TL;DR
This paper investigates the energy efficiency of dense linear algebra routines when operating at near-threshold voltages (NTV) to save power, a technique that increases error rates. The authors develop an "iso-energy" model to determine how much error detection and correction overhead can be tolerated while maintaining the same energy efficiency as error-free execution, finding that software resilience can indeed complement hardware reliability in low-power scenarios.
What Problem Does This Paper Address?
Modern computing systems face a growing "energy wall," where power consumption limits performance scaling. Near-threshold voltage (NTV) computing offers a way to significantly reduce power by operating processors at very low voltages, close to their minimum operating threshold. However, this comes at the cost of increased susceptibility to soft errors (SE), which can lead to silent data corruption (SDC). For critical applications like dense linear algebra (DLA) used in scientific computing, these errors can have serious consequences. This paper addresses the fundamental question: can the energy savings from NTV computing be maintained even when implementing software-based resilience mechanisms to combat these increased error rates? Specifically, it explores the trade-off between energy efficiency gains from voltage-frequency scaling (VFS) and the overhead introduced by fault tolerance techniques for DLA kernels.
Methods
The core of this research lies in developing an "iso-energy" model. This model aims to quantify the conditions under which a fault-tolerant version of a DLA routine, operating at low voltage and potentially experiencing errors, can achieve the same energy efficiency as its error-free counterpart running at a higher, safer voltage.
The authors distinguish between two types of overheads:
- Detection Overhead ($O_d$): This is a fixed cost incurred for error detection, independent of the number of errors.
- Correction Overhead ($O_c$): This cost is proportional to the error rate, as correction is only needed when errors occur.
They model CPU power as a combination of dynamic and static components, fitting these models to experimental data from ARM processors. Using these power models, they derive iso-energy formulas that relate the acceptable correction overhead (and thus error rate) to voltage-frequency configurations and detection overheads.
The study focuses on three key DLA kernels:
- GEMV (Matrix-vector multiplication): Treated as a memory-bound operation.
- GEMM (Matrix-matrix multiplication): Treated as a compute-bound operation.
- GETRF (LU factorization with partial pivoting): Built on top of GEMM.
They analyze different scenarios:
- Undervolting: Voltage is reduced while frequency remains constant.
- General VFS: Both voltage and frequency are reduced simultaneously, with various relationships between them (constant, linear, quadratic, and modeled).
Datasets
The paper does not introduce new benchmark datasets. Instead, it uses standard dense linear algebra operations (GEMV, GEMM, GETRF) and evaluates their performance and energy consumption on real hardware.
Key Insights and Ideas
- Software Resilience is Viable for NTV: The study demonstrates that software-based fault tolerance can indeed complement hardware reliability in low-power, near-threshold voltage scenarios, allowing for energy savings without sacrificing overall energy efficiency.
- Detection Overhead is the Bottleneck: The fixed cost of error detection is a significant factor. Even without errors, voltage must be reduced to a certain extent just to compensate for this detection overhead.
- Memory-Bound vs. Compute-Bound Differences: The energy efficiency trade-offs differ significantly between memory-bound (GEMV) and compute-bound (GEMM, GETRF) operations. For compute-bound tasks, frequency scaling has a more direct impact on performance, making it easier to compensate for overheads. Memory-bound operations are more sensitive to memory bandwidth, limiting their scalability with frequency and making it harder to achieve energy efficiency gains through VFS when error correction is involved.
- Iso-Energy Curves are Informative: The developed iso-energy curves provide a clear visualization of the acceptable error correction overhead for different voltage scaling levels and detection overheads, offering practical guidance for system design.
- Model Extrapolation is Necessary but Carefully Handled: The authors acknowledge that directly measuring power at very low voltages is challenging and rely on extrapolation. They mitigate this by focusing on extrapolation examples within 85-95% of the lowest nominal voltage, a range where their models are more likely to be accurate.
Key Results
- The paper presents experimental results on ARM multicore processors (Cortex-A7, A15, A53, A57) showing performance and energy efficiency metrics for GEMV, GEMM, and GETRF across various voltage-frequency configurations.
- The developed iso-energy models show that for compute-bound operations (GEMM, GETRF), the relationship between voltage, frequency, and power is simpler (static power is negligible at low voltages). This makes it easier to find iso-energy configurations.
- For memory-bound GEMV, the static power component is more significant, making it harder to compensate for detection and correction overheads through VFS alone.
- The study provides specific examples, such as for the Cortex-A15, showing that even with no errors, voltage must be reduced by about 5% to compensate for a 10% detection overhead. With a 28% correction overhead, voltage can be reduced to 85% of its nominal value to maintain the same energy efficiency.
Limitations and Future Work
- Extrapolation of Power Models: The authors acknowledge that their study relies on extrapolating power models to very low voltages, which introduces some uncertainty. They mitigate this by focusing on values within 85-95% of the lowest nominal voltage.
- Simplified Fault Model: The paper focuses on Silent Data Corruption (SDC) and does not delve deeply into other fault models or the reliability of the error detection/correction mechanisms themselves, which could also be subject to errors.
- Theoretical vs. Practical Implementation: The detection+correction strategies discussed are presented as theoretical proposals, and the actual implementation details are not the focus.
- Future Work: The authors suggest that a detailed analysis of the reliability of error detection and correction mechanisms is a clear target for future research aiming for practical and efficient resilient implementations. They also note that their methodology can be extended to other architectures (e.g., Intel, AMD) and other DLA operations.
Summary updated