Slow FOR loop for 2nd iteration and after

I'm running a code that takes seismograms and stacks them according to their nearest neighbor. The first time through this loop it runs in .28 seconds (the line "disp('time for RFbeamer loop')" near the bottom), but each iteration afterwards takes almost 2 seconds to run. It stays at about 2 seconds each time after, it doesn't keep increasing in time. Only the variables that went in the first time go in for the rest of the iterations. With that said, I cannot figure out why the 2nd-nth iterations take almost 2 seconds longer!

for ij=1:m;
  tic
  centroid=GOOD(ij).name
  numleft=num2str(m-ij);
  disp(['Number of months left is ' months_left '. Current low pass frequency is ' freq ' Hz. Number of stations left for part 2 is ' numleft '.'])
  load(centroid);
  sample_before=(data(3).SACT0/data(3).dt)-(window(1)/data(3).dt);
  sample_after=(data(3).SACT0/data(3).dt)+(window(2)/data(3).dt);
  centroid_vertical_data=data(3).data(sample_before:sample_after);
  ttest=zeros(size(timelist));
  dtest2=ttest;
  dtest=ttest;
  dtime=abs(timelist-timelist(ij));
  dlats=6371*(stlats-stlats(ij))*pi/180;
  dlons=(6371*cos(stlats*pi/180)).*(stlons-stlons(ij))*pi/180;
  dd=sqrt(dlats.^2+dlons.^2);
  inbeam1=find(dtime<15);
  ttest(inbeam1)=ones(1,length(inbeam1));

    inbeam_hole=find(dd<=ri);
    dtest(inbeam_hole)=ones(1,length(inbeam_hole));
    ttest_hole=ttest.*dtest;
    inbeam_hole2=find(ttest_hole>0);

    inbeam_donut=find(dd>ri & dd<=ro);
    dtest2(inbeam_donut)=ones(1,length(inbeam_donut));
    ttest_donut=ttest.*dtest2;
    inbeam_donut2=find(ttest_donut>0);

    icount=0;

    for i=1:length(inbeam_hole2);
        fname1=GOOD(inbeam_hole2(i)).name;
        load(fname1);
        fname1_vertical_data=data(3).data(sample_before:sample_after);
        corrn=max(conv(detrend(centroid_vertical_data),fliplr(detrend(fname1_vertical_data))));
        corrd1=max(conv(detrend(centroid_vertical_data),fliplr(detrend(centroid_vertical_data))));
        corrd2=max(conv(detrend(fname1_vertical_data),fliplr(detrend(fname1_vertical_data))));
        corr=corrn/sqrt((corrd1)*corrd2);
        if corr>=corrcoeff;
            icount=icount+1;
            MT_hole(icount).name=fname1;
        end
    end

    jcount=0;

    for j=1:length(inbeam_donut2);
        fname2=GOOD(inbeam_donut2(j)).name;
        load(fname2);
        fname2_vertical_data=data(3).data(sample_before:sample_after);
        corrn=max(conv(detrend(centroid_vertical_data),fliplr(detrend(fname2_vertical_data))));
        corrd1=max(conv(detrend(centroid_vertical_data),fliplr(detrend(centroid_vertical_data))));
        corrd2=max(conv(detrend(fname2_vertical_data),fliplr(detrend(fname2_vertical_data))));
        corr=corrn/sqrt((corrd1)*corrd2);
        if corr>=corrcoeff;
            jcount=jcount+1;
            MT_donut(jcount).name=fname2;
        end
    end

    whos
    toc
    disp('time for RFbeamer loop')
    pause
    tic
    beambld_9comp_2011(MT_hole,MT_donut,centroid,ri,ro)
    toc
    disp('time for beambld loop')
    pause
    clear MT_donut
    number_in_hole=icount;
    number_in_donut=jcount;
    keep GOOD m ij timelist stlats stlons freq months_left ri ro reference_dt cutwindow corrcoeff window
end

Sorry for the long code. Any help for my problem would be greatly appreciated, (along with any advice for speeding it up in general)!

Answer 1

[Jan]

You can use the profiler to find out, which lines are taking more time in the second run. E.g. create one profile report when the program stops after the first iteration, and one with running two iterations.