Fault-Injection-Based Assessment of Fault Tolerance Measures with FAIL*

Motivation

Persistent Memory is a group of technologies that enables fast, byte addressable main memory with the ability to maintain its state without a constant power supply. Prominent examples are Intel Optane products or battery backed NVDIMM modules from Viking Technology. As data in this memory is meant to persist, reliability is an important concern since an error recovery cannot be done via a simple system reboot without losing information. However, all currently available technologies of persistent memory suffer from transient hardware faults. Although hardware solutions exist, they are not always sufficient for the task or are not available due to economic reasons.

The DeNoVo Malloc memory allocator, that is being developed at the IBR, provides not only a familiar malloc interface to persistent memory but also protects all the memory it manages transparently from transient faults. Although microbenchmarks have shown that DeNoVo malloc’s data encryption can reliably detect more than 7 bit-flips per double word, it is hard to assess the reliability gain from DeNoVo malloc for a whole application without tool support.

Task Description

The goal of this thesis is to assess and improve the reliability gain from DeNoVo malloc with the help of the fault injection tool FAIL*.
This involves:

• Adapting the DeNoVo malloc library to a minimal Linux kernel or barebone execution.
• Writing suitable benchmark applications.
• Conducting fault injection campaigns with FAIL*
• Use the insights from the fault injection campaigns to improve DeNoVo malloc

Links

• Fail*
• Intel 3D XPoint
• Viking ArxCis-NV Non-Volatile Nemory