Skip to content

Design

The Hamartia Error and Detector API was created in order to gain a better understanding of how errors manifest using the Monte Carlo approach. However, the API can be used for other error/detector-related modelling as well since it is a general API.

Motivation

  • Simple fault models are currently used
    • Such as flipping a bit in the result
    • More precise models should be used/evaluated
  • Fault models have to be reimplemented each time
    • There should be a common, shared interface
    • Can be used across projects and injectors
  • Injection should be fast and have low-overhead
    • Separate levels of abstraction
    • Faults should be guaranteed (at some level), focus on the error manifestation
  • Allow for large-scale experiments and simulations

Design Decisions

Monte-Carlo Injection

We adopt the Monte Carlo method to injection since the search space is very vast (in most cases). Monte Carlo method will help to approximate the effects of different errors in this space and hopefully lead to a better understanding of what errors look like so that they can be better detected and corrected. However, there are built-in settings for narrowing the injection space and targeting certain parts of a program if needed.

Guaranteed Errors with Nested Monte Carlo

All error models will create a guaranteed error (unmasked). This is verified and generated through an iterative approach (random). The main idea for only injecting unmasked errors is that errors which are masked will not cause any effects on the program. However, we do track statistics on how many iterations it took to get a unmasked error. See our SC18 paper for more details.

Unified and Cross-language Interface

If there is an easy to implement interface for all error/detector models, researchers can share models and run them with different injectors. We currently support instruction-level injection and RTL gate-level injection with C++ and Python; however, the API is fairly generic enough where it could be extended to more circuit-level injection (or other targets) in the future.

flow