Reliability has emerged as a first class design concern, as a result of an
exponential increase in the number of transistors on the chip, and lowering of
operating and threshold voltages with each new process generation.
Radiation-induced transient faults are a significant source of soft errors in
current and future process generations. Techniques to mitigate their effect come
at a significant cost of area, power, performance, and design effort.
Architectural Vulnerability Factor (AVF) modeling has been proposed to easily
estimate the processor's soft error rates, and to enable the designers to make
appropriate cost/reliability trade-offs early in the design cycle. Using cycle-accurate
microarchitectural or logic gate-level simulations, AVF modeling captures the
masking effect of program execution on the visibility of soft errors at the
output. AVF modeling is used to identify structures in the processor that have
the highest contribution to the overall Soft Error Rate (SER) while running
typical workloads, and used to guide the design of SER mitigation mechanisms.
The precise mechanisms of interaction between the workload and the
microarchitecture that together determine the overall AVF is not well studied in
literature, beyond qualitative analyses. Consequently, there is no known
methodology for ensuring that the workload suite used for AVF modeling offers
sufficient SER coverage. Additionally, owing to the lack of an intuitive model,
AVF modeling is reliant on detailed microarchitectural simulations for
understanding the impact of scaling processor structures, or design space
exploration studies. Microarchitectural simulations are time-consuming, and do
not easily provide insight into the mechanisms of interactions between the
workload and the microarchitecture to determine AVF, beyond aggregate
statistics.
These aforementioned challenges are addressed in this dissertation by developing
two methodologies.
First, beginning with a systematic analysis of the factors affecting the occupancy of
corruptible state in a processor, a methodology is developed that
generates a synthetic workload for a given microarchitecture such that the SER
is maximized. As it is impossible for every bit in the processor to
simultaneously contain corruptible state, the worst-case realizable SER
while running a workload is less than the sum of their circuit-level fault rates.
The knowledge of the worst-case SER enables efficient design trade-offs by
allowing the architect to validate the coverage of the workload suite and select
an appropriate design point, and to identify structures that may potentially have
high contribution to SER. The methodology
induces 1.4X higher SER in the core as compared to the highest SER induced
by SPEC CPU2006 and MiBench programs.
Second, a first-order analytical model is proposed, which is developed from
the first principles of out-of-order superscalar execution that models the AVF
induced by a workload in microarchitectural structures, using inexpensive
profiling. The central component of this model is a methodology to estimate the
occupancy of correct-path state in various structures in the core. Owing to its
construction, the model provides fundamental insight into the precise mechanism
of interaction between the workload and the microarchitecture to determine AVF.
The model is used to cheaply perform
sizing studies for structures in the core, design space exploration, and workload
characterization for AVF. The model is used to quantitatively explain results
that may appear counter-intuitive from aggregate performance metrics. The Mean
Absolute Error in determining AVF of a 4-wide out-of-order superscalar processor
using model is less than 7% for each structure, and the Normalized Mean Square
Error for determining overall SER is 9.0%, as compared to cycle-accurate microarchitectural simulation. / text
| Identifer | oai:union.ndltd.org:UTEXAS/oai:repositories.lib.utexas.edu:2152/ETD-UT-2012-05-5018 |
| Date | 11 July 2012 |
| Creators | Nair, Arun Arvind |
| Source Sets | University of Texas |
| Language | English |
| Detected Language | English |
| Type | thesis |
| Format | application/pdf |
Page generated in 0.0018 seconds