Fredagskollokvium: Mikolaj Szydlarski: Accelerating Cosmic Microwave Background Data Analysis
Mikolaj Szydlarski, postdoktor, Institutt for teoretisk astrofysikk, UiO.
Abstract: Estimation of sky signals from sequences of time order data is one of the key steps in the Cosmic Microwave Background (CMB) data analysis, commonly referred to as a map-making. Some of the most popular and general methods proposed for this task involve solving a general least square problem with non-diagonal noise weights given by a block-diagonal matrix with blocks describing noise correlations of segments of the data over which it was stationary. Given sizes of current and anticipated CMB data sets, composed of as many as tens and hundreds billions of samples, solving such problems requires significant computational power, potentially frustrating applications of such methods to actual data, even on the largest available supercomputers.
In this work we discuss approaches to improving performance of the map-making solvers in order to permit their application to even the largest, anticipated data sets. In this context, we study new iterative algorithms based on a conjugate gradient (CG) approach enhanced with a novel, parallel, two-level preconditioner (2lvl-PCG). We discuss in detail their implementation for massively parallel computational platforms, their performance, and in particular dependence of their convergence rate on parameters characterizing analyzed data set.
We apply the proposed solvers to examples of simulated, non-polarized and polarized CMB observations and a set of idealized scanning strategies with a sky coverage ranging from nearly a full sky down to small sky patches.
We find that our new solver outperforms by as much as a factor 5 standard solvers as used today and does so in terms of both the convergence rate and time to the solution and without increasing significantly memory consumption or volume of the inter-processor communication. We therefore conclude that the proposed approach is well suited to address successfully challenges posed by new and forthcoming CMB data sets.