1. Dynamic Prediction of Architectural Vulnerability
Kristen Walcott, Greg Humphreys, Sudhanva Gurumurthi
University of Virginia {walcott, humper,
Challenge
As soft errors become more of a problem, protection will be needed even for every day PCs.
Providing total redundancy is too expensive and assumes that AVF is 100%.
Our work shows that AVF varies over time.
Transient faults due to particle strikes are a key challenge in microprocessor design.
As transistor counts increase exponentially, per-chip faults are a growing burden.
Spatial and temporal redundancy techniques are used to protect against faults.
Redundancy techniques assume that any fault will result in a visible program error (i.e., the Architectural Vulnerability Factor (AVF)) is 100 percent.
Over-design can hurt performance and drain power.
Rising Problem
Dynamic AVF Prediction
Outliers
Intel Corporation
FIT = Failure in Time = 1 failure in a billion hours
Prediction Results
Bit
Read?
Bit has error protection
benign fault
no error
yes no
Does bit matter?
Particle Strike
Causes Bit Flip!
Detection
only
Detection &
Correction
Silent Data Corruption
True Detected Unrecoverable Error
False Detected Unrecoverable Error
We identify strong correlations between structural AVF values and a small set of processor metrics.
Using linear and quadratic regression, we came up with an AVF characterization that uses only a few variables. These characterizations can be used to predict AVF accurately!
What bits matter?
Future Work
With an accurate predictor, redundancy may be turned on only when vulnerability is high. Preliminary results show that partial redundancy provides a significant performance boost over full redundancy. Next we will perform a more rigorous exploration of the design space of partial redundant multithreading implementations and investigate redundancy toggling policies.
Calculating Vulnerability
AVFbit = Probability Bit Matters =
# of Visible Errors
# of Bit Flips from Particle Strikes
Computer Science
at the UNIVERSITY of VIRGINIA