Towards Dependability in Everyday Software Using Software Telemetry

Abstract

Application-level software dependability is difficult to ensure. Thus it's typically used only in custom systems and is achieved using one-of-a-kind, handcrafted solutions. We are interested in understanding whether and how these techniques can be applied to more common, lower-end systems. To this end, we have adapted a condition-based maintenance (CBM) approach called the multivariate state estimation technique (MSET). This approach automatically creates sophisticated statistical models that predict system failure well before failures occur, leading to simpler and more successful recoveries. We have packaged this approach in the Software Dependability Framework (SDF). The SDF consists of instrumentation and data management libraries, a CBM module, performance visualization tools, and a software architecture that supports system designers. Finally, we evaluated our framework on a simple video game application. Our results suggest that we can cheaply and reliably predict impending runtime failures and respond to them in time to improve the system's dependability