Configurable Detection of SDC-causing Errors in Programs

Abstract

Silent Data Corruption (SDC) is a serious reliability issue in many domains, including embedded systems. However, current protection techniques are brittle, and do not allow programmers to trade o performance for SDC coverage. Further, many of them require tens of thousands of fault injection experiments, which are highly time-intensive. In this thesis, we propose two empirical models, namely SDCTune and SDCAuto, to predict the SDC proneness of a program's data. Both models are based on static and dynamic features of the program alone, and do not require fault injections to be performed. The dierence between the two models is that SDCTune is built using a manual tuning process, while SDCAuto is built using a machine learning algorithm. We then develop an algorithm using both models to selectively protect the most SDC-prone data in the program subject to a given performance overhead bound. Our results show that both models are accurate at predicting the relative SDC rate of an application. And in terms of eciency of detection (i.e., ratio of SDC coverage provided to performance overhead), our technique outperforms full duplication by a factor of 0.78x to 1.65x with SDCTune model, and 0.62x to 0.96x with SDCAuto model.

Identifying SDC-Causing Instructions Based on Random Forests Algorithm

“…Step 3 strengthens the identified vulnerable blocks. The work [16] introduces a prediction model named SDCAuto to predict the SDC proneness of a program's instructions. SDCAuto is built using CART algorithm, requiring little to no human intervention.…”

“…2 illustrates the diagram of the work [16]. The work [16] first compiles the source code into LLVM IR, and extracts instruction features based on LLVM IR file. Then, it obtains the SDC proneness for each instruction with the help of SDCAuto model.…”

“…Then, it obtains the SDC proneness for each instruction with the help of SDCAuto model. [16] Finally, detectors are inserted into the source code for protecting the most SDC-prone instructions for the user-specified overhead bound.…”

“…Recently, a lot of works [10][11][12][13][14][15][16] try to improve the static injection framework. CriticalFault [10] applies vulnerability analysis to avoid the derated fault injections.…”

“…The work [16] presents a selective protection technique that allows users to selectively protect these SDC-prone data. The main idea of work [16] is predicting the SDC proneness of a program's data firstly, then selectively protects the most SDC-prone instructions of the program for the user-specified overhead bound. Since the prediction model is built based on the Classification and Regression Tree (CART) algorithm, the process of prediction do not need to perform fault injections.…”