Fault Tolerant Memory Allocation Interface for Software-based Checksums

BBB-Link: https://bbb.tu-dresden.de/b/mat-xin-oyh-xzn

Presentation Language: English

With the continued scaling of semiconductor technologies, faults have become an increasingly relevant factor for system reliability, as they can silently corrupt the program state and lead to incorrect system behaviour. This thesis investigates a software-based approach to memory fault detection by enforcing checksums through a memory allocation interface. A custom safe-pointer abstraction was designed and implemented in C++ for a bare-metal execution environment, to protect dynamically allocated memory regions while keeping the application largely unchanged. The proposed interface can be integrated directly at the application level without requiring specialized hardware support or a specific toolchain, while also allowing developers to selectively protect vulnerable data structures and tailor protection mechanisms based on application-specific knowledge. To this end, multiple checksum application and verification strategies are examined in depth, differing in the underlying checksum computation, the data granularity considered, and when verification is performed. The implementation leverages experimental C++26 compile-time reflection features to support different data types without manual type-specific handling. The effectiveness of the approach is evaluated using the FAIL* fault injection framework in combination with benchmarks from TACLeBench. The results demonstrate that the proposed memory allocation interface with in-memory checksums can substantially reduce silent data corruptions by deliberately managing trade-offs between performance overhead, detection latency, and fault tolerance through selective application strategies