Fingerprinting:
Low-overhead Error Detection on Commodity Microprocessors
Tuesday October 18, 2005
Hamerschlag Hall D-210
4:00 pm
Jared Smolens
Carnegie Mellon University
Future microprocessors will be increasingly susceptible to soft
errors from many sources, including decreasing node capacitances
and process variation. Detecting these errors in a timely manner
and cost-efficient manner is a key problem in designing high-reliability
computer systems. In this talk, I present two forms of fingerprinting,
a low-overhead error detection technique that detects architectural
differences in execution across redundant processor cores. First,
microarchitectural fingerprints capture a cumulative summary of
values on internal data and control path nodes in the microprocessor.
These can be implemented with small modifications to existing design-for-test
hardware, but require processor cores to operate in lock-step. Alternatively,
architectural fingerprints summarize a processor's recent architectural
state updates in a small signature for low bandwidth comparison,
without capturing timing-dependent behavior. These require modest
additions to the microprocessor's retirement stages, but eliminate
the lock-step requirement. In this talk, I will show how these fingerprinting
mechanisms work and how they can be implemented in an existing commodity
microprocessor design.
Jared Smolens is a PhD candidate in the ECE department at Carnegie
Mellon University, where he is advised by Prof. James Hoe. He received
his BS and MS in ECE from Carnegie Mellon University. His research
interests are in multiprocessor and microprocessor architectures,
fault tolerance, and performance modeling. He has spent this past
summer working to prove out fingerprinting at Intel.
|