Initially, one or several individuals are set to be
infectious and the transition rates(possibility for
infecting and recovering) are fixed

 If we observe the number of individuals of
each state, assuming all the individuals are
randomly connected, we should expect a
theoretical result as below
 The green one, which indicates the number of
infected individuals, is the data that we
observe in the actual simulation

 Instead of solving the problem in continuous
domain, we use discrete domain which is
compatible with hardware system. Moreover,
we use a uniform time step for simulation,
which lead to a fixed infecting and recovering
probability for every step.
 The most important component for this project
is a high performance random number
generator(RNG). In this case, we use a 64-bit
XOR feedback shift register as the RNG, which
is easy to implement in FPGA

 Since we need a easily changeable code for
implementing, a MATLAB program is designed
with which users can change just the parameters
and get auto-generated Verilog code.
 If SIR cell is in S state, any connected infected
individual could infect it with specified random
number
 If SIR cell is in I state, it could recover with
specified random number
 P_StoI and P_ItoR are fixed based on parameters, if
RNG gives out a number less than fixed P, the
comparer gives out 1 which enables the next stage.

 First of all, we test the algorithm in MATLAB
simulation
 MATLAB simulation gives out a similar result
as theoretical one which has a sharp rising and
a slow decaying.
 If we double the possibility for both infecting
and recovering, the shapes of the figures stay
similar but the time steps decrease a half,
which shows the validation of the algorithm

 For verification, there are three steps:
 First, with the same RNG, FPGA and MATLAB
should give out exactly same results, which proves
the Verilog code authentic
 Second, test the RNG that we used, especially
compare it with the MATLAB rand() function
which we assume is good enough
 Last, repeat the simulation on FPGA, with the
same seeds for RNG, the result should not change
 To examine the FPGA version, we build a same
RNG on MATLAB and give them the same seeds
as FPGA. Also, setting a smaller group of
individual makes it easier to identify them

 We designed a hardware solution to accelerate the
simulation for SIR model and implemented it on Altera
DE2-115 board
 MATLAB simulation was introduced to validate the
mathematical algorithm for the simulation
 Verilog code was automatically generated from a MATLAB
program which made it easier to change the parameters
 We used a 64-bit XOR feedback shift register as random
number generator in FPGA and prove it to be good enough
for simulation
 Verilog code was verified by comparing to the MATLAB
version with the same RNG
 The final implementation was stable during repeats
 The hardware solution gave a 5000 times better speed with
120 individuals than software

 1]A. Hoogland, J. Spaa, B. Selman and A.
Compagner, A special-purpose processor for
the Monte Carlo simulation of ising spin
systems, Journal of Computational Physics,
Volume 51, Issue 2, August 1983, Pages 250-260
 [2]http://www.cse.wustl.edu/~jain/cse567-
08/ftp/k_27trg.pdf
 [3]http://en.wikipedia.org/wiki/Compartmen
tal_models_in_epide miology

Acceleration of stochastic algorithm on fpga system

More Related Content

Viewers also liked (17)

Similar to Acceleration of stochastic algorithm on fpga system (9)

Recently uploaded (20)

Acceleration of stochastic algorithm on fpga system