Soft errors: the hardware-software interface

A recent report from the ITRS identifies soft errors, as one of the most important reliability challenges for the coming decades. Soft errors are transient errors caused by several effects e.g., voltage fluctuations, wire-cross talks, and cosmic particle strikes; and manifest as a temporary switch of the logic value of a transistor. While it is not possible to prove nor disprove that a certain error happened due to soft errors, several fiscal disasters e.g., Sun server crashes in 2000, and HP server crashes in 2005, have been attributed to soft errors. Industry has moved from the position of ignoring soft errors to adding design efforts for protection from them. For instance, in the recently announced nVIDIA's Fermi GPUs, the L1 cache, L2 cache and register files are ECC protected. Although the soft error rate is about once-per year today, it is expected to reach alarming levels of once-per-day in about a decade or two. Researchers are busy finding cost-effective solutions to protect computing devices from soft errors.

This tutorial will attempt to cover the entire gamut of soft error protection techniques, but will particularly focus on the soft error mitigation techniques at the hardware/software interface. Much time will be spent on microarchitectural, compiler, and hybrid compiler-microarchitectural techniques for soft error mitigation. This tutorial will be particularly useful for budding researchers who are fascinated by soft errors, and want to explore this as their research direction. For such researchers, this tutorial will be a one-stop-shop to acquire knowledge of and analyze seminal research work in the field of soft error mitigation, at several design layers. For developers who have been working on soft errors at different levels, this will give them a picture of what can be done at other levels, so that they can provide complementary cross-layer protection. Finally, researchers and developers working on other aspects of system design can learn how soft errors are going to affect them.