DNS means you are performing a 3D unsteady simulation, what are such residuals?

How do you monitor in time the development of the flow?

The mean of residuals would be low value since it depends on number of iterations involved in DNS simulation scenario. I.e. in your case fully turbulent and 3D unsteady flows in pipe.

The number of cells for your coarse grid are also high, considering this solution method, the tolerance for residuals can be further reduced , of order 1e-10 which can possibly reduce the magnitude of residuals at each iteration.

The step size for each iteration can be another factor which affects magnitude of residuals during iteration, check for these changes.

Try with double precision numbers, defined at the begining of the runing solution

Check the Courant number on the report, it will give you an idea, if you need to increase the mesh or reduce the timestep

I am interested in your question but it needs a lot more explanation. Let me explain what I'm wondering about that I'd need to know to think about your question: If you are solving a F(u)=b, the residual is

Residual = b-F(u_approximate).

This is a vector with a lot of components so plots are usually some sort of aggregate statistic.

So what is your system?? Sometimes, people solve F(u)=b by time stepping: (u_n+1 - u_n)/k + F(u_n)=b then the residual means the discrete time derivative. Sometimes codes are written to be very memory efficient so they calculate something they call a residual that is just some easy to get data that serves as an optimistic proxy.

When the residuals are getting converged , check whether you can tweak with the boundary and/ or initial conditions for the problem in such a way that the residuals still remain converged and output consistent values.