Using spmd or parfor in Matlab

Question

I am currently trying to run experiments in parallel using MATLAB 2011b that are very time-consuming. I am wondering if someone could help me 'translate' the following block of generic (non-working) parfor code into something that will work in the spmd code.

amountOfOptions = 8;
startStockPrice = 60 + 40 * rand(1,amountOfOptions);        
strike = 70 + 20 * rand(1,amountOfOptions);                 
v = 0.35 + 0.3 * rand(1,amountOfOptions);                   
IV = 0.25 + 0.1 * rand(1,amountOfOptions);                  
sigma = 0.15 + 0.65 * rand(1,amountOfOptions);              
riskFreeRate = 0.05 + 0.1 * rand(1,amountOfOptions);        
tn = fix(1 + 3 * rand(1,amountOfOptions)); 
tic;
for g=1:amountOfOptions
        for i=1:10                          
        N = i*5;              
        Cti = zeros(1,N);                       
        Sti = zeros(1,N);                       
        B = zeros(1,N);                         
        d1_ti = zeros(1,N);    
        delta_t = zeros(1,N);
        ctn = 0;
        cmtn = 0;
        result = 0;
        t = (1:N)/N;        
        dt = 1/N;                         
        c_mt0 = 0;                                      
      for j=1:10
            B = sigma(g)*randn(1,N);      
                 part1 = startStockPrice(g)*normcdf((log(startStockPrice(g)/strike(g))+(riskFreeRate(g)+(0.5*(IV(g))^2))*(tn))/(v(g)*sqrt(tn)),0,sigma(g));
                 part2 = exp(-riskFreeRate(g)*tn)*strike(g)*normcdf((log(startStockPrice(g)/strike(g))+(riskFreeRate(g)-(0.5*(IV(g))^2))*(tn))/(IV(g)*sqrt(tn)));
            c_mt0 = part1 - part2;          
            Sti(1) = startStockPrice(g);       
                for j = 2:N-1
                   Sti(j)=Sti(j-1)*exp( (riskFreeRate(g)-dt*0.5*sigma(g)^2) * t(j)*dt + sigma(g)*B(j));
                end                                                               
            Sti(N) = Sti(N-1)*exp( (riskFreeRate(g)-dt*0.5*sigma(g)^2) * t(N)*dt + sigma(g)*B(N));

                    parfor i = 1:N-1
                         d1ti(i) = (log(Sti(i)/strike(g)) +  (riskFreeRate(g) + v(g).^2/2) * (tn - t(i))) / (v(g) * sqrt(tn - t(i)));
                    end 
                    parfor i = 1:N-1 
                        Cti(i) = Sti(i).*normcdf((d1ti(i)),0,sigma(g)) - exp(-riskFreeRate(g).*(tn(g) - t(i))).*strike(g).*normcdf(((d1ti(i) - v(g)*sqrt(tn(g) - t(i)))) , 0 ,sigma(g));  
                    end
                        if((Sti(N) - strike(g)) > 0) 
                            ctn = Sti(N) - strike(g);
                        else
                            ctn = 0;
                        end
                    parfor i = 1:N-1
                         delta_t(i) = normcdf((d1ti(i)),0,sigma(g)); 
                    end
           cmtn = ctn - c_mt0*exp(riskFreeRate(g)*tn(g));                                                  
           result= cmtn + result;
        end
        result= result/10;                                                               
      end
end
time = toc;

I've always used `parfor` over `spmd` because it's more logical for me. Since `parfor` requires that each iteration within the loop be independent of all other iterations, you need to ensure that is the case. Please create a minimum example for me. Do you want a toy example instead? — krisdestruction, Apr 22 '15 at 16:27
Example using parfor on official site [link](http://www.mathworks.com/help/distcomp/examples/simple-benchmarking-of-parfor-using-blackjack.html) _italic_ **bold** `code` — Султан Гашимов, Apr 22 '15 at 16:33
Anyways let me know if my solution worked for you. If so, kindly mark it as the accepted answer :) — krisdestruction, Apr 22 '15 at 16:33
Yeah never post full code in the comments. Just update your answer. — krisdestruction, Apr 22 '15 at 16:33

krisdestruction · Accepted Answer · 2015-04-23T00:27:06.110

I've always used parfor over spmd because it's more logical for me. Since parfor requires that each iteration within the loop be independent of all other iterations. It's as easy as encapsulating it using the following method.

% Initial Variables
amountOfOptions = 8;
startStockPrice = 60 + 40 * rand(1,amountOfOptions);        
strike = 70 + 20 * rand(1,amountOfOptions);                 
v = 0.35 + 0.3 * rand(1,amountOfOptions);                   
IV = 0.25 + 0.1 * rand(1,amountOfOptions);                  
sigma = 0.15 + 0.65 * rand(1,amountOfOptions);              
riskFreeRate = 0.05 + 0.1 * rand(1,amountOfOptions);        
tn = fix(1 + 3 * rand(1,amountOfOptions)); 

% Open Parpool
try
    parpool;
catch
end

% Use parfor
parfor i = 1:amountOfOptions
    [startStockPrice(i),strike(i),v(i),IV(i),sigma(i),riskFreeRate(i),tn(i)] = fun( startStockPrice(i),strike(i),v(i),IV(i),sigma(i),riskFreeRate(i),tn(i) );
end

Then you can create the encapsulating function fun that will accept all the parameters and process/reoutput them. It will have the following definition/header:

function [startStockPrice,strike,v,IV,sigma,riskFreeRate,tn] = fun( startStockPrice,strike,v,IV,sigma,riskFreeRate,tn );

@СултанГашимов Anytime :) – krisdestruction Apr 22 '15 at 16:46 — krisdestruction, Apr 22 '15 at 16:46
Can the downvoter care to comment before they downvote? – krisdestruction Apr 23 '15 at 00:25 — krisdestruction, Apr 23 '15 at 00:25

Using spmd or parfor in Matlab

1 Answers1