Try this:
k=0;
for p=1:1:7
t=cputime;
dt(p)=10^-p;
k=0:1.35/dt(p);
xs = 2/(sqrt(pi))*sum(exp(-(k*dt(p)).^2)*dt(p))
e(p)=cputime-t;
end
semilogx(1./dt, e)
You were plotting timestep as if it were a consequence of execution time. Also, as p increases, dt decreases, so if you plot dt then it is getting smaller and smaller and so your datapoints were getting further left, which is more difficult for people to understand. If you plot against 1./dt then you are plotting time as a consequence of number of data samples used, which is much more natural for people.
