Analysis of the architecture and organization of protein structures is a major challenge to better understand protein flexibility, folding, functions and interactions with their partners and to design new drugs. Protein structures are often described as series of alpha-helices and beta-sheets, or at a higher level as an arrangement of protein domains. Due to the lack of an intermediate vision which could give a good understanding and description of protein structure architecture, we have proposed a novel intermediate view, the Protein Units (PUs). They are novel level of protein structure description between secondary structures and domains. A PU is defined as a compact sub-region of the 3D structure corresponding to one sequence fragment, defined by a high number of intra-PU contacts and a low number of inter-PU contacts. The methodology to obtain PUs from the protein structures is named Protein Peeling (PP). For the algorithm, the protein structures are described as a succession of Ca. The distances between Ca are translated into contact probabilities using a logistic function. Protein Peeling only uses this contact probability matrix. An optimization procedure, based on the Matthews' coefficient correlation (MCC) between contacts probability sub matrices, defines optimal cutting points that separate the region examined into two or three PUs. The process is iterated until the compactness of the resulting PUs reaches a given limit. An index assesses the compactness quality and relative independence of each PU. Protein Peeling is a tool to better understand and analyze the organization of protein structures. We have developed a dedicated bioinformatic web server: Protein Peeling 2 (PP2). Given the 3D coordinates of a protein, it proposes an automatic identification of protein units (PUs). The interface component consists of a web page (HTML) and common gateway interface (CGI). The user can set many parameters and upload a given structure in PDB file format to a perl core instance. This last component is a module that embeds all the information necessary for two others softwares (mainly coded in C to perform most of the computation tasks and R for the analysis). Results are given both textually and graphically using JMol applet and PyMol software. The server can be accessed from http://www.dsimb.inserm.fr/dsimb_tools/peeling/. Only one equivalent on line methodology is available.