Map-making is one of the crucial steps in the analysis of the CMB data
sets which can be done by applying the maximum likelihood approach.
This approach yields a solution in a form of a generalized least
square problem, which several software package such as MADmap solve to
produce an estimate of the sky. Enormous sizes of data sets
anticipated from the next generation of the CMB experiments require
massively parallel implementations of the map-making algorithm
suitable for high performance computing (HPC) systems. In this context
communication between
computational nodes is quickly becoming one of the major challenges,
which need to be robustly addressed to ensure scalability of the
map-making codes. Indeed, the analysis of the MADmap code, which uses
the preconditioned conjugate gradient (PCG) algorithm to compute the
map of the sky, shows that the cost of the communication required by
such an algorithm is usually significant, and on occasions dominant,
as compared to the cost of the entire procedure. For example,
Cantalupo et al (2010) find that for a simulated data set of a
Planck-like experiment 24.0% of run time is typically spent on
performing communication. We present a study of so-called
communication-avoiding, iterative solvers, such as the s-step PCG
method, as applied to generalized least squares systems in the context
of generic and CMB map-making applications.