A hybrid self-diagnosis mechanism with defective nodes locating and attack detection for parallel computing systems

Abstract

In recent years parallel computing has been widely employed for both science research and commercial applications. For parallel systems such as many-core or computer clusters, it is inevitable to have one or more computing node failures due to random errors or injected attacks. Usually a diagnosis mechanism is able to locate several defective nodes through a number of tests and the analysis of those test signatures (syndromes). Although this covers the cases caused by random errors, sophisticated attacks are still able to manipulate the outputs of each node, so that they will be masked and pass the diagnosis. Therefore in this paper we propose a hybrid self-diagnosis mechanism. We adopt a new type of analysis with the linear syndromes, which are able to locate up to a certain number of defective nodes caused by random errors. In addition to this, we introduce a new type of robust analysis of the non-linear syndromes, which is capable of detecting the attacks undetectable by the linear syndromes at a probability close to one. Moreover, since this hybrid self-diagnosis mechanism is on the data level which makes little distinction among different operating systems and programming languages, it can be migrated onto any other platforms conveniently.